We propose the task of Panoptic Scene Completion~(PSC) which extends the recently popular Semantic Scene Completion (SSC) task with instance-level information to produce a richer understanding of the 3D scene. Our PSC proposal utilizes a hybrid mask-based technique on the non-empty voxels from sparse multi-scale completions. Whereas the SSC literature overlooks uncertainty which is critical for robotics applications, we instead propose an efficient ensembling to estimate both voxel-wise and instance-wise uncertainties along PSC. This is achieved by building on a multi-input multi-output (MIMO) strategy, while improving performance and yielding better uncertainty for little additional compute. Additionally, we introduce a technique to aggregate permutation-invariant mask predictions. Our experiments demonstrate that our method surpasses all baselines in both Panoptic Scene Completion and uncertainty estimation on three large-scale autonomous driving datasets. Our code and data will be made public.

3D reconstruction from a single-view is challenging because of the ambiguity from monocular cues and lack of information about occluded regions. Neural radiance fields (NeRF), while popular for view synthesis and 3D reconstruction, are typically reliant on multi-view images. Existing methods for single-view 3D reconstruction with NeRF rely on either data priors to hallucinate views of occluded regions, which may not be physically accurate, or shadows observed by RGB cameras, which are difficult to detect in ambient light and low albedo backgrounds. We propose using time-of-flight data captured by a single-photon avalanche diode to overcome these limitations. Our method models two-bounce optical paths with NeRF, using lidar transient data for supervision. By leveraging the advantages of both NeRF and two-bounce light measured by lidar, we demonstrate that we can reconstruct visible and occluded geometry without data priors or reliance on controlled ambient lighting or scene albedo. In addition, we demonstrate improved generalization under practical constraints on sensor spatial- and temporal-resolution. We believe our method is a promising direction as single-photon lidars become ubiquitous on consumer devices, such as phones, tablets, and headsets.

The perception of motion behavior in a dynamic environment holds significant importance for autonomous driving systems, wherein class-agnostic motion prediction methods directly predict the motion of the entire point cloud. While most existing methods rely on fully-supervised learning, the manual labeling of point cloud data is laborious and time-consuming. Therefore, several annotation-efficient methods have been proposed to address this challenge. Although effective, these methods rely on weak annotations or additional multi-modal data like images, and the potential benefits inherent in the point cloud sequence are still underexplored.To this end, we explore the feasibility of self-supervised motion prediction with only unlabeled LiDAR point clouds. Initially, we employ an optimal transport solver to establish coarse correspondences between current and future point clouds as the coarse pseudo motion labels. Training models directly using such coarse labels leads to noticeable spatial and temporal prediction inconsistencies. To mitigate these issues, we introduce three simple spatial and temporal regularization losses, which facilitate the self-supervised training process effectively. Experimental results demonstrate the significant superiority of our approach over the state-of-the-art self-supervised methods.Code will be available.

A unified and versatile LiDAR segmentation model with strong robustness and generalizability is desirable for safe autonomous driving perception. This work presents M3Net, a one-of-a-kind framework for fulfilling multi-task, multi-dataset, multi-modality LiDAR segmentation in a universal manner using just a single set of parameters. To better exploit data volume and diversity, we first combine large-scale driving datasets acquired by different types of sensors from diverse scenes and then conduct alignments in three spaces, namely data, feature, and label spaces, during the training. As a result, M3Net is capable of taming heterogeneous data for training state-of-the-art LiDAR segmentation models. Extensive experiments on twelve LiDAR segmentation datasets verify our effectiveness. Notably, using a shared set of parameters, M3Net achieves 75.1%, 83.1%, and 72.4% mIoU scores, respectively, on the official benchmarks of SemanticKITTI, nuScenes, and Waymo Open.

In this paper, we introduce the first large-scale video prediction model in the autonomous driving discipline. To eliminate the restriction of high-cost data collection and empower the generalization ability of our model, we acquire massive data from the web and pair it with diverse and high-quality text descriptions. The resultant dataset accumulates over 2000 hours of driving videos, spanning areas all over the world with diverse weather conditions and traffic scenarios. Inheriting the merits from recent latent diffusion models, our model, dubbed GenAD, handles the challenging dynamics in driving scenes with novel temporal reasoning blocks. We showcase that it can generalize to various unseen driving datasets in a zero-shot manner, surpassing general or driving-specific video prediction counterparts. Furthermore, GenAD can be adapted into an action-conditioned prediction model or a motion planner, holding great potential for real-world driving applications.

In contrast to extensive studies on general vision, pre-training for scalable visual autonomous driving remains seldom explored. Visual autonomous driving applications require features encompassing semantics, 3D geometry, and temporal information simultaneously for joint perception, prediction, and planning, posing dramatic challenges for pre-training. To resolve this, we bring up a new pre-training task termed as visual point cloud forecasting - predicting future point clouds from historical visual input. The key merit of this task captures the synergic learning of semantics, 3D structures, and temporal dynamics. Hence it shows superiority in various downstream tasks. To cope with this new problem, we present ViDAR, a general model to pre-train downstream visual encoders. It first extracts historical embeddings by the encoder. These representations are then transformed to 3D geometric space via a novel Latent Rendering operator for future point cloud prediction. Experiments demonstrate significant improvements in downstream tasks, e.g., 3.1% NDS on 3D detection, ∼10% error reduction on motion forecasting, and ∼15% less collision rate on planning. Code and models would be public.

We tackle semi-supervised object detection based on motion cues. Recent results suggest that heuristic-based clustering methods in conjunction with object trackers can be used to pseudo-label instances of moving objects and use these as supervisory signals to train 3D object detectors in Lidar data without manual supervision. We re-think this approach and suggest that both, object detection, as well as motion-inspired pseudo-labeling, can be tackled in a data-driven manner. We leverage recent advances in scene flow estimation to obtain point trajectories from which we extract long-term, class-agnostic motion patterns. Revisiting correlation clustering in the context of message passing networks, we learn to group those motion patterns to cluster points to object instances. By estimating the full extent of the objects, we obtain per-scan 3D bounding boxes that we use to supervise a Lidar object detection network. Our method not only outperforms prior heuristic-based approaches (57.5 AP, +14 improvement over prior work), more importantly, we show we can pseudo-label and train object detectors across datasets.

Autonomous driving (AV) systems rely on robust perception models as a cornerstone of safety assurance. However, objects encountered on the road exhibit a long-tailed distribution, with rare or unseen categories posing challenges to a deployed perception model. This necessitates an expensive process of continuously curating and annotating data with significant human effort. We propose to leverage recent advances in vision-language and large language models to design an Automatic Data Engine (AIDE) that automatically identifies issues, efficiently curates data, improves the model through auto-labeling, and verifies the model through generation of diverse scenarios. This process operates iteratively, allowing for continuous self-improvement of the model. We further establish a benchmark for open-world detection on AV datasets to comprehensively evaluate various learning paradigms, demonstrating our method's superior performance at a reduced cost.

Point cloud analysis has achieved outstanding performance by transferring point cloud pre-trained models. However, existing methods for model adaptation usually update all model parameters, i.e., full fine-tuning paradigm, which is inefficient as it relies on high computational costs (e.g., training GPU memory) and massive storage space. In this paper, we aim to study parameter-efficient transfer learning for point cloud analysis with an ideal trade-off between task performance and parameter efficiency. To achieve this goal, we first freeze the parameters of the default pre-trained models and then propose the Dynamic Adapter, which generates a dynamic scale for each point token, considering the token significance to the downstream task. We further seamlessly integrate Dynamic Adapter with Prompt Tuning (DAPT) by constructing Internal Prompts, capturing the instance-specific features for interaction. Extensive experiments conducted on five challenging datasets demonstrate that the proposed DAPT achieves superior performance compared to the full fine-tuning counterparts while significantly reducing the trainable parameters and training GPU memory by 95\% and 35\%, respectively. The code will be made available.

Vision-based roadside 3D object detection has attracted rising attention in autonomous driving domain, since it encompasses inherent advantages in reducing blind spots and expanding perception range. While previous work mainly focuses on accurately estimating depth or height for 2D-to-3D mapping, ignoring the position approximation error in the voxel pooling process. Inspired by this insight, we propose a novel voxel pooling strategy to reduce such error, dubbed BEVSpread. Specifically, instead of bringing the image features contained in a frustum point to a single BEV grid, BEVSpread considers each frustum point as a source and spreads the image features to the surrounding BEV grids with adaptive weights. To achieve superior propagation performance, a specific weight function is designed to dynamically control the decay speed of the weights according to distance and depth. Aided by customized CUDA parallel acceleration, BEVSpread achieves comparable inference time as the original voxel pooling. Extensive experiments on two large-scale roadside benchmarks demonstrate that, as a plug-in, BEVSpread can significantly improve the performance of existing frustum-based BEV methods by a large margin of (1.12, 5.26, 3.01) AP in vehicle, pedestrian and cyclist.

State-of-the-art approaches for autonomous driving integrate multiple sub-tasks of the overall driving task into a single pipeline that can be trained in an end-to-end fashion by passing latent representations between the different modules.In contrast to previous approaches that rely on a unified grid to represent the belief state of the scene, we propose dedicated representations to disentangle dynamic agents and static scene elements. This allows us to explicitly compensate for the effect of both ego and object motion between consecutive time steps and to flexibly propagate the belief state through time. Furthermore, dynamic objects can not only attend to the input camera images, but also directly benefit from the inferred static scene structure via a novel dynamic-static cross-attention. Extensive experiments on the challenging nuScenes benchmark demonstrate the benefits of the proposed dual-stream design, especially for modelling highly dynamic agents in the scene, and highlight the improved temporal consistency of our approach. Our method titled DualAD not only outperforms independently trained single-task networks, but also improves over previous state-of-the-art end-to-end models by a large margin on all tasks along the functional chain of driving.

Diffusion models (DMs) excel in photo-realistic image synthesis, but their adaptation to LiDAR scene generation poses a substantial hurdle. This is primarily because DMs operating in the point space struggle to preserve the curve-like patterns and 3D geometry of LiDAR scenes, which consumes much of their representation power. In this paper, we propose \textbf{LiDAR Diffusion Models} (LiDMs) to generate LiDAR-realistic scenes from a latent space tailored to capture the realism of LiDAR scenes by incorporating geometric priors into the learning pipeline. Our method targets three major desiderata: pattern realism, geometry realism, and object realism. Specifically, we introduce curve-wise compression to simulate real-world LiDAR patterns, point-wise coordinate supervision to learn scene geometry, and patch-wise encoding for a full 3D object context. With these three core designs, our method achieves competitive performance on unconditional LiDAR generation in 64-beam scenario and state of the art on conditional LiDAR generation, while maintaining high efficiency compared to point-based DMs (up to 107$\times$ faster). Furthermore, by compressing LiDAR scenes into a latent space, we enable the controllability of DMs with various conditions such as semantic maps, camera views, and text prompts. %Our code and pretrained weights are available at https://github.com/hancyran/LiDAR-Diffusion.

In autonomous driving, predicting future events in advance and evaluating the foreseeable risks empowers autonomous vehicles to plan their actions, enhancing safety and efficiency on the road. To this end, we propose Drive-WM, the first driving world model compatible with existing end-to-end planning models. Through a joint spatial-temporal modeling facilitated by view factorization, our model is the first to generate high-fidelity multiview videos. Building on its powerful generation ability, we showcase the potential of applying the world model for safe driving planning for the first time. Our Drive-WM enables driving into multiple futures based on distinct driving maneuvers, and determines the optimal trajectory according to the image-based rewards. Evaluation on real-world driving datasets verifies that our method could generate high-quality, consistent, and controllable multiview videos, opening up possibilities for real-world simulations and safe planning.

Autonomous driving is a complex and challenging task that aims at safe motion planning through scene understanding and reasoning. While vision-only autonomous driving methods have recently achieved notable performance, through enhanced scene understanding, several key issues, including lack of reasoning, low generalization performance and long-tail scenarios, still need to be addressed. In this paper, we present VLP, a novel Vision-Language-Planning framework that exploits language models to bridge the gap between linguistic understanding and autonomous driving. VLP enhances autonomous driving systems by strengthening both the source memory foundation and the self-driving car's contextual understanding. VLP achieves state-of-the-art end-to-end planning performance on the challenging NuScenes dataset by achieving 35.9\% and 60.5\% reduction in terms of average L2 error and collision rates, respectively, compared to the previous best method. Moreover, VLP shows improved performance in challenging long-tail scenarios and strong generalization capabilities when faced with new urban environments.

Computer vision techniques play a central role in the perception stack of autonomous vehicles. Such methods are employed to perceive the vehicle surroundings given sensor data. 3D LiDAR sensors are commonly used to collect sparse 3D point clouds from the scene. However, compared to human perception, such systems struggle to deduce the unseen parts of the scene given those sparse point clouds. In this matter, the scene completion task aims at predicting the gaps in the LiDAR measurements to achieve a more complete scene representation. Given the promising results of recent diffusion models as generative models for images, we propose extending them to achieve scene completion from a single 3D LiDAR scan. Previous works used diffusion models over range images extracted from LiDAR data, directly applying image-based diffusion methods. Distinctly, we propose to directly operate on the points, reformulating the noising and denoising diffusion process such that it can efficiently work at scene scale. Together with our approach, we propose a regularization loss to stabilize the noise predicted during the denoising process. Our experimental evaluation shows that our method can complete the scene given a single LiDAR scan as input, producing a scene with more details compared to state-of-the-art scene completion methods. We believe that our proposed diffusion process formulation can support further research in diffusion models applied to scene-scale point cloud data.

LiDAR semantic segmentation (LSS) is a critical task in autonomous driving and has achieved promising progress. However, prior LSS methods are conventionally investigated and evaluated on datasets within the same domain in clear weather. The robustness of LSS models in unseen scenes and all weather conditions is crucial for ensuring safety and reliability in real applications. To this end, we propose UniMix, a universal method that enhances the adaptability and generalizability of LSS models. UniMix first leverages physically valid adverse weather simulation to construct a Bridge Domain, which serves to bridge the domain gap between the clear weather scenes and the adverse weather scenes. Then, a Universal Mixing operator is defined regarding spatial, intensity, and semantic distributions to create the intermediate domain with mixed samples from given domains. Integrating the proposed two techniques into a teacher-student framework, UniMix efficiently mitigates the domain gap and enables LSS models to learn weather-robust and domain-invariant representations. We devote UniMix to two main setups: 1) unsupervised domain adaption, adapting the model from the clear weather source domain to the adverse weather target domain; 2) domain generalization, learning a model that generalizes well to unseen scenes in adverse weather. Extensive experiments validate the effectiveness of UniMix across different tasks and datasets, all achieving superior performance over state-of-the-art methods. The code will be released.

Semantic scene completion, also known as semantic occupancy prediction, can provide dense geometric and semantic information for autonomous vehicles, which attracts the increasing attention of both academia and industry. Unfortunately, existing methods usually formulate this task as a voxel-wise classification problem and treat each voxel equally in 3D space during training. As the hard voxels have not been paid enough attention, the performance in some challenging regions is limited. The 3D dense space typically contains a large number of empty voxels, which are easy to learn but require amounts of computation due to handling all the voxels uniformly for the existing models. Furthermore, the voxels in the boundary region are more challenging to differentiate than those in the interior. In this paper, we propose HASSC approach to train the semantic scene completion model with hardness-aware design. The global hardness from the network optimization process is defined for dynamical hard voxel selection. Then, the local hardness with geometric anisotropy is adopted for voxel-wise refinement. Besides, self-distillation strategy is introduced to make training process stable and consistent. Extensive experiments show that our HASSC scheme can effectively promote the accuracy of the baseline model without incurring the extra inference cost. Source code is available at: https://github.com/songw-zju/HASSC.

Trajectory prediction is fundamental in computer vision and autonomous driving, particularly for understanding pedestrian behavior and enabling proactive decision-making. Existing approaches in this field often assume precise and complete observational data, neglecting the challenges associated with out-of-view objects and the noise inherent in sensor data due to limited camera range, physical obstructions, and the absence of ground truth for denoised sensor data. Such oversights are critical safety concerns, as they can result in missing essential, non-visible objects. To bridge this gap, we present a novel method for out-of-sight trajectory prediction that leverages a vision-positioning technique. Our approach denoises noisy sensor observations in an unsupervised manner and precisely maps sensor-based trajectories of out-of-sight objects into visual trajectories. This method has demonstrated state-of-the-art performance in out-of-sight noisy sensor trajectory denoising and prediction on the Vi-Fi and JRDB datasets. By enhancing trajectory prediction accuracy and addressing the challenges of out-of-sight objects, our work significantly contributes to improving the safety and reliability of autonomous driving in complex environments. Our work represents the first initiative towards Out-Of-Sight Trajectory prediction (OOSTraj), setting a new benchmark for future research.

Currently, high-definition (HD) map construction leans towards a lightweight online generation tendency, which aims to preserve timely and reliable road scene information. However, map elements contain strong shape priors. Subtle and sparse annotations make current detection-based frameworks ambiguous in locating relevant feature scopes and cause the loss of detailed structures in prediction. To alleviate these problems, we propose MGMap, a mask-guided approach that effectively highlights the informative regions and achieves precise map element localization by introducing the learned masks. Specifically, MGMap employs learned masks based on the enhanced multi-scale BEV features from two perspectives. At the instance level, we propose the Mask-activated instance (MAI) decoder, which incorporates global instance and structural information into instance queries by the activation of instance masks. At the point level, a novel position-guided mask patch refinement (PG-MPR) module is designed to refine point locations from a finer-grained perspective, enabling the extraction of point-specific patch information. Compared to the baselines, our proposed MGMap achieves a notable improvement of around 10 mAP for different input modalities. Extensive experiments also demonstrate that our approach showcases strong robustness and generalization capabilities. Our code can be found at https://github.com/xiaolul2/MGMap.

Multi-agent trajectory prediction is essential in autonomous driving, risk avoidance, and traffic flow control. However, the \textbf{heterogeneous traffic density} on \textbf{interactions}, which caused by physical laws, social norms and so on, is often overlooked in existing methods. When the density varies, the number of agents involved in interactions and the corresponding interaction probability change dynamically. To tackle this issue, we propose a new method, called \underline{\textbf{D}}ensity-\underline{\textbf{A}}daptive Model based on \underline{\textbf{M}}otif \underline{\textbf{M}}atrix for Multi-Agent Trajectory Prediction (DAMM), to gain insights into multi-agent systems. Here we leverage the \textbf{motif matrix} to represent dynamic connectivity in a higher-order pattern, and distill the interaction information from the perspectives of the spatial and the temporal dimensions. Specifically, in spatial dimension, we utilize multi-scale feature fusion to adaptively select the optimal range of neighbors participating in interactions for each time slot. In temporal dimension, we extract the temporal interaction features and adapt a pyramidal pooling layer to generate the interaction probability for each agent. Experimental results demonstrate that our approach surpasses state-of-the-art methods on autonomous driving dataset.

Predicting the future occupancy of the surrounding environment is a vital task for autonomous driving. However, current best-performing single-modality or multi-modality fusion methods can only predict uniform snapshots of future occupancy states and still require strictly synchronous sensory data for sensor fusion. We propose a StreamingFlow framework to lift these strong limitations. StreamingFlow is a novel BEV occupancy predictor that ingests asynchronous multi-sensor data streams for fusion and performs streaming forecasting of the future occupancy map at any future timestamps. By integrating neural ordinary differential equations (N-ODE) onto recurrent neural networks, StreamingFlow learns derivatives of BEV features over temporal horizons, updates the implicit sensor's BEV feature as part of the fusion process, and propagates BEV states to the desired future time point. Extensive experiments on two large-scale datasets, nuScenes and Lyft L5, demonstrate that StreamingFlow significantly outperforms previous vision-based, lidar-based methods, and shows competitive performance compared to state-of-the-art fusion-based methods with a much lighter model.

LiDAR-based 3D object detection plays an essential role in autonomous driving. Existing high-performing 3D object detectors usually build dense feature maps in the backbone network and prediction head. However, the computational costs introduced by the dense feature maps grow quadratically as the perception range increases, making these models hard to scale up to long-range detection. Some recent works have attempted to construct fully sparse detectors to solve this issue; nevertheless, the resulting models either rely on a complex multi-stage pipeline or exhibit inferior performance. In this work, we propose SAFDNet, a straightforward yet highly effective architecture, tailored for fully sparse 3D object detection. In SAFDNet, an adaptive feature diffusion strategy is designed to address the center feature missing problem. We conducted extensive experiments on Waymo Open, nuScenes, and Argoverse2 datasets. SAFDNet performed slightly better than the previous SOTA on the first two datasets but much better on the last dataset, which features long-range detection, verifying the efficacy of SAFDNet in scenarios where long-range detection is required. Notably, on Argoverse2, SAFDNet surpassed the previous best hybrid detector HEDNet by 2.6\% mAP while being 2.1$\times$ faster, and yielded 2.1\% mAP gains over the previous best sparse detector FSDv2 while being 1.3$\times$ faster. The code will be available at https://github.com/zhanggang001/HEDNet.

This paper presents a novel aerial-to-ground feature aggregation strategy, tailored for the task of cross-view image-based geo-localization. Conventional vision-based methods heavily rely on matching ground-view image features with a pre-recorded image database, often through establishing planar homography correspondences via a planar ground assumption. As such, they tend to ignore features that are off-ground and not suited for handling visual occlusions, leading to unreliable localization in challenging scenarios. We propose a Top-to-Ground Aggregation (T2GA) module that capitalizes aerial orthographic views to aggregate features down to the ground level, leveraging reliable off-ground information to improve feature alignment. Furthermore, we introduce a Cycle Domain Adaptation (CycDA) loss that ensures feature extraction robustness across domain changes. Additionally, an Equidistant Re-projection (ERP) loss is introduced to equalize the impact of all keypoints on orientation error, leading to a more extended distribution of keypoints which benefits orientation estimation. On both KITTI and Ford Multi-AV datasets, our method consistently achieves the lowest mean longitudinal and lateral translations across different settings and obtains the smallest orientation error when the initial pose is less accurate, a more challenging setting. Further, it can complete an entire route through continual vehicle pose estimation with initial vehicle pose given only at the starting point.

Improving the detection of distant 3d objects is an important yet challenging task. For camera-based 3D perception, the annotation of 3d bounding relies heavily on LiDAR for accurate depth information. As such, the distance of annotation is often limited due to the sparsity of LiDAR points on distant objects, which hampers the capability of existing detectors for long-range scenarios. We address this challenge by considering only 2D box supervision for distant objects since they are easy to annotate. We propose LR3D, a framework that learns to recover the missing depth of distant objects. LR3D adopts an implicit projection head to learn the generation of mapping between 2D boxes and depth using the 3D supervision on close objects. This mapping allows the depth estimation of distant objects conditioned on their 2D boxes, making long-range 3D detection with 2D supervision feasible. Experiments show that without distant 3D annotations, LR3D allows camera-based methods to detect distant objects (over 200m) with comparable accuracy to full 3D supervision. Our framework is general, and could widely benefit 3D detection methods to a large extent.

End-to-end autonomous driving recently emerged as a promising research direction to target autonomy from a full-stack perspective. Along this line, many of the latest works follow an open-loop evaluation setting on nuScenes to study the planning behavior. In this paper, we delve deeper into the problem by conducting thorough analyses and demystifying more devils in the details. We initially observed that the nuScenes dataset, characterized by relatively simple driving scenarios, leads to an under-utilization of perception information in end-to-end models incorporating ego status, such as the ego vehicle's velocity. These models tend to rely predominantly on the ego vehicle's status for future path planning. Beyond the limitations of the dataset, we also note that current metrics do not comprehensively assess the planning quality, leading to potentially biased conclusions drawn from existing benchmarks. To address this issue, we introduce a new metric to evaluate whether the predicted trajectories adhere to the road. We further propose a simple baseline able to achieve competitive results without relying on perception annotations. Given the current limitations on the benchmark and metrics, we suggest the community reassess relevant prevailing research and be cautious about whether the continued pursuit of state-of-the-art would yield convincing and universal conclusions. Code and models are available at https://github.com/NVlabs/BEV-Planner.

For safe motion planning in real-world, autonomous vehicles require behavior prediction models that are reliable and robust to distribution shifts. The recent studies suggest that the existing learning-based trajectory prediction models do not posses such characteristics and are susceptible to small perturbations that are not present in the training data, largely due to overfitting to spurious correlations while learning. In this paper, we propose a causal disentanglement representation learning approach aiming to separate invariant (causal) and variant (spurious) features for more robust learning. Our method benefits from a novel intervention mechanism in the latent space that estimates potential distribution shifts resulted from spurious correlations using uncertain feature statistics, hence, maintaining the realism of interventions. To facilitate learning, we propose a novel invariance objective based on the variances of the distributions over uncertain statistics to induce the model to focus on invariant representations during training.We conduct extensive experiments on two large-scale autonomous driving datasets and show that besides achieving state-of-the-art performance, our method can significantly improve prediction robustness to various distribution shifts in driving scenes. We further conduct ablative studies to evaluate the design choices in our proposed framework.

In autonomous driving, behavior prediction is fundamental for safe motion planning, hence the security and robustness of prediction models against adversarial attacks are of paramount importance. We propose a novel adversarial backdoor attack against trajectory prediction models as a means of studying their potential vulnerabilities. Our attack affects the victim at training time via naturalistic, hence stealthy, poisoned samples crafted using a novel two-step approach. First, the triggers are crafted by perturbing the trajectory of attacking vehicle and then disguised by transforming the scene using a bi-level optimization technique. The proposed attack does not depend on a particular model architecture and operates in a black-box manner, thus can be effective without any knowledge of the victim model. We conduct extensive empirical studies using state-of-the-art prediction models on two benchmark datasets using metrics customized for trajectory prediction. We show that the proposed attack is highly effective, as it can significantly hinder the performance of prediction models, unnoticeable by the victims, and efficient as it forces the victim to generate malicious behavior even under constrained conditions. Via ablative studies, we analyze the impact of different attack design choices followed by an evaluation of existing defence mechanisms against the proposed attack.

Neural radiance fields (NeRFs) have gained popularity in the autonomous driving (AD) community. Recent methods show NeRFs' potential for closed-loop simulation, enabling testing of AD systems, and as an advanced training data augmentation technique. However, existing methods often require long training times, dense semantic supervision, or lack generalizability. This, in turn, hinders the application of NeRFs for AD at scale. In this paper, we propose NeuRAD, a robust novel view synthesis method tailored to dynamic AD data. Our method features simple network design, extensive sensor modeling for both camera and lidar --- including rolling shutter, beam divergence and ray dropping --- and is applicable to multiple datasets out of the box. We verify its performance on five popular AD datasets, achieving state-of-the-art performance across the board. To encourage further development, we openly release the NeuRAD source code.

Bird's eye view (BEV) representation has emerged as a dominant solution for describing 3D space in autonomous driving scenarios. However, objects in the BEV representation typically exhibit small sizes, and the associated point cloud context is inherently sparse, which leads to great challenges for reliable 3D perception. In this paper, we propose IS-Fusion, an innovative multimodal Fusion framework that jointly captures the Instance- and Scene-level contextual information. IS-Fusion essentially differs from existing approaches that only focus on the BEV scene-level fusion by explicitly incorporating instance-level multimodal information, thus facilitating the instance-centric tasks like 3D object detection. It comprises a Hierarchical Scene Fusion (HSF) module and an Instance-Guided Fusion (IGF) module. HSF applies Point-to-Grid and Grid-to-Region transformers to capture the multimodal scene context at different granularities. This leads to enriched scene-level features that are later used to generate high-quality instance features. Next, IGF explores the relationships between these instances, and aggregates the local multimodal context for each instance. These instances then serve as guidance to enhance the scene feature and yield an instance-aware BEV representation. On the challenging nuScenes benchmark, IS-Fusion outperforms all the published multimodal works to date. Notably, it achieves 72.8% mAP on the nuScenes validation set, outperforming prior art like BEVFusion by 4.3% mAP. Code shall be released.

Autonomous systems need to process large-scale, sparse, and irregular point clouds with limited compute resources. Consequently, it is essential to develop LiDAR perception methods that are both efficient and effective. Although naive-ly enlarging 3D kernel size can enhance performance, it will also lead to a cubically-increasing overhead. Therefore, it is crucial to develop streamlined 3D large kernel designs that eliminate redundant weights and work effectively with larger kernels. In this paper, we propose an efficient and effective Large Sparse Kernel 3D Neural Network (LSK3DNet) that leverages dynamic pruning to amplify the 3D kernel size. Our method comprises two core components: Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS). SDS dynamically prunes and regrows volumetric weights from the beginning to learn a large sparse 3D kernel. It not only boosts performance but also significantly reduces model size and computational cost. Moreover, CWS selects the most important channels for 3D convolution during training and subsequently prunes the redundant channels to accelerate inference for 3D vision tasks. We demonstrate the effectiveness of LSK3DNet on three benchmark datasets and five tracks compared with classical models and large kernel designs. Notably, LSK3DNet achieves the state-of-the-art performance on SemanticKITTI (i.e., 75.6% on single-scan and 63.4% on multi-scan), with roughly 40% model size reduction and 60% computing operations reduction compared to the naive large 3D kernel model.

Three-dimensional object detection is one of the key tasks in autonomous driving. To reduce costs in practice, low-cost multi-view cameras for 3D object detection are proposed to replace the expansive LiDAR sensors. However, relying solely on cameras is difficult to achieve highly accurate and robust 3D object detection. An effective solution to this issue is combining multi-view cameras with the economical millimeter-wave radar sensor to achieve more reliable multi-modal 3D object detection. In this paper, we introduce RCBEVDet, a radar-camera fusion 3D object detection method in the bird's eye view (BEV). Specifically, we first design RadarBEVNet for radar BEV feature extraction. RadarBEVNet consists of a dual-stream radar backbone and a Radar Cross-Section (RCS) aware BEV encoder. In the dual-stream radar backbone, a point-based encoder and a transformer-based encoder are proposed to extract radar features, with an injection and extraction module to facilitate communication between the two encoders. The RCS-aware BEV encoder takes RCS as the object size prior to scattering the point feature in BEV. Besides, we present the Cross-Attention Multi-layer Fusion module to automatically align the multi-modal BEV feature from radar and camera with the deformable attention mechanism, and then fuse the feature with channel and spatial fusion layers. Experimental results show that RCBEVDet achieves new state-of-the-art radar-camera fusion results on nuScenes and view-of-delft (VoD) 3D object detection benchmarks. Furthermore, RCBEVDet achieves better 3D detection results than all real-time camera-only and radar-camera 3D object detectors with a faster inference speed at 21~28 FPS. The source code and models will be released to the public.

Recent temporal LiDAR-based 3D object detectors achieve promising performance based on the two-stage proposal-based approach. They generate 3D box candidates from the first-stage dense detector, followed by different temporal aggregation methods. However, these approaches require per-frame objects or whole point clouds, posing challenges related to memory bank utilization. Moreover, point clouds and trajectory features are combined solely based on concatenation, which may neglect effective interactions between them. In this paper, we propose a point-trajectory transformer with long short-term memory for efficient temporal 3D object detection. To this end, we only utilize point clouds of current-frame objects and their historical trajectories as input to minimize the memory bank storage requirement. Furthermore, we introduce modules to encode trajectory features, focusing on long short-term and future-aware perspectives, and then effectively aggregate them with point cloud features. We conduct extensive experiments on the large-scale Waymo dataset to demonstrate that our approach performs well against state-of-the-art methods. The source codes and trained models will be made publicly available.

Adapting driving behavior to new environments, customs, and laws is a long-standing problem in autonomous driving, precluding the widespread deployment of autonomous vehicles (AVs). In this paper, we present LLaDA, a simple yet powerful tool that enables human drivers and autonomous vehicles alike to drive everywhere by adapting their tasks and motion plans to traffic rules in new locations. LLaDA achieves this by leveraging the impressive zero-shot generalizability of large language models (LLMs) in interpreting the traffic rules in the local driver handbook. Through an extensive user study, we show that LLaDA's instructions are useful in disambiguating in-the-wild unexpected situations. We also demonstrate LLaDA's ability to adapt AV motion planning policies in real-world datasets; LLaDA outperforms baseline planning approaches on all our metrics. Please check our website for more details: https://boyiliee.github.io/llada/.

We tackle the problem of 3D point cloud localization based on a few natural linguistic descriptions and introduce a novel neural network, Text2Loc, that fully interprets the semantic relationship between points and text. Text2Loc follows a coarse-to-fine localization pipeline: text-submap global place recognition, followed by fine localization. In global place recognition, relational dynamics among each textual hint are captured in a hierarchical transformer with max-pooling (HTM), whereas a balance between positive and negative pairs is maintained using text-submap contrastive learning. Moreover, we propose a novel matching-free fine localization method to further refine the location predictions, which completely removes the need for complicated text-instance matching and is lighter, faster, and more accurate than previous methods. Extensive experiments show that Text2Loc improves the localization accuracy by up to 2 times over the state-of-the-art on the KITTI360Pose dataset. Our project page is publicly available here.

Recently, unsupervised 3D object detection for autonomous driving gained great attention. The prevalent approaches primarily follow cluster-based pseudo-label generation and iterative self-training processes. However, the challenge arises due to the sparsity of LiDAR scans, which leads to pseudo-labels with erroneous size and position, resulting in subpar detection performance. To tackle this problem, this paper introduces a Commonsense Prototype-based Detector, termed CPD, for unsupervised 3D object detection. CPD first constructs Commonsense Prototype (CProto) characterized by high-quality bounding box and dense points, based on commonsense intuition. Subsequently, CPD refines the low-quality pseudo-labels by leveraging the size prior from CProto. Furthermore, CPD enhances the detection accuracy of sparsely scanned objects by the geometric knowledge from CProto. CPD outperforms state-of-the-art unsupervised 3D detectors on Waymo Open Dataset (WOD), PandaSet, and KITTI datasets by a large margin. For the first time, our unsupervised CPD surpasses some weakly supervised approaches. Besides, by training CPD on WOD and testing on KITTI, CPD attains 90.85% and 81.01% 3D Average Precision on easy and moderate car classes, respectively. These achievements position CPD in close proximity to fully supervised detectors, highlighting the significance of our method. The code will be made publicly available.

This work proposes the first online asymmetric semi-supervised framework, namely A-Teacher, for LiDAR-based 3D object detection. Our motivation stems from the observation that 1) existing symmetric teacher-student methods for semi-supervised 3D object detection have characterized simplicity, but impede the distillation performance between teacher and student because of the demand for an identical model structure and input data format. 2) The offline asymmetric methods with a complex teacher model, constructed differently, can generate more precise pseudo-labels, but is challenging to jointly optimize the teacher and student model. Consequently, in this paper, we devise a different path from the conventional paradigm, which can harness the capacity of a strong teacher while preserving the advantages of online teacher model updates. The essence is the proposed attention-based refinement model that can be seamlessly integrated into a vanilla teacher. The refinement model works in the divide-and-conquer manner that respectively handles three challenging scenarios including 1) objects detected in the current timestamp but with suboptimal box quality, 2) objects are missed in the current timestamp but are detected in past or future frames, 3) objects are neglected in all frames. It is worth noting that even while tackling these complex cases, our model retains the efficiency of the online teacher-student semi-supervised framework. Experimental results on Waymo show that our method outperforms previous state-of-the-art HSSDA for 4.7 on mAP (L1) while consuming fewer training resources.

Many existing motion prediction approaches rely on symbolic perception outputs to generate agent trajectories, such as bounding boxes, road graph information and traffic lights. This symbolic representation is a high-level abstraction of the real world, which may render the motion prediction model vulnerable to perception errors (e.g., failures in detecting open-vocabulary obstacles) while missing salient information from the scene context (e.g., poor road conditions). An alternative paradigm is end-to-end learning from raw sensors. However, this approach suffers from the lack of interpretability and requires significantly more training resources. In this work, we propose tokenizing the visual world into a compact set of scene elements and then leveraging pre-trained image foundation models and LiDAR neural networks to encode all the scene elements in an open-vocabulary manner. The image foundation model enables our scene tokens to encode the general knowledge of the open world while the LiDAR neural network encodes geometry information. Our proposed representation can efficiently encode the multi-frame multi-modality observations with a few hundred tokens and is compatible with most transformer-based architectures. To evaluate our method, we have augmented Waymo Open Motion Dataset with camera embeddings. Experiments over Waymo Open Motion Dataset show that our approach leads to significant performance improvements over the state-of-the-art.

While behavior cloning has recently emerged as a highly successful paradigm for autonomous driving, humans rarely learn to perform complex tasks, such as driving, via imitation or behavior cloning alone. In contrast, learning in humans often involves additional detailed guidance throughout the interactive learning process, i.e., where feedback, often via language, provides detailed information as to which part of their trial was performed incorrectly or suboptimally and why. Motivated by this observation, we introduce an efficient feedback-based framework for improving behavior-cloning-based training of sensorimotor driving agents. Our key insight is to leverage recent advances in Large Language Models (LLMs) to provide corrective fine-grained feedback regarding the underlying reason behind driving prediction failures. Moreover, our introduced network architecture is efficient, enabling the first sensorimotor end-to-end training and evaluation of LLM-based driving models. The resulting agent achieves state-of-the-art performance in open-loop evaluation on nuScenes, outperforming prior state-of-the-art by over 5.4% and 14.3% in accuracy and collision rate, respectively. In CARLA, our camera-based agent improves by 16.6% in driving score over prior LIDAR-based approaches.

The perception of autonomous vehicles using radars has attracted increased research interest due its ability to operate in fog and bad weather. However, training radar models is hindered by the cost and difficulty of annotating large-scale radar data. To overcome this bottleneck, we propose a self-supervised learning framework to leverage the large amount of unlabeled radar data to pre-train radar-only embeddings for self-driving perception tasks. The proposed method combines radar-to-radar and radar-to-vision contrastive losses to learn a general representation from unlabeled radar heatmaps paired with their corresponding camera images. When used for downstream object detection, we demonstrate that the proposed self-supervision framework can improve the accuracy of state-of-the-art supervised baselines by $5.8\%$ in mAP. Code is available at https://github.com/yiduohao/Radical.

Perceiving the world and forecasting its future state is a critical task for self-driving. Supervised approaches leverage annotated object labels to learn a model of the world --- traditionally with object detections and trajectory predictions, or temporal bird's-eye-view (BEV) occupancy fields. However, these annotations are expensive and typically limited to a set of predefined categories that do not cover everything we might encounter on the road. Instead, we learn to perceive and forecast a continuous 4D (spatio-temporal) occupancy field with self-supervision from LiDAR data. This unsupervised world model can be easily and effectively transferred to downstream tasks. We tackle point cloud forecasting by adding a lightweight learned renderer and achieve state-of-the-art performance in Argoverse 2, nuScenes, and KITTI. To further showcase its transferability, we fine-tune our model for BEV semantic occupancy forecasting and show that it outperforms the fully supervised state-of-the-art, especially when labeled data is scarce. Finally, when compared to prior state-of-the-art on spatio-temporal geometric occupancy prediction, our 4D world model achieves a much higher recall of objects from classes relevant to self-driving.

Conventional radar feature extraction faces limitations due to low spatial resolution, noise, multipath reflection, the presence of ghost targets, and motion blur. Such limitations can be exacerbated by nonlinear object motion, particularly from an ego-centric viewpoint. It becomes evident that to address these challenges, the key lies in exploiting temporal feature relation over an extended horizon and enforcing spatial motion consistence for effective association. To this end, this paper proposes SIRA (Scalable Inter-frame Relation and Association) with two designs. First, inspired by Swin Transformer, we introduce extended temporal relation, generalizing the existing temporal relation layer from two consecutive frames to multiple inter-frames with temporally regrouped window attention for scalability. Second, we propose motion consistency track with the concept of a pseudo-tracklet generated from observational data for better trajectory prediction and subsequent object association. Our approach achieves $58.11$ mAP$@0.5$ for oriented object detection and $47.79$ MOTA for multiple object tracking on the Radiate dataset, surpassing previous state-of-the-art by a margin of $+4.11$ mAP$@0.5$ and $+9.94$ MOTA, respectively.

Vision-based perception for autonomous driving requires an explicit modeling of a 3D space, where 2D latent representations are mapped and subsequent 3D operators are applied. However, operating on dense latent spaces introduces a cubic time and space complexity, which limits scalability in terms of perception range or spatial resolution. Existing approaches compress the dense representation using projections like Bird's Eye View (BEV) or Tri-Perspective View (TPV). Although efficient, these projections result in information loss, especially for tasks like semantic occupancy prediction. To address this, we propose SparseOcc, an efficient occupancy network inspired by sparse point cloud processing. It utilizes a lossless sparse latent representation with three key innovations. Firstly, a 3D sparse diffuser performs latent completion using spatially decomposed 3D sparse convolutional kernels. Secondly, a feature pyramid and sparse interpolation enhance scales with information from others. Finally, the transformer head is redesigned as a sparse variant. SparseOcc achieves a remarkable 74.9% reduction on FLOPs over the dense baseline. Interestingly, it also improves accuracy, from 12.8% to 14.1% mIOU, which in part can be attributed to the sparse representation's ability to avoid hallucinations on empty voxels.

Absolute pose regression (APR) estimates global pose in an end-to-end manner, achieving impressive results in learn-based LiDAR localization. However, compared to the top-performing methods reliant on 3D-3D correspondence matching, APR’s accuracy still has room for improvement. We recognize APR’s lack of geometrically robust features learning and iterative denoising process leads to suboptimal results. In this paper, we propose DiffLoc, a novel framework that formulates LiDAR localization as a conditional generation of poses. First, we propose to utilize the foundation model and static-object-aware pool to learn geometrically robust features. Second, we creatively incorporate the iterative denoising process into APR via a diffusion model conditioned on the learned geometrically robust features. In addition, due to the unique nature of diffusion models, we propose to adapt our models to two additional applications: (1) using multiple inferences to evaluate pose uncertainty, and (2) seamlessly introducing geometric constraints on denoising steps to improve prediction accuracy. Extensive experiments conducted on the Oxford Radar RobotCar and NCLT datasets demonstrate that DiffLoc outperforms better than the state-of-the-art methods. Especially on the NCLT dataset, we achieve 35% and 34.7% improvement on position and orientation accuracy, respectively. Our code will be released upon acceptance.

This paper addresses the critical challenges of sparsity and occlusion in LiDAR-based 3D object detection. Current methods often rely on supplementary modules or specific architectural designs, potentially limiting their applicability to new and evolving architectures. To our knowledge, we are the first to propose a versatile technique that seamlessly integrates into any existing framework for 3D Object Detection, marking the first instance of Weak-to-Strong generalization in 3D computer vision. We introduce a novel framework, X-Ray Distillation with Object-Complete Frames, suitable for both supervised and semi-supervised settings, that leverages the temporal aspect of point cloud sequences. This method extracts crucial information from both previous and subsequent LiDAR frames, creating Object-Complete frames that represent objects from multiple viewpoints, thus addressing occlusion and sparsity. Given the limitation of not being able to generate Object-Complete frames during online inference, we utilize Knowledge Distillation within a Teacher-Student framework. This technique encourages the strong Student model to emulate the behavior of the weaker Teacher, which processes simple and informative Object-Complete frames, effectively offering a comprehensive view of objects as if seen through X-ray vision. Our proposed methods surpass state-of-the-art in semi-supervised learning by 1-1.5 mAP and enhance the performance of five established supervised models by 1-2 mAP on standard autonomous driving datasets, even with default hyperparameters. Code for Object-Complete frames is available here: https://github.com/sakharok13/X-Ray-Teacher-Patching-Tools.

Trajectory prediction is a challenging problem that requires considering interactions among multiple actors and the surrounding environment.While data-driven approaches have been used to address this complex problem, they suffer from unreliable predictions under distribution shift during test time.Accordingly, several online learning methods have been proposed using regression loss from the ground truth of observed data leveraging the auto-labeling nature of trajectory prediction task.We tackle two issues of the previous methods.First, we tackle that online adaptation is done with a few GT samples in a delayed time stamp.It makes underfitting or overfitting, so previous works only optimized the last layers of motion decoder.We employ the masked autoencoder (MAE) for representation learning to encourage complex interaction modeling in shifted test distribution for updating deeper layers.Second, we emphasize that driving data comes sequentially during test time, unlike random shuffles during training time.Therefore, we can access a specific past motion pattern for each actor instance.To address this, we propose an actor-specific token memory, enabling the learning of personalized motion pattern during test-time training.Our proposed method has been validated across various challenging cross-dataset distribution shift scenarios including nuScenes, Lyft, Waymo, and Interaction. Our method surpasses the performance of existing state-of-the-art online learning methods in terms of both prediction accuracy and computational efficiency.

Scene simulation in autonomous driving has gained significant attention because of its huge potential for generating customized data. However, existing editable scene simulation approaches face limitations in terms of user interaction efficiency, multi-camera photo-realistic rendering and external digital assets integration. To address these challenges, this paper introduces ChatSim, the first system that enables editable photo-realistic 3D driving scene simulations via natural language commands with external digital assets. To enable editing with high command flexibility, ChatSim leverages a large language model (LLM) agent collaboration framework. To generate photo-realistic outcomes, ChatSim employs a novel multi-camera neural radiance field method. Furthermore, to unleash the potential of extensive high-quality digital assets, ChatSim employs a novel multi-camera lighting estimation method to achieve scene-consistent assets' rendering. Our experiments on Waymo Open Dataset demonstrate that ChatSim can handle complex language commands and generate corresponding photo-realistic scene videos.

We present a highly scalable self-training framework for incrementally adapting vision-based end-to-end autonomous driving policies in a semi-supervised manner, i.e., over a continual stream of incoming video data. To facilitate large-scale model training (e.g., open web or unlabeled data), we do not assume access to ground-truth labels and instead estimate pseudo-label policy targets for each video. Our framework comprises three key components: knowledge distillation, a sample purification module, and an exploration and knowledge retention mechanism. First, given sequential image frames, we pseudo-label the data and estimate uncertainty using an ensemble of inverse dynamics models. The uncertainty is used to select the most informative samples to add to an experience replay buffer. We specifically select high-uncertainty pseudo-labels to facilitate the exploration and learning of new and diverse driving skills. However, in contrast to prior work in continual learning that assumes ground-truth labeled samples, the uncertain pseudo-labels can introduce significant noise. Thus, we also pair the exploration with a label refinement module, which makes use of consistency constraints to re-label the noisy exploratory samples and effectively learn from diverse data. Trained as a complete never-ending learning system, we demonstrate state-of-the-art performance on training from domain-changing data as well as millions of images from the open web.

End-to-end motion planning models equipped with deep neural networks have shown great potential for enabling full autonomous driving. However, the oversized neural networks render them impractical for deployment on resource-constrained systems, which unavoidably requires more computational time and resources during reference. To handle this, knowledge distillation offers a promising approach that compresses models by enabling a smaller student model to learn from a larger teacher model. Nevertheless, how to apply knowledge distillation to compress motion planners has not been explored so far. In this paper, we propose PlanKD, the first knowledge distillation framework tailored for compressing end-to-end motion planners. First, considering that driving scenes are inherently complex, often containing planning-irrelevant or even noisy information, transferring such information is not beneficial for the student planner. Thus, we design an information bottleneck based strategy to only distill planning-relevant information, rather than transfer all information indiscriminately. Second, different waypoints in an output planned trajectory may hold varying degrees of importance for motion planning, where a slight deviation in certain crucial waypoints might lead to a collision. Therefore, we devise a safety-aware waypoint-attentive distillation module that assigns adaptive weights to different waypoints based on the importance, to encourage the student to accurately mimic more crucial waypoints, thereby improving overall safety. Experiments demonstrate that our PlanKD can boost the performance of smaller planners by a large margin, and significantly reduce their reference time.

Scene flow estimation, which aims to predict per-point 3D displacements of dynamic scenes, is a fundamental task in the computer vision field. However, previous works commonly suffer from unreliable correlation caused by locally constrained searching ranges, and struggle with accumulated inaccuracy arising from the coarse-to-fine structure. To alleviate these problems, we propose a novel uncertainty-aware scene flow estimation network (DifFlow3D) with the diffusion probabilistic model. Iterative diffusion-based refinement is designed to enhance the correlation robustness and resilience to challenging cases, e.g. dynamics, noisy inputs, repetitive patterns, etc. To restrain the generation diversity, three key flow-related features are leveraged as conditions in our diffusion model. Furthermore, we also develop an uncertainty estimation module within diffusion to evaluate the reliability of estimated scene flow. Our DifFlow3D achieves state-of-the-art performance, with 24.0% and 29.1% EPE3D reduction respectively on FlyingThings3D and KITTI 2015 datasets. Notably, our method achieves an unprecedented millimeter-level accuracy (0.0078m in EPE3D) on the KITTI dataset. Additionally, our diffusion-based refinement paradigm can be readily integrated as a plug-and-play module into existing scene flow networks, significantly increasing their estimation accuracy. Codes are released at https://github.com/IRMVLab/DifFlow3D.

Despite significant recent progress in the field of autonomous driving, modern methods still struggle and can incur serious accidents when encountering long-tail unforeseen events and challenging urban scenarios. On the one hand, large language models (LLM) have shown impressive reasoning capabilities that approach “Artificial General Intelligence”. On the other hand, previous autonomous driving methods tend to rely on limited-format inputs ($\textit{e.g.}$, sensor data and navigation waypoints), restricting the vehicle's ability to understand language information and interact with humans. To this end, this paper introduces LMDrive, a novel language-guided, end-to-end, closed-loop autonomous driving framework. LMDrive uniquely processes and integrates multi-modal sensor data with natural language instructions, enabling interaction with humans and navigation software in realistic instructional settings. To facilitate further research in language-based closed-loop autonomous driving, we also publicly release the corresponding dataset which includes approximately 64K instruction-following data clips, and the LangAuto benchmark that tests the system's ability to handle complex instructions and challenging driving scenarios. Extensive closed-loop experiments are conducted to demonstrate LMDrive's effectiveness. To the best of our knowledge, we're the very first work to leverage LLMs for closed-loop end-to-end autonomous driving. More demo videos and codes can be found at our webpage: https://github.com/opendilab/LMDrive

In rapidly-evolving domains such as autonomous driving, the use of multiple sensors with different modalities is crucial to ensure high operational precision and stability. To correctly exploit the provided information by each sensor in a single common frame, it is essential for these sensors to be accurately calibrated. In this paper, we leverage the ability of Neural Radiance Fields (NeRF) to represent different sensors modalities in a common volumetric representation to achieve robust and accurate spatio-temporal sensor calibration.By designing a partitioning approach based on the visible part of the scene for each sensor, we formulate the calibration problem using only the overlapping areas. This strategy results in a more robust and accurate calibration that is less prone to failure. We demonstrate that our approach works on outdoor urban scenes by validating it on multiple established driving datasets. Results show that our method is able to get better accuracy and robustness compared to existing methods.

Autonomous driving (AD) has made significant strides in recent years. However, existing frameworks struggle to interpret and execute spontaneous user instructions, such as "overtake the car ahead." Large Language Models (LLMs) have demonstrated impressive reasoning capabilities showing potential to bridge this gap. In this paper, we present LaMPilot, a novel framework that integrates LLMs into AD systems, enabling them to follow user instructions by generating code that leverages established functional primitives. We also introduce LaMPilot-Bench, the first benchmark dataset specifically designed to quantitatively evaluate the efficacy of language model programs in AD. Adopting the LaMPilot framework, we conduct extensive experiments to assess the performance of off-the-shelf LLMs on LaMPilot-Bench. Our results demonstrate the potential of LLMs in handling diverse driving scenarios and following user instructions in driving. To facilitate further research in this area, we release our code and data at GitHub.com/PurdueDigitalTwin/LaMPilot.

Sparse LiDAR point clouds cause severe loss of detail of static structures and reduce the density of static points available for navigation. Reduced density can be detrimental to navigation under several scenarios. We observe that despite high sparsity, in most cases, the global topology of LiDAR outlining the static structures can be inferred. We utilize this property to obtain a backbone skeleton of a static LiDAR scan in the form of a single connected component that is a proxy to its global topology. We utilize the backbone to augment new points along static structures to overcome sparsity. Newly introduced points could correspond to existing static structures or to static points that were earlier obstructed by dynamic objects. To the best of our knowledge, we are the first to use this strategy for sparse LiDAR point clouds. Existing solutions close to our approach fail to identify and preserve the global static LiDAR topology and generate sub-optimal points. We propose GLiDR, a Graph Generative network that is topologically regularized using 0-dimensional Persistent Homology (PH) constraints. This enables GLiDR to introduce newer static points along a topologically consistent global static LiDAR backbone. GLiDR generates precise static points using 32× sparser dynamic scans and performs better than the baselines across three datasets. The newly introduced static points allow GLiDR to outperform LiDAR-based navigation using SLAM in several settings. GLiDR generate a valuable byproduct - an accurate binary segmentation mask of static and dynamic objects that is helpful for navigation planning and safety in constrained environment.

Detecting objects in 3D under various (normal and adverse) weather conditions is essential for safe autonomous driving systems. Recent approaches have focused on employing weather-insensitive 4D radar sensors and leveraging them with other modalities, such as LiDAR. However, they fuse multi-modal information without considering the sensor characteristics and weather conditions, and lose some height information which could be useful for localizing 3D objects. In this paper, we propose a novel framework for robust LiDAR and 4D radar-based 3D object detection. Specifically, we propose a 3D-LRF module that considers the distinct patterns they exhibit in 3D space (e.g., precise 3D mapping of LiDAR and wide-range, weather-insensitive measurement of 4D radar) and extract fusion features based on their 3D spatial relationship. Then, our weather-conditional radar-flow gating network modulates the information flow of fusion features depending on weather conditions, and obtains enhanced feature that effectively incorporates the strength of two domains under various weather conditions. The extensive experiments demonstrate that our model achieves SoTA performance for 3D object detection under various weather conditions.

Learning 3D scene flow from LiDAR point clouds presents significant difficulties, including poor generalization from synthetic datasets to real scenes, scarcity of real-world 3D labels, and poor performance on real sparse LiDAR point clouds. We present a novel approach from the perspective of auto-labelling, aiming to generate a large number of 3D scene flow pseudo labels for real-world LiDAR point clouds. Specifically, we employ the assumption of rigid body motion to simulate potential object-level rigid movements in autonomous driving scenarios. By updating different motion attributes for multiple anchor boxes, the rigid motion decomposition is obtained for the whole scene. Furthermore, we developed a novel 3D scene flow data augmentation method for global and local motion. By perfectly synthesizing target point clouds based on augmented motion parameters, we easily obtain lots of 3D scene flow labels in point clouds highly consistent with real scenarios. On multiple real-world datasets including LiDAR KITTI, nuScenes, and Argoverse, our method outperforms all previous supervised and unsupervised methods without requiring manual labelling. Impressively, our method achieves a tenfold reduction in EPE3D metric on the LiDAR KITTI dataset, reducing it from $0.190m$ to a mere $0.008m$ error.

Many query-based approaches for 3D Multi-Object Tracking (MOT) adopt the tracking-by-attention paradigm, utilizing track queries for identity-consistent detection and object queries for identity-agnostic track spawning. Tracking-by-attention, however, entangles detection and tracking queries in one embedding for both the detection and tracking task, which is sub-optimal. Other approaches resemble the tracking-by-detection paradigm, detecting objects using decoupled track and detection queries followed by a subsequent association. These methods, however, do not leverage synergies between the detection and association task. Combining the strengths of both paradigms, we introduce ADA-Track, a novel end-to-end framework for 3D MOT from multi-view cameras. We introduce a learnable data association module based on edge-augmented cross-attention, leveraging appearance and geometric features. Furthermore, we integrate this association module into the decoder layer of a DETR-based 3D detector, enabling simultaneous DETR-like query-to-image cross-attention for detection and query-to-query cross-attention for data association. By stacking these decoder layers, queries are refined for the detection and association task alternately, effectively harnessing the task dependencies. We evaluate our method on the nuScenes dataset and demonstrate the advantage of our approach compared to the two previous paradigms. Code is available at https://github.com/dsx0511/ADA-Track.

Bird's-eye View (BeV) representations have emerged as the de-facto shared space in driving applications, offering a unified space for sensor data fusion and supporting various downstream tasks. However, conventional models use grids with fixed resolution and range and face computational inefficiencies due to the uniform allocation of resources across all cells. To address this, we propose PointBeV, a novel sparse BeV segmentation model operating on sparse BeV cells instead of dense grids. This approach offers precise control over memory usage, enabling the use of long temporal contexts and accommodating memory-constrained platforms. PointBeV employs an efficient two-pass strategy for training, enabling focused computation on regions of interest. At inference time, it can be used with various memory/performance trade-offs and flexibly adjusts to new specific use cases. PointBeV achieves state-of-the-art results on the nuScenes dataset for vehicle, pedestrian, and lane segmentation, showcasing superior performance in static and temporal settings despite being trained solely with sparse signals. We release our code with two new efficient modules used in the architecture: Sparse Feature Pulling, designed for the effective extraction of features from images to BeV, and Submanifold Attention, which enables efficient temporal modeling. The code is available at https://github.com/valeoai/PointBeV.

Vision-centric perception systems for autonomous driving have gained considerable attention recently due to their cost-effectiveness and scalability, especially compared to LiDAR-based systems. However, these systems often struggle in low-light conditions, potentially compromising their performance and safety. To address this, our paper introduces LightDiff, a domain-tailored framework designed to enhance the low-light image quality for autonomous driving applications. Specifically, we employ a multi-condition controlled diffusion model. LightDiff works without any human-collected paired data, leveraging a dynamic data degradation process instead. It incorporates a novel multi-condition adapter that adaptively controls the input weights from different modalities, including depth maps, RGB images, and text captions, to effectively illuminate dark scenes while maintaining context consistency. Furthermore, to align the enhanced images with the detection model's knowledge, LightDiff employs perception-specific scores as rewards to guide the diffusion training process through reinforcement learning. Extensive experiments on the nuScenes datasets demonstrate that LightDiff can significantly improve the performance of several state-of-the-art 3D detectors in night-time conditions while achieving high visual quality scores, highlighting its potential to safeguard autonomous driving.

Autonomous driving stands as a pivotal domain in computer vision, shaping the future of transportation. Within this paradigm, the backbone of the system plays a crucial role in interpreting the complex environment. However, a notable challenge has been the loss of clear supervision when it comes to Bird's Eye View elements. To address this limitation, we introduce CLIP-BEVFormer, a novel approach that leverages the power of contrastive learning techniques to enhance the multi-view image-derived BEV backbones with ground truth information flow. We conduct extensive experiments on the challenging nuScenes dataset and showcase significant and consistent improvements over the SOTA. Specifically, CLIP-BEVFormer achieves an impressive 8.5\% and 9.2\% enhancement in terms of NDS and mAP, respectively, over the previous best BEV model on the 3D object detection task.

To be updatedTrajectory prediction plays a vital role in various applications, including autonomous driving, robotics, and scene understanding. Existing approaches mainly focus on developing compact neural networks to increase prediction precision on public datasets, typically employing a standardized input duration. However, a notable issue arises when these models are evaluated with varying observation lengths, leading to a significant performance drop, a phenomenon we term the Observation Length Shift. To tackle this issue, we introduce a general and effective framework, the FlexiLength Network (FLN), to enhance the robustness of existing trajectory prediction techniques against varying observation periods. Specifically, FLN integrates trajectory data with diverse observation lengths, incorporates FlexiLength Calibration (FLC) to acquire temporal invariant representations, and employs FlexiLength Adaptation (FLA) to further refine these representations for more accurate future trajectory predictions. Comprehensive experiments on multiple datasets, \ie, ETH/UCY, nuScenes, and Argoverse 1, demonstrate the effectiveness and flexibility of our proposed FLN framework.

In the context of autonomous driving, the significance of effective feature learning is widely acknowledged. While conventional 3D self-supervised pre-training methods have shown widespread success, most methods follow the ideas originally designed for 2D images. In this paper, we present UniPAD, a novel self-supervised learning paradigm applying 3D volumetric differentiable rendering. UniPAD implicitly encodes 3D space, facilitating the reconstruction of continuous 3D shape structures and the intricate appearance characteristics of their 2D projections. The flexibility of our method enables seamless integration into both 2D and 3D frameworks, enabling a more holistic comprehension of the scenes. We manifest the feasibility and effectiveness of UniPAD by conducting extensive experiments on various 3D perception tasks. Our method significantly improves lidar-, camera-, and lidar-camera-based baseline by 9.1, 7.7, and 6.9 NDS, respectively. Notably, our pretraining pipeline achieves 73.2 NDS for 3D object detection and 79.4 mIoU for 3D semantic segmentation on the nuScenes validation set, achieving state-of-the-art results in comparison with previous methods.

Social relations have substantial impacts on the potential trajectories of each individual. Modeling these dynamics has been a central solution for more precise and accurate trajectory forecasting. However, previous works ignore the importance of `social depth', meaning the influences flowing from different degrees of social relations. In this work, we propose HighGraph, a graph-based relational reasoning method that captures the higher-order dynamics of social interactions. First, we construct a collision-aware relation graph based on the agents' observed trajectories. Upon this graph structure, we build our core module that aggregates the agent features from diverse social distances. As a result, the network is able to model more complex relations, thereby yielding more accurate and socially acceptable trajectories. Our HighGraph is a plug-and-play module that can be easily applied to any current trajectory predictors. Extensive experiments with ETH/UCY and SDD datasets demonstrate that our HighGraph noticeably improves the previous state-of-the-art baselines both quantitatively and qualitatively.

Predicting the trajectories of road agents is essential for autonomous driving systems. The recent mainstream methods follow a static paradigm, which predicts the future trajectory by using a fixed duration of historical frames. These methods make the predictions independently even at adjacent time steps, which leads to potential instability and temporal inconsistency. As successive time steps have largely overlapping historical frames, their forecasting should have intrinsic correlation, such as overlapping predicted trajectories should be consistent, or be different but share the same motion goal depending on the road situation. Motivated by this, in this work, we introduce HPNet, a novel dynamic trajectory forecasting method. Aiming for stable and accurate trajectory forecasting, our method leverages not only historical frames including maps and agent states, but also historical predictions. Specifically, we newly design a Historical Prediction Attention module to automatically encode the dynamic relationship between successive predictions. Besides, it also extends the attention range beyond the currently visible window benefitting from the use of historical predictions. The proposed Historical Prediction Attention together with the Agent Attention and Mode Attention is further formulated as the Triple Factorized Attention module, serving as the core design of HPNet. Experiments on the Argoverse and INTERACTION datasets show that HPNet achieves state-of-the-art performance, and generates accurate and stable future trajectories. Our code are available at \url{https://github.com/XiaolongTang23/HPNet}.

LiDAR localization is a fundamental task in robotics and computer vision, which estimates the pose of a LiDAR point cloud within a global map. Scene Coordinate Regression (SCR) has demonstrated state-of-the-art performance in this task. In SCR, a scene is represented as a neural network, which outputs the world coordinates for each point in the input point cloud. However, SCR treats all points equally during localization, ignoring the fact that not all objects are beneficial for localization. For example, dynamic objects and repeating structures often negatively impact SCR. To address this problem, we introduce LiSA, the first method that incorporates semantic awareness into SCR to boost the localization robustness and accuracy. To avoid extra computation or network parameters during inference, we distill the knowledge from a segmentation model to the original SCR network. Experiments show the superior performance of LiSA on standard LiDAR localization benchmarks compared to state-of-the-art methods. Applying knowledge distillation not only preserves high efficiency but also achieves higher localization accuracy than introducing extra semantic segmentation modules. We also analyze the benefit of semantic information for LiDAR localization. Our code will be released upon acceptance.

Predicting the future motion of surrounding agents is essential for autonomous vehicles (AVs) to operate safely in dynamic, human-robot-mixed environments. Context information, such as road maps and surrounding agents' states, provides crucial geometric and semantic information for motion behavior prediction. To this end, recent works explore two-stage prediction frameworks where coarse trajectories are first proposed, and then used to select critical context information for trajectory refinement. However, they either incur a large amount of computation or bring limited improvement, if not both. In this paper, we introduce a novel scenario-adaptive refinement strategy, named SmartRefine, to refine prediction with minimal additional computation. Specifically, SmartRefine can comprehensively adapt refinement configurations based on each scenario's properties, and smartly chooses the number of refinement iterations by introducing a quality score to measure the prediction quality and remaining refinement potential of each scenario. SmartRefine is designed as a generic and flexible approach that can be seamlessly integrated into most state-of-the-art motion prediction models. Experiments on Argoverse (1 \& 2) show that our method consistently improves the prediction accuracy of multiple state-of-the-art prediction models. Specifically, by adding SmartRefine to QCNet, we outperform all published ensemble-free works on the Argoverse 2 leaderboard (single agent track) at submission. Comprehensive studies are also conducted to ablate design choices and explore the mechanism behind multi-iteration refinement. Codes are available at https://github.com/opendilab/SmartRefine/.

Recent self-training techniques have shown significant improvements in unsupervised domain adaptation for 3D object detection (3D UDA).These techniques typically select pseudo labels (i.e. 3D boxes) to supervise the target domain. However, this selection process inevitably introduces unreliable 3D boxes, as the 3D points inside them cannot be definitively assigned as foreground or background points.Previous techniques reweight these unreliable boxes to reduce unreliability, but these boxes could still poison the training process.To resolve this problem, in this paper, we propose a pseudo label refinery framework.Specifically, in the selection process, to mitigate the unreliability of pseudo boxes, we propose a complementary augmentation strategy.This strategy either removes all points within each unreliable box or alternatively replaces it with a high-confidence box.Furthermore, the point numbers of instances in high-beam datasets are considerably higher than those in low-beam datasets, also degrading the quality of pseudo labels during the training process. We alleviate this issue by generating additional proposals and aligning RoI features across different domains. Experimental results demonstrate that our method effectively enhances the quality of pseudo labels and consistently surpasses the state-of-the-art methods on six autonomous driving benchmarks.

Collaborative perception allows for information sharing between multiple agents, such as vehicles and infrastructure, to obtain a comprehensive view of the environment through communication and fusion. Current research on multi-agent collaborative perception systems often assumes ideal communication and perception environments and neglects the effect of real-world noise such as pose noise, motion blur, and perception noise. To address this gap, in this paper, we propose a novel motion-aware robust communication network (MRCNet) that mitigates noise interference and achieves accurate and robust collaborative perception. MRCNet consists of two main components: multi-scale robust fusion (MRF) addresses pose noise by developing cross-semantic multi-scale enhanced aggregation to fuse features of different scales, while motion enhanced mechanism (MEM) captures motion context to compensate for information blurring caused by moving objects. Experimental results on popular collaborative 3D object detection datasets demonstrate that MRCNet outperforms competing methods in noisy scenarios with improved perception performance using less bandwidth.

Training deep models for LiDAR semantic segmentation is challenging due to the inherent sparsity of point clouds. Utilizing temporal data is a natural remedy against the sparsity problem as it makes the input signal denser. However, previous multi-frame fusion algorithms fall short in utilizing sufficient temporal information due to the memory constraint, and they also ignore the informative temporal images. To fully exploit rich information hidden in long-term temporal point clouds and images, we present the Temporal Aggregation Network, termed TASeg. Specifically, we propose a Temporal LiDAR Aggregation and Distillation (TLAD) algorithm, which leverages historical priors to assign different aggregation steps for different classes. It can largely reduce memory and time overhead while achieving higher accuracy. Besides, TLAD trains a teacher injected with gt priors to distill the model, further boosting the performance. To make full use of temporal images, we design a Temporal Image Aggregation and Fusion (TIAF) module, which can greatly expand the camera FOV and enhance the present features. Temporal LiDAR points in the camera FOV are used as mediums to transform temporal image features to the present coordinate for temporal multi-modal fusion. Moreover, we develop a Static-Moving Switch Augmentation (SMSA) algorithm, which utilizes sufficient temporal information to enable objects to switch their motion states freely, thus greatly increasing static and moving training samples. Extensive experiments are conducted on SemanticKITTI and nuScenes benchmarks. Notably, our TASeg ranks 1$^{st}$ on three challenging tracks, i.e., SemanticKITTI single-scan track, multi-scan track and nuScenes LiDAR segmentation track, strongly demonstrating the superiority of our method. Codes will be made publicly available upon publication.

Current sparsely-supervised object detection methods largely depend on high threshold settings to derive high-quality pseudo labels from detector predictions. However, hard instances within point clouds frequently display incomplete structures, causing decreased confidence scores in their assigned pseudo-labels. Previous methods inevitably result in inadequate positive supervision for these instances. To address this problem, we propose a novel Hard INsTance Enhanced Detector (HINTED), for sparsely-supervised 3D object detection. Firstly, we design a self-boosting teacher (SBT) model to generate more potential pseudo-labels, enhancing the effectiveness of information transfer. Then, we introduce a mixed-density student (MDS) model to concentrate on hard instances during the training phase, thereby improving detection accuracy. Our extensive experiments on the KITTI dataset validate our method's superior performance. Compared with leading sparsely-supervised methods, HINTED significantly improves the detection performance on hard instances, notably outperforming fully-supervised methods in detecting challenging categories like cyclists. HINTED also significantly outperforms the state-of-the-art semi-supervised method on challenging categories. The code will be made publicly available.

Knowledge distillation (KD) possesses immense potential to accelerate the deep neural networks (DNNs) for LiDAR-based 3D detection. However, in most of prevailing approaches, the suboptimal teacher models and insufficient student architecture investigations limit the performance gains. To address these issues, we propose a simple yet effective Category-aware Knowledge Distillation and Pruning (CaKDP) framework for compressing 3D detectors. Firstly, CaKDP transfers the knowledge of two-stage detector to one-stage student one, mitigating the impact of inadequate teacher models. To bridge the gap between the heterogeneous detectors, we investigate their differences, and then introduce the student-motivated category-aware KD to align the category prediction between distillation pairs. Secondly, we propose a category-aware pruning scheme to obtain the customizable architecture of compact student model. The method calculates the category prediction gap before and after removing each filter to evaluate the importance of filters, and retains the important filters. Finally, to further improve the student performance, a modified IOU-aware refinement module with negligible computations is leveraged to remove the redundant false positive predictions. Experiments demonstrate that CaKDP achieves the compact detector with high performance. For example, on WOD, CaKDP accelerates CenterPoint by half while boosting L2 mAPH by 1.61%. The code is available at https://github.com/zhnxjtu/CaKDP.

Diffusion models excel at modeling complex and multimodal trajectory distributions for decision-making and control.Reward-gradient guided denoising has been recently proposed to generate trajectories that maximize both a differentiable reward function and the likelihood under the data distribution captured by a diffusion model.Reward-gradient guided denoising requires a differentiable reward function fitted to both clean and noised samples, limiting its applicability as a general trajectory optimizer. In this paper, we propose DiffusionES, a method that combines gradient-free optimization with trajectory denoising to optimize black-box non-differentiable objectives while staying in the data manifold.Diffusion-ES samples trajectories during evolutionary search from a diffusion model and scores them using a black-box reward function. It mutates high-scoring trajectories using a truncated diffusion process that applies a small number of noising and denoising steps, allowing for much more efficient exploration of the solution space.We show that DiffusionES achieves state-of-the-art performance on nuPlan, an established closed-loop planning benchmark for autonomous driving.Diffusion-ES outperforms existing sampling-based planners, reactive deterministic or diffusion-based policies, and reward-gradient guidance.Additionally, we show that unlike prior guidance methods, our method can optimize non-differentiable language-shaped reward functions generated by few-shot LLM prompting.When guided by a human teacher that issues instructions to follow, our method can generate novel, highly complex behaviors, such as aggressive lane weaving, which are not present in the training data.This allows us to solve the hardest nuPlan scenarios which are beyond the capabilities of existing trajectory optimization methods and driving policies.

LiDAR Upsampling is a challenging task for the perception systems of robots and autonomous vehicles, due to the sparse and irregular structure of large-scale scene contexts. Recent works propose to solve this problem by converting LiDAR data from 3D Euclidean space into an image super-resolution problem in 2D image space. Although their methods can generate high-resolution range images with fine-grained details, the resulting 3D point clouds often blur out details and predict invalid points. In this paper, we propose TULIP, a new method to reconstruct high-resolution LiDAR point clouds from low-resolution LiDAR input. We also follow a range image-based approach but specifically modify the patch and window geometries of a Swin-Transformer-based network to better fit the characteristics of range images. We conducted several experiments on three different public real-world and simulated datasets. TULIP outperforms state-of-the-art methods in all relevant metrics and generates robust and more realistic point clouds than prior works. We plan to publish the code upon acceptance.

Knowledge of lane topology is a core problem in autonomous driving. Aerial imagery can provide high resolution, quickly updatable lane source data but detecting lanes from such data has so far been an expensive manual process or, where automated solutions exist, undrivable and requiring of downstream processing. We propose a method for large-scale lane topology extraction from aerial imagery while ensuring that the resulting lanes are realistic and drivable by introducing a novel Bézier Graph shared parameterisation of Bézier curves. We develop a transformer-based model to predict these Bézier Graphs from input aerial images, demonstrating competitive results on the UrbanLaneGraph dataset. We demonstrate that our method generates realistic lane graphs which require both minimal input, and minimal downstream processing. We make our code publicly available at https://github.com/driskai/BGFormer

Stereo rectification is widely considered ``solved'' due to the abundance of traditional approaches to perform rectification. However, autonomous vehicles and robots in-the-wild require constant re-calibration due to exposure to various environmental factors, including vibration, and structural stress, when cameras are arranged in a wide-baseline configuration. Conventional rectification methods fail in these challenging scenarios: especially for larger vehicles, such as autonomous freight trucks and semi-trucks, the resulting incorrect rectification severely affects the quality of downstream tasks that use stereo/multi-view data. To tackle these challenges, we propose an online rectification approach that operates at real-time rates while achieving high accuracy. We propose a novel learning-based online calibration approach that utilizes stereo correlation volumes built from a feature representation obtained from cross-image attention. Our model is trained to minimize vertical optical flow as proxy rectification constraint, and predicts the relative rotation between the stereo pair. The method is real-time and even outperforms conventional methods used for offline calibration, and substantially improves downstream stereo depth, post-rectification. We release two public datasets, a synthetic and experimental wide baseline dataset, to foster further research.

Microscopic traffic simulation plays a crucial role in transportation engineering by providing insights into individual vehicle behavior and overall traffic flow. However, creating a realistic simulator that accurately replicates human driving behaviors in various traffic conditions presents significant challenges. Traditional simulators relying on heuristic models often fail to deliver accurate simulations due to the complexity of real-world traffic environments. Due to the covariate shift issue, existing imitation learning-based simulators often fail to generate stable long-term simulations. In this paper, we propose a novel approach called learner-aware supervised imitation learning to address the covariate shift problem in multi-agent imitation learning. By leveraging a variational autoencoder simultaneously modeling the expert and learner state distribution, our approach augments expert states such that the augmented state is aware of learner state distribution. Our method, applied to urban traffic simulation, demonstrates significant improvements over existing state-of-the-art baselines in both short-term microscopic and long-term macroscopic realism when evaluated on the real-world dataset pNEUMA.

Vectorized High-Definition (HD) map construction requires predictions of the category and point coordinates of map elements (e.g. road boundary, lane divider, pedestrian crossing, etc.). State-of-the-art methods are mainly based on point-level representation learning for regressing accurate point coordinates. However, this pipeline has limitations in obtaining element-level information and handling element-level failures, e.g. erroneous element shape or entanglement between elements. To tackle the above issues, we propose a simple yet effective HybrId framework named HIMap to sufficiently learn and interact both point-level and element-level information. Concretely, we introduce a hybrid representation called HIQuery to represent all map elements, and propose a point-element interactor to interactively extract and encode the hybrid information of elements, e.g. point position and element shape, into the HIQuery. Additionally, we present a point-element consistency constraint to enhance the consistency between the point-level and element-level information. Finally, the output point-element integrated HIQuery can be directly converted into map elements’ class, point coordinates, and mask. We conduct extensive experiments and consistently outperform previous methods on both nuScenes and Argoverse2 datasets. Notably, our method achieves 77.8 mAP on the nuScenes dataset, remarkably superior to previous SOTAs by 8.3 mAP at least.

Object detection in radar imagery with neural networks shows great potential for improving autonomous driving. However, obtaining annotated datasets from real radar images, crucial for training these networks, is challenging, especially in scenarios with long-range detection and adverse weather and lighting conditions where radar performance excels. To address this challenge, we present RadSimReal, an innovative physical radar simulation capable of generating synthetic radar images with accompanying annotations for various radar types and environmental conditions, all without the need for real data collection. Remarkably, our findings demonstrate that training object detection models on RadSimReal data and subsequently evaluating them on real-world data produce performance levels comparable to models trained and tested on real data from the same dataset, and even achieves better performance when testing across different real datasets. RadSimReal offers advantages over other physical radar simulations that it does not necessitate knowledge of the radar design details, which are often not disclosed by radar suppliers, and has faster run-time. This innovative tool has the potential to advance the development of computer vision algorithms for radar-based autonomous driving applications.

Building accurate maps is a key building block to enable reliable localization, planning, and navigation of autonomous vehicles. We propose a novel approach for building accurate 3D maps of dynamic environments utilizing a sequence of LiDAR scans. To this end, we propose encoding the 4D scene into a novel spatio-temporal implicit neural map representation by fitting a time-dependent truncated signed distance function to each point. Using our representation, we can extract the static map by filtering the dynamic parts. Our neural representation is based on sparse feature grids, a globally shared decoder, and time-dependent basis functions, which can be jointly optimized in an unsupervised fashion. To learn this representation from a sequence of LiDAR scans, we design a simple yet efficient loss function to supervise the map optimization in a piecewise way. We evaluate our approach on various scenes containing moving objects in terms of the reconstruction quality of static maps and the segmentation of dynamic point clouds. The experimental results demonstrate that our method is capable of removing the dynamic part of the input point clouds while reconstructing accurate and complete large-scale 3D maps outperforming several state-of-the-art methods for static map generation and scene reconstruction.

Safety and robustness are crucial factors in developing trustworthy autonomous vehicles. One essential aspect of addressing these factors is to equip vehicles with the capability to predict future trajectories for all moving objects in the surroundings and quantify prediction uncertainties. In this paper, we propose the Sequential Neural Variational Agent (SeNeVA), a generative model that describes the distribution of future trajectories for a single moving object. Our approach can distinguish Out-of-Distribution data while quantifying uncertainty and achieving competitive performance compared to state-of-the-art methods on the Argoverse 2 and INTERACTION datasets. Specifically, a 0.446 meters minimum Final Displacement Error, a 0.203 meters minimum Average Displacement Error, and a 5.35\% Miss Rate are achieved on the INTERACTION test set. Extensive qualitative and quantitative analysis is also provided to evaluate the proposed model.

Embodied AI, such as autonomous vehicles, suffers from insufficient, long-tailed data because it must be obtained from the physical world. In fact, data must be continuously obtained in a series of small batches, and the model must also be continuously trained to achieve generalizability and scalability by improving the biased data distribution. This paper addresses the training cost and catastrophic forgetting problems when continuously updating models to adapt to incoming small batches from various environments for real-world motion prediction in autonomous driving. To this end, we propose a novel continual motion prediction (CMP) learning framework based on sparse meta-representation learning and an optimal memory buffer retention strategy. In meta-representation learning, a model explicitly learns a sparse representation of each driving environment, from road geometry to vehicle states, by training to reduce catastrophic forgetting based on an augmented modulation network with sparsity regularization. Also, in the adaptation phase, We develop an Optimal Memory Buffer Retention strategy that smartly preserves diverse samples by focusing on representation similarity. This approach handles the nuanced task distribution shifts characteristic of motion prediction datasets, ensuring our model stays responsive to evolving input variations without requiring extensive resources. The experiment results demonstrate that the proposed method shows superior adaptation performance to the conventional continual learning approach, which is developed using a synthetic dataset for the continual learning problem.

Recent works have proposed end-to-end autonomous vehicle (AV) architectures comprised of differentiable modules, achieving state-of-the-art driving performance. While they provide advantages over the traditional perception-prediction-planning pipeline (e.g., removing information bottlenecks between components and alleviating integration challenges), they do so using a diverse combination of tasks, modules, and their interconnectivity. As of yet, however, there has been no systematic analysis of the necessity of these modules or the impact of their connectivity, placement, and internal representations on overall driving performance. Addressing this gap, our work conducts a comprehensive exploration of the design space of end-to-end modular AV stacks. Our findings culminate in the development of PARA-Drive: a fully parallel end-to-end AV architecture. PARA-Drive not only achieves state-of-the-art performance in perception, prediction, and planning, but also significantly enhances runtime speed by nearly 3x, without compromising on interpretability or safety.

We present ChatScene, a unified and effective framework that leverages the capabilities of Large Language Models (LLMs) to guide the generation of safety-critical scenarios for autonomous vehicles. This framework distinctively transforms textually described scenarios generated by LLMs into domain-specific languages, which then generate actual code for prediction and control in simulators, facilitating the creation of diverse and complex scenarios within the CARLA simulation environment. A key part of our approach is a comprehensive knowledge retrieval component, which efficiently translates specific textual descriptions into corresponding domain-specific code snippets by training a knowledge database containing the scenario description and code pairs. Extensive experimental results underscore the efficacy of ChatScene in improving the safety of autonomous vehicles. For instance, the scenarios generated by ChatScene show a 15% increase in collision rates compared to state-of-the-art baselines when tested against different reinforcement learning-based ego vehicles. Furthermore, we show that by using our generated safety-critical scenarios to fine-tune different autonomous agents, they can achieve a 9% reduction in collision rates, surpassing current SOTA methods. ChatScene effectively bridges the gap between textual descriptions of traffic scenarios and practical CARLA simulations.

In the field of 3D object detection for autonomous driving, LiDAR-Camera (LC) fusion is the top-performing sensor configuration. Still, LiDAR is relatively high cost, which hinders adoption of this technology for consumer automobiles. Alternatively, camera and radar are commonly deployed on vehicles already on the road today, but performance of Camera-Radar (CR) fusion falls behind LC fusion. In this work, we propose Camera-Radar Knowledge Distillation (CRKD) to bridge the performance gap between LC and CR detectors with a novel cross-modality KD framework. We use the Bird's-Eye-View (BEV) representation as the shared feature space to enable effective knowledge distillation. To accommodate the unique cross-modality KD path, we propose four distillation losses to help the student learn crucial features from the teacher model. We present extensive evaluations on the nuScenes dataset to demonstrate the effectiveness of the proposed CRKD framework. The project page for CRKD is https://song-jingyu.github.io/CRKD.

Collaborative perception empowers each agent to improve its perceptual ability through the exchange of perceptual messages with other agents. It inherently results in a fundamental trade-off between perception ability and communication cost. To address this bottleneck issue, our core idea is to optimize the collaborative messages from two key aspects: representation and selection. The proposed codebook-based message representation enables the transmission of integer codes, rather than high-dimensional feature maps. The proposed information-filling-driven message selection optimizes local messages to collectively fill each agent's information demand, preventing information overflow among multiple agents. By integrating these two designs, we propose CodeFilling, a novel communication-efficient collaborative perception system, which significantly advances the perception-communication trade-off and is inclusive to both homogeneous and heterogeneous collaboration settings. We evaluate CodeFilling in both a real-world dataset, DAIR-V2X, and a new simulation dataset, OPV2VH+. Results show that CodeFilling outperforms previous SOTA Where2comm on DAIR-V2X/OPV2VH+ with 1,333/1,206 times lower communication volume. Our code is available at~\url{https://github.com/PhyllisH/CodeFilling}.

The inherent noisy and sparse characteristics of radar data pose challenges in finding effective representations for 3D object detection. In this paper, we propose RadarDistill, a novel knowledge distillation (KD) method, which can improve the representation of radar data by leveraging LiDAR data. RadarDistill successfully transfers desirable characteristics of LiDAR features into radar features using three key components: Cross-Modality Alignment (CMA), Activation-based Feature Distillation (AFD), and Proposal-based Feature Distillation (PFD). CMA enhances the density of radar features by employing multiple layers of dilation operations, effectively addressing the challenge of inefficient knowledge transfer from LiDAR to radar. AFD selectively transfers knowledge based on regions of the LiDAR features, with a specific focus on areas where activation intensity exceeds a predefined threshold. PFD similarly guides the radar network to selectively mimic features from the LiDAR network within the object proposals. Our comparative analyses conducted on the nuScenes datasets demonstrate that RadarDistill achieves state-of-the-art (SOTA) performance for radar-only object detection task, recording 20.5% in mAP and 43.7% in NDS. Also, RadarDistill significantly improves the performance of the camera-radar fusion model.

Scene flow characterizes the 3D motion between two LiDAR scans captured by an autonomous vehicle at nearby timesteps.Prevalent methods consider scene flow as point-wise unconstrained flow vectors that can be learned by either large-scale training beforehand or time-consuming optimization at inference. However, these methods do not take into account that objects in autonomous driving often move rigidly. We incorporate this rigid-motion assumption into our design, where the goal is to associate objects over scans and then estimate the locally rigid transformations. We propose ICP-Flow, a learning-free flow estimator.The core of our design is the conventional Iterative Closest Point (ICP) algorithm, which aligns the objects over time and outputs the corresponding rigid transformations. Crucially, to aid ICP, we propose a histogram-based initialization that discovers the most likely translation, thus providing a good starting point for ICP. The complete scene flow is then recovered from the rigid transformations.We outperform state-of-the-art baselines, including supervised models, on the Waymo dataset and perform competitively on Argoverse-v2 and nuScenes. Further, we train a feedforward neural network, supervised by the pseudo labels from our model, and achieve top performance among all models capable of real-time inference. We validate the advantage of our model on scene flow estimation with longer temporal gaps, up to 0.4 seconds where other models fail to deliver meaningful results.

Semantic segmentation in bird’s eye view (BEV) plays a crucial role in autonomous driving. Previous methods usually follow an end-to-end pipeline, directly predicting the BEV segmentation map from monocular RGB inputs. However, the challenge arises when the RGB inputs and BEV targets represent distinct perspectives, making the direct point-to-point translation hard to optimize. In this paper, we decompose the original BEV segmentation task into two stages, namely BEV map reconstruction and cross-view feature alignment. In the first stage, we train a BEV autoencoder to perfectly reconstruct the BEV segmentation label maps given corrupted noisy latent representation, which urges the decoder to learn fundamental knowledge of typical BEV scene patterns. The second stage involves mapping RGB input images into the pretrained BEV latent space of Stage I, directly optimizing the correlations between the two views at the feature level. Our approach simplifies the complexity of combining perception and generation into a single step, equipping the model to handle intricate and challenging scenes effectively. Besides, we propose to transform the BEV segmentation map from the Cartesian coordinate system to the polar coordinate system. This conversion allows us to correspond each RGB image column directly to BEV segmentation targets, one by one. Extensive experiments on the nuScenes and Argoverse datasets show our model clearly outperforms state-of-the-art models.

Vision-centric autonomous driving has recently raised wide attention due to its lower cost. Pre-training is essential for extracting a universal representation. However, current vision-centric pre-training typically relies on either 2D or 3D pre-text tasks, overlooking the temporal characteristics of autonomous driving as a 4D scene understanding task. In this paper, we address this challenge by introducing a world model-based autonomous driving 4D representation learning framework, dubbed DriveWorld, which is capable of pre-training from multi-camera driving videos in a spatio-temporal fashion. Specifically, we propose a Memory State-Space Model for spatio-temporal modelling, which consists of a Dynamic Memory Bank module for learning temporal-aware latent dynamics to predict future changes and a Static Scene Propagation module for learning spatial-aware latent statics to offer comprehensive scene contexts. We additionally introduce a Task Prompt to decouple task-aware features for various downstream tasks. The experiments demonstrate that DriveWorld delivers promising results on various autonomous driving tasks. When pre-trained with OpenScene dataset, DriveWorld achieves a 7.5% increase in mAP for 3D object detection, a 3.0% increase in IoU for online mapping, a 5.0% increase in AMOTA for multi-object tracking, a 0.1m decrease in minADE for motion forecasting, a 3.0% increase in IoU for occupancy prediction, and a 0.34m reduction in average L2 error for planning. Code will be made publicly available.

High-definition (HD) maps have played an integral role in the development of modern autonomous vehicle (AV) stacks, albeit with high associated labeling and maintenance costs. As a result, many recent works have proposed methods for estimating HD maps online from sensor data, enabling AVs to operate outside of previously-mapped regions. However, current online map estimation approaches are developed in isolation of their downstream tasks, complicating their integration in AV stacks. In particular, they do not produce uncertainty or confidence estimates. In this work, we extend multiple state-of-the-art online map estimation methods to additionally output uncertainty estimates and show how they enable more tightly integrating online mapping with trajectory forecasting. In doing so, we find that incorporating uncertainty yields up to 50% faster training convergence and up to 15% better prediction performance on the real-world nuScenes driving dataset.

Leveraging vast training data, multimodal large language models (MLLMs) have demonstrated formidable general visual comprehension capabilities and achieved remarkable performance across various tasks. However, their performance in visual document understanding still leaves much room for improvement. This discrepancy is primarily attributed to the fact that visual document understanding is a fine-grained prediction task. In natural scenes, MLLMs typically use low-resolution images, leading to a substantial loss of visual information. Furthermore, general-purpose MLLMs do not excel in handling document-oriented instructions. In this paper, we propose a High-Resolution Visual Document Assistant (HRVDA), which bridges the gap between MLLMs and visual document understanding. This model employs a content filtering mechanism and an instruction filtering module to separately filter out the content-agnostic visual tokens and instruction-agnostic visual tokens, thereby achieving efficient model training and inference for high-resolution images. In addition, we construct a document-oriented visual instruction tuning dataset and apply a multi-stage training strategy to enhance the model's document modeling capabilities. Extensive experiments demonstrate that our model achieves state-of-the-art performance across multiple document understanding datasets, while maintaining training efficiency and inference speed comparable to low-resolution models.

Recently, the advent of Large Visual-Language Models (LVLMs) has received increasing attention across various domains, particularly in the field of visual document understanding (VDU). Different from conventional vision-language tasks, VDU is specifically concerned with text-rich scenarios containing abundant document elements. Nevertheless, the importance of fine-grained features remains largely unexplored within the community of LVLMs, leading to suboptimal performance in text-rich scenarios. In this paper, we abbreviate it as the fine-grained feature collapse issue. With the aim of filling this gap, we propose a contrastive learning framework, termed Document Object COntrastive learning (DoCo), specifically tailored for the downstream tasks of VDU. DoCo leverages an auxiliary multimodal encoder to obtain the features of document objects and align them to the visual features generated by the vision encoder of LVLM, which enhances visual representation in text-rich scenarios. It can represent that the contrastive learning between the visual holistic representations and the multimodal fine-grained features of document objects can assist the vision encoder in acquiring more effective visual cues, thereby enhancing the comprehension of text-rich documents in LVLMs. We also demonstrate that the proposed DoCo serves as a plug-and-play pre-training method, which can be employed in the pre-training of various LVLMs without inducing any increase in computational complexity during the inference process. Extensive experimental results on multiple benchmarks of VDU reveal that LVLMs equipped with our proposed DoCo can achieve superior performance and mitigate the gap between VDU and generic vision-language tasks.

Before developing a Document Layout Analysis (DLA) model in real-world applications, conducting comprehensive robustness testing is essential. However, the robustness of DLA models remains underexplored in the literature. To address this, we are the first to introduce a {robustness benchmark} for DLA models, which includes 450K document images of three datasets. To cover realistic corruptions, we propose a perturbation taxonomy with ${12}$ common document perturbations with 3 severity levels inspired by real-world document processing. Additionally, to better understand document perturbation impacts, we propose two metrics, Mean Perturbation Effect (mPE) for perturbation assessment and Mean Robustness Degradation (mRD) for robustness evaluation. Furthermore, we introduce a self-titled model, \ie, Robust Document Layout Analyzer (RoDLA), which improves attention mechanisms to boost extraction of robust features. Experiments on the proposed benchmarks (PubLayNet-P, DocLayNet-P, and M$^6$Doc-P) demonstrate that RoDLA obtains state-of-the-art mRD scores of $115.7$, $135.4$, and $150.4$, respectively. Compared to previous methods, RoDLA achieves notable improvements in mAP of $+3.8\%$, $+7.1\%$ and $+12.1\%$, respectively.

Scene text recognition (STR) in the wild frequently encounters challenges when coping with domain variations, font diversity, shape deformations, etc. A straightforward solution is performing model fine-tuning tailored to a specific scenario, but it is computationally intensive and requires multiple model copies for various scenarios. Recent studies indicate that large language models (LLMs) can learn from a few demonstration examples in a training-free manner, termed ``In-Context Learning" (ICL). Nevertheless, applying LLMs as a text recognizer is unacceptably resource-consuming. Moreover, our pilot experiments on LLMs show that ICL fails in STR, mainly attributed to the insufficient incorporation of contextual information from diverse samples in the training stage. To this end, we introduce E$^2$STR, a STR model trained with context-rich scene text sequences, where the sequences are generated via our proposed in-context training strategy. E$^2$STR demonstrates that a regular-sized model is sufficient to achieve effective ICL capabilities in STR. Extensive experiments show that E$^2$STR exhibits remarkable training-free adaptation in various scenarios and outperforms even the fine-tuned state-of-the-art approaches on public benchmarks.

The rebroadcasting of screen-recaptured document images introduces a significant risk to the confidential documents processed in government departments and commercial companies. However, detecting recaptured document images subjected to distortions from online social networks (OSNs) is challenging since the common forensics cues, such as moiré pattern, are weakened during transmission.In this work, we first devise a pixel-level distortion model of the screen-recaptured document image to identify the robust features of color artifacts.Then, we extract a chromaticity map from the recaptured image to highlight the presence of color artifacts even under low-quality samples.Based on the prior understanding, we design a chromaticity map adapter (CMA) to efficiently extract the chromaticity map, fed into the transformer backbone as multi-modal prompt tokens.To evaluate the performance of the proposed method, we collect a recaptured office document image dataset with over 10K diverse samples.Experimental results demonstrate that the proposed CMA method outperforms a SOTA approach (with RGB modality only), reducing the average EER from 26.82\% to 16.78\%.Robustness evaluation shows that our method achieves 0.8688 and 0.7554 AUCs under samples with JPEG compression (QF=70) and resolution as low as 534$\times$503 pixels.

In recent years, text-image joint pre-training techniques have shown promising results in various tasks. However, in Optical Character Recognition (OCR) tasks, aligning text instances with their corresponding text regions in images poses a challenge, as it requires effective alignment between text and OCR-Text (referring to the text in images as OCR-Text to distinguish from the text in natural language) rather than a holistic understanding of the overall image content. In this paper, we propose a new pre-training method called OCR-Text Destylization Modeling (ODM) that transfers diverse styles of text found in images to a uniform style based on the text prompt. With ODM, we achieve better alignment between text and OCR-Text and enable pre-trained models to adapt to the complex and diverse styles of scene text detection and spotting tasks. Additionally, we have designed a new labeling generation method specifically for ODM and combined it with our proposed Text-Controller module to address the challenge of annotation costs in OCR tasks, allowing a larger amount of unlabeled data to participate in pre-training. Extensive experiments on multiple public datasets demonstrate that our method significantly improves performance and outperforms current pre-training methods in scene text detection and spotting tasks. Code is available at https://github.com/PriNing/ODM.

The increasing use of transformer-based large language models brings forward the challenge of processing long sequences.In document visual question answering (DocVQA), leading methods focus on the single-page setting, while documents can span hundreds of pages.We present GRAM, a method that seamlessly extends pre-trained single-page models to the multi-page setting, without requiring computationally-heavy pretraining.To do so, we leverage a single-page encoder for local page-level understanding, and enhance it with document-level designated layers and learnable tokens, facilitating the flow of information across pages for global reasoning.To enforce our model to utilize the newly introduced document-level tokens, we propose a tailored bias adaptation method.For additional computational savings during decoding, we introduce an optional compression stage using our C-Former model, which reduces the encoded sequence length, thereby allowing a tradeoff between quality and latency.Extensive experiments showcase GRAM's state-of-the-art performance on the benchmarks for multi-page DocVQA, demonstrating the effectiveness of our approach.

Modularity plays a crucial role in the development and maintenance of complex systems. While end-to-end text spotting efficiently mitigates the issues of error accumulation and sub-optimal performance seen in traditional two-step methodologies, the two-step methods continue to be favored in many competitions and practical settings due to their superior modularity. In this paper, we introduce Bridging Text Spotting, a novel approach that resolves the error accumulation and suboptimal performance issues in two-step methods while retaining modularity. To achieve this, we adopt a well-trained detector and recognizer that are developed and trained independently and then lock their parameters to preserve their already acquired capabilities. Subsequently, we introduce a Bridge that connects the locked detector and recognizer through a zero-initialized neural network. This zero-initialized neural network, initialized with weights set to zeros, ensures seamless integration of the large receptive field features in detection into the locked recognizer. Furthermore, since the fixed detector and recognizer cannot naturally acquire end-to-end optimization features, we adopt the Adapter to facilitate their efficient learning of these features. We demonstrate the effectiveness of the proposed method through extensive experiments: Connecting the latest detector and recognizer through Bridging Text Spotting, we achieved an accuracy of 83.3% on Total-Text, 69.8% on CTW1500, and 89.5% on ICDAR 2015. The code is available at https://github.com/mxin262/Bridging-Text-Spotting.

The laws of model size, data volume, computation and model performance have been extensively studied in the field of Natural Language Processing (NLP). However, the scaling laws in Scene Text Recognition (STR) have not yet been investigated. To address this, we conducted comprehensive studies that involved examining the correlations between performance and the scale of models, data volume and computation in the field of text recognition. Conclusively, the study demonstrates smooth power laws between performance and model size, as well as training data volume, when other influencing factors are held constant. Additionally, we have constructed a large-scale dataset called REBU-Syn, which comprises 6 million real samples and 18 million synthetic samples. Based on our scaling law and new dataset, we have successfully trained a scene text recognition model, achieving a new state-of-the-art on 6 common test benchmarks with a top-1 average accuracy of 97.42%. The models and dataset are publicly available at large-ocr-model.github.io.

Recently, leveraging large language models (LLMs) or multimodal large language models (MLLMs) for document understanding has been proven very promising. However, previous works that employ LLMs/MLLMs for document understanding have not fully explored and utilized the document layout information, which is vital for precise document understanding. In this paper, we propose LayoutLLM, an LLM/MLLM based method for document understanding. The core of LayoutLLM is a layout instruction tuning strategy, which is specially designed to enhance the comprehension and utilization of document layouts. The proposed layout instruction tuning strategy consists of two components: Layout-aware Pre-training and Layout-aware Supervised Fine-tuning. To capture the characteristics of document layout in Layout-aware Pre-training, three groups of pre-training tasks, corresponding to document-level, region-level and segment-level information, are introduced. Furthermore, a novel module called layout chain-of-thought (LayoutCoT) is devised to enable LayoutLLM to focus on regions relevant to the question and generate accurate answers. LayoutCoT is effective for boosting the performance of document understanding. Meanwhile, it brings a certain degree of interpretability, which could facilitate manual inspection and correction. Experiments on standard benchmarks show that the proposed LayoutLLM significantly outperforms existing methods that adopt open-source 7B LLMs/MLLMs for document understanding.

Recently, visually-situated text parsing (VsTP) has experienced notable advancements, driven by the increasing demand for automated document understanding and the emergence of Generative Large Language Models (LLMs) capable of processing document-based questions. Various methods have been proposed to address the challenging problem of VsTP. However, due to the diversified targets and heterogeneous schemas, previous works usually design task-specific architectures and objectives for individual tasks, which inadvertently leads to modal isolation and complex workflow. In this paper, we propose a unified paradigm for parsing visually-situated text across diverse scenarios. Specifically, we devise a universal model, called OmniParser, which can simultaneously handle three typical visually-situated text parsing tasks: text spotting, key information extraction, and table recognition. In OmniParser, all tasks share the unified encoder-decoder architecture, the unified objective: point-conditioned text generation, and the unified input & output representation: prompt & structured sequences. Extensive experiments demonstrate that the proposed OmniParser achieves state-of-the-art (SOTA) or highly competitive performances on 7 datasets for the three visually-situated text parsing tasks, despite its unified, concise design. The code is available at https://github.com/AlibabaResearch/AdvancedLiterateMachinery.

Document image restoration is a crucial aspect of Document AI systems, as the quality of document images significantly influences the overall performance. Prevailing methods address distinct restoration tasks independently, leading to intricate systems and the incapability to harness the potential synergies of multi-task learning. To overcome this challenge, we propose DocRes, a generalist model that unifies five document image restoration tasks including dewarping, deshadowing, appearance enhancement, deblurring, and binarization. To instruct DocRes to perform various restoration tasks, we propose a novel visual prompt approach called Dynamic Task-Specific Prompt (DTSPrompt). The DTSPrompt for different tasks comprises distinct prior features, which are additional characteristics extracted from the input image. Beyond its role as a cue for task-specific execution, DTSPrompt can also serve as supplementary information to enhance the model's performance. Moreover, DTSPrompt is more flexible than prior visual prompt approaches as it can be seamlessly applied and adapted to inputs with high and variable resolutions. Experimental results demonstrate that DocRes achieves competitive or superior performance compared to existing state-of-the-art task-specific models. This underscores the potential of DocRes across a broader spectrum of document image restoration tasks. The source code will be publicly available at http://anonymous.for.review.

Existing scene text detectors generally focus on accurately detecting single-level (i.e., word-level, line-level, or paragraph-level) text entities without exploring the relationships among different levels of text entities. To comprehensively understand scene texts, detecting multi-level texts while exploring their contextual information is critical. To this end, we propose a unified framework (dubbed LayoutFormer) for hierarchical text detection, which simultaneously conducts multi-level text detection and predicts the geometric layouts for promoting scene text understanding. In LayoutFormer, WordDecoder, LineDecoder, and ParaDecoder are proposed to be responsible for word-level text prediction, line-level text prediction, and paragraph-level text prediction, respectively. Meanwhile, WordDecoder and ParaDecoder adaptively learn word-line and line-paragraph relationships, respectively. In addition, we propose a Prior Location Sampler to be used on multi-scale features to adaptively select a few representative foreground features for updating text queries. It can improve hierarchical detection performance while significantly reducing the computational cost. Comprehensive experiments verify that our method achieves state-of-the-art performance on single-level and hierarchical text detection.

In this paper, we solve a novel challenging generation task, i.e. Handwritten Mathematical Expression Generation (HMEG) from symbolic sequences. Since symbolic sequences are naturally graph-structured data, we formulate HMEG as a graph-to-image (G2I) generation problem. Unlike the generation of natural images, HMEG requires critic layout clarity for synthesizing correct and recognizable formulas, but has no real masks available to supervise the learning process. To alleviate this challenge, we propose a novel end-to-end G2I generation pipeline (i.e. $graph \rightarrow layout \rightarrow mask \rightarrow image$), which requires no real masks or nondifferentiable alignment between layouts and masks. Technically, to boost the capacity of predicting detailed relations among adjacent symbols, we propose a Less-is-More (LiM) learning strategy. In addition, we design a differentiable layout refinement module, which maps bounding boxes to pixel-level soft masks, so as to further alleviate ambiguous layout areas. Our whole model, including layout prediction, mask refinement, and image generation, can be jointly optimized in an end-to-end manner. Experimental results show that, our model can generate high-quality HME images, and outperforms previous generative methods. Besides, a series of ablations study demonstrate effectiveness of the proposed techniques. Finally, we validate that our generated images promisingly boosts the performance of HME recognition models, through data augmentation. Our code will be released after peer review.

Event-based semantic segmentation (ESS) is a fundamental yet challenging task for event camera sensing. The difficulties in interpreting and annotating event data limit its scalability. While domain adaptation from images to event data can help to mitigate this issue, there exist data representational differences that require additional effort to resolve. In this work, for the first time, we synergize information from image, text, and event-data domains and introduce OpenESS to enable scalable ESS in an open-world, annotation-efficient manner. We achieve this goal by transferring the semantically rich CLIP knowledge from image-text pairs to event streams. To pursue better cross-modality adaptation, we propose a frame-to-event contrastive distillation and a text-to-event semantic consistency regularization. Experimental results on popular ESS benchmarks showed our approach outperforms existing methods. Notably, we achieve 53.93% and 43.31% mIoU on DDD17 and DSEC-Semantic without using either event or frame labels.

Applying a pre-trained large model to downstream tasks is prohibitive under resource-constrained conditions.Recent dominant approaches for addressing efficiency issues involve adding a few learnable parameters to the fixed backbone model.This strategy, however, leads to more challenges in loading large models for downstream fine-tuning with limited resources.In this paper, we propose a novel method for increasing the parameter efficiency of pre-trained models by introducing an intermediate pre-training stage.To this end, we first employ low-rank approximation to compress the original large model and then devise a feature distillation module and a weight perturbation regularization module.These modules are specifically designed to enhance the low-rank model.In particular, we update only the low-rank model while freezing the backbone parameters during pre-training. This allows for direct and efficient utilization of the low-rank model for downstream fine-tuning tasks.The proposed method achieves both efficiencies in terms of required parameters and computation time while maintaining comparable results with minimal modifications to the backbone architecture.Specifically, when applied to three vision-only and one vision-language Transformer models, our approach often demonstrates a merely $\sim$0.6 point decrease in performance while reducing the original parameter size by 1/3 to 2/3.The code has been released in the supplementary material for reproduction.

Vision-Language Transformers (VLTs) have shown great success recently, but are meanwhile accompanied by heavy computation costs, where a major reason can be attributed to the large number of visual and language tokens. Existing token pruning research for compressing VLTs mainly follows a single-modality-based scheme yet ignores the critical role of aligning different modalities for guiding the token pruning process, causing the important tokens for one modality to be falsely pruned in another modality branch. Meanwhile, existing VLT pruning works also lack the flexibility to dynamically compress each layer based on different input samples. To this end, we propose a novel framework named Multimodal Alignment-Guided Dynamic Token Prunning (MADTP) for accelerating various VLTs. Specifically, we first introduce a well-designed Multi-modality Alignment Guidance (MAG) module that can align features of the same semantic concept from different modalities, to ensure the pruned tokens are less important for all modalities. We further design a novel Dynamic Token Pruning (DTP) module, which can adaptively adjust the token compression ratio in each layer based on different input instances. Extensive experiments on various benchmarks demonstrate that MADTP significantly reduces the computational complexity of kinds of multimodal models while preserving competitive performance. Notably, when applied to the BLIP model in the NLVR2 dataset, MADTP can reduce the GFLOPs by 80% with less than 4% performance degradation.

Knowledge distillation is an effective method for training small and efficient deep learning models. However, the efficacy of a single method can degenerate when transferring to other tasks, modalities, or even other architectures. To address this limitation, we propose a novel constrained feature distillation method. This method is derived from a small set of core principles, which results in two emerging components: an orthogonal projection and a task-specific normalisation. Equipped with both of these components, our transformer models can outperform all previous methods on ImageNet and reach up to 4.4% relative improvement over the previous state-of-the-art methods. To further demonstrate the generality of our method, we apply it to object detection and image generation, whereby we obtain consistent and substantial performance improvements over state-of-the-art. We provide code in the supplementary.

Knowledge distillation involves transferring soft labels from a teacher to a student using a shared temperature-based softmax function. However, the assumption of a shared temperature between teacher and student implies a mandatory exact match between their logits in terms of logit range and variance. This side-effect limits the performance of student, considering the capacity discrepancy between them and the finding that the innate logit relations of teacher are sufficient for student to learn. To address this issue, we propose setting the temperature as the weighted standard deviation of logit and performing a plug-and-play $\mathcal{Z}$-score pre-process of logit standardization before applying softmax and Kullback-Leibler divergence. Our pre-process enables student to focus on essential logit relations from teacher rather than requiring a magnitude match, and can improve the performance of existing logit-based distillation methods. We also show a typical case where the conventional setting of sharing temperature between teacher and student cannot reliably yield an authentic distillation evaluation; nonetheless, this challenge is successfully alleviated by our $\mathcal{Z}$-score. We extensively evaluate our method for various student and teacher models on CIFAR-100 and ImageNet, showing its significant superiority. The vanilla knowledge distillation powered by our pre-process can achieve favorable performance against state-of-the-art methods, and other distillation variants can obtain considerable gain with the assistance of our pre-process.

Vision Transformer (ViT) has emerged as a prominent backbone for computer vision. For more efficient ViTs, recent works lessen the quadratic cost of the self-attention layer by pruning or fusing the redundant tokens. However, these works faced the speed-accuracy trade-off caused by the loss of information. Here, we argue that token fusion needs to consider diverse relations between tokens to minimize information loss. In this paper, we propose a Multi-criteria Token Fusion, that gradually fuses the tokens based on multi-criteria (\ie, similarity, informativeness, and size of fused tokens). Further, we utilize the one-step-ahead attention, which is the improved approach to capture the informativeness of the tokens.By training the model equipped with MCTF using a token reduction consistency, we achieve the best speed-accuracy trade-off in the image classification (ImageNet1K). Experimental results prove that MCTF consistently surpasses the previous reduction methods with and without training. Specifically, DeiT-T and DeiT-S with MCTF reduce FLOPs by about 44% while improving the performance (+0.5%, and +0.3%) over the base model, respectively. We also demonstrate the applicability of MCTF in various Vision Transformers (e.g., T2T-ViT, LV-ViT), achieving at least 31% speedup without performance degradation.

The large-scale visual pretraining has significantly improve the performance of large vision models. However, we observe the \emph{low FLOPs pitfall} that the existing low-FLOPs models cannot benefit from large-scale pretraining. In this paper, we introduce a novel design principle, termed ParameterNet, aimed at augmenting the number of parameters in large-scale visual pretraining models while minimizing the increase in FLOPs. We leverage dynamic convolutions to incorporate additional parameters into the networks with only a marginal rise in FLOPs. The ParameterNet approach allows low-FLOPs networks to take advantage of large-scale visual pretraining. Furthermore, we extend the ParameterNet concept to the language domain to enhance inference results while preserving inference speed. Experiments on the large-scale ImageNet-22K have shown the superiority of our ParameterNet scheme. For example, ParameterNet-600M can achieve higher accuracy than the widely-used Swin Transformer (81.6\% \emph{vs.} 80.9\%) and has much lower FLOPs (0.6G \emph{vs.} 4.5G). The code will be released at \url{https://parameternet.github.io/}.

Diffusion models have recently gained unprecedented attention in the field of image synthesis due to their remarkable generative capabilities. Notwithstanding their prowess, these models often incur substantial computational costs, primarily attributed to the sequential denoising process and cumbersome model size. Traditional methods for compressing diffusion models typically involve extensive retraining, presenting cost and feasibility challenges. In this paper, we introduce DeepCache, a novel training-free paradigm that accelerates diffusion models from the perspective of model architecture. DeepCache capitalizes on the inherent temporal redundancy observed in the sequential denoising steps of diffusion models, which caches and retrieves features across adjacent denoising stages, thereby curtailing redundant computations. Utilizing the property of the U-Net, we reuse the high-level features while updating the low-level features in a very cheap way. This innovative strategy, in turn, enables a speedup factor of 2.3$\times$ for Stable Diffusion v1.5 with only a 0.05 decline in CLIP Score, and 4.1$\times$ for LDM-4-G with a slight decrease of 0.22 in FID on ImageNet. Our experiments also demonstrate DeepCache's superiority over existing pruning and distillation methods that necessitate retraining and its compatibility with current sampling techniques. Furthermore, we find that under the same throughput, DeepCache effectively achieves comparable or even marginally improved results with DDIM or PLMS.

This work presents Adaptive Local-then-Global Merging (ALGM), a token reduction method for semantic segmentation networks that use plain Vision Transformers.ALGM merges tokens in two stages: (1) In the first network layer, it merges similar tokens within a small local window and (2) halfway through the network, it merges similar tokens across the entire image. This is motivated by an analysis in which we found that, in those situations, tokens with a high cosine similarity can likely be merged without a drop in segmentation quality. With extensive experiments across multiple datasets and network configurations, we show that ALGM not only significantly improves the throughput by up to 100%, but can also enhance the mean IoU by up to +1.1, thereby achieving a better trade-off between segmentation quality and efficiency than existing methods. Moreover, our approach is adaptive during inference, meaning that the same model can be used for optimal efficiency or accuracy, depending on the application.

The remarkable performance of Vision Transformers (ViTs) typically requires an extremely large training cost. Existing methods have attempted to accelerate the training of ViTs, yet typically disregard method universality with accuracy dropping.Meanwhile, they break the training consistency of the original transformers, including the consistency of hyper-parameters, architecture, and strategy, which prevents them from being widely applied to different Transformer networks. In this paper, we propose a novel token growth scheme Token Expansion (termed ToE) to achieve consistent training acceleration for ViTs. We introduce an ``initialization-expansion-merging'' pipeline to maintain the integrity of the intermediate feature distribution of original transformers, preventing the loss of crucial learnable information in the training process. ToE can not only be seamlessly integrated into the training and fine-tuning process of transformers (e.g., DeiT and LV-ViT), but also effective for efficient training frameworks (e.g., EfficientTrain), without twisting the original training hyper-parameters, architecture, and introducing additional training strategies. Extensive experiments demonstrate that ToE achieves about 1.3x faster for the training of ViTs in a lossless manner, or even with performance gains over the full-token training baselines.

Dataset distillation reduces the storage and computational consumption of training a network by generating a small surrogate dataset that encapsulates rich information of the original large-scale one. However, previous distillation methods heavily rely on the sample-wise iterative optimization scheme. As the images-per-class (IPC) setting or image resolution grows larger, the necessary computation will demand overwhelming time and resources. In this work, we intend to incorporate generative diffusion techniques for computing the surrogate dataset. Observing that key factors for constructing an effective surrogate dataset are representativeness and diversity, we design additional minimax criteria in the generative training to enhance these facets for the generated images of diffusion models. We present a theoretical model of the process as hierarchical diffusion control demonstrating the flexibility of the diffusion process to target these criteria without jeopardizing the faithfulness of the sample to the desired distribution. The proposed method achieves state-of-the-art validation performance while demanding much less computational resources. Under the 100-IPC setting on ImageWoof, our method requires less than one-twentieth the distillation time of previous methods, yet yields even better performance. Source code and generated data are available in https://github.com/vimar-gu/MinimaxDiffusion.

Recent transformer-based architectures have shown impressive results in the field of image segmentation. Thanks to their flexibility, they obtain outstanding performance in multiple segmentation tasks, such as semantic and panoptic, under a single unified framework. To achieve such impressive performance, these architectures employ intensive operations and require substantial computational resources, which are often not available, especially on edge devices.To fill this gap, we propose Prototype-based Efficient MaskFormer (PEM), an efficient transformer-based architecture that can operate in multiple segmentation tasks. PEM proposes a novel prototype-based cross-attention which leverages the redundancy of visual features to restrict the computation and improve the efficiency without harming the performance. In addition, PEM introduces an efficient multi-scale feature pyramid network, capable of extracting features that have high semantic content in an efficient way, thanks to the combination of deformable convolutions and implicit attention mechanisms.We benchmark the proposed PEM architecture on two tasks, semantic and panoptic segmentation, evaluated on two different datasets, Cityscapes and ADE20K. PEM demonstrates outstanding performance on every task and dataset, outperforming task-specific architectures while being comparable and even better than computationally-expensive baselines. The code will be released upon acceptance.

Recent success of pre-trained foundation vision-language models makes Open-Vocabulary Segmentation (OVS) possible. Despite the promising performance, this approach introduces heavy computational overheads for two challenges: 1) large model sizes of the backbone; 2) expensive costs during the fine-tuning. These challenges hinder this OVS strategy from being widely applicable and affordable in real-world scenarios. Although traditional methods such as model compression and efficient fine-tuning can address these challenges, they often rely on heuristics. This means that their solutions cannot be easily transferred and necessitate re-training on different models, which comes at a cost. In the context of efficient OVS, we target achieving performance that is comparable to or even better than prior OVS works based on large vision-language foundation models, by utilizing smaller models that incur lower training costs. The core strategy is to make our efficiency principled and thus seamlessly transferable from one OVS framework to others without further customization. Comprehensive experiments on diverse OVS benchmarks demonstrate our superior trade-off between segmentation accuracy and computation costs over previous works.

Few-shot model compression aims to compress a large model into a more compact one with only a tiny training set (even without labels). Block-level pruning has recently emerged as a leading technique in achieving high accuracy and low latency in few-shot CNN compression. But, few-shot compression for Vision Transformers (ViT) remains largely unexplored, which presents a new challenge. In particular, the issue of sparse compression exists in traditional CNN few-shot methods, which can only produce very few compressed models of different model sizes. This paper proposes a novel framework for few-shot ViT compression named DC-ViT. Instead of dropping the entire block, DC-ViT selectively eliminates the attention module while retaining and reusing portions of the MLP module. DC-ViT enables dense compression, which outputs numerous compressed models that densely populate the range of model complexity. DC-ViT outperforms state-of-the-art few-shot compression methods by a significant margin of 10 percentage points, along with lower latency in the compression of ViT and its variants.

Large pretrained models are increasingly crucial in modern computer vision tasks. These models are typically used in downstream tasks by end-to-end finetuning, which is highly memory-intensive for tasks with high-resolution data, e.g., video understanding, small object detection, and point cloud analysis. In this paper, we propose Dynamic Reversible Dual-Residual Networks, or Dr2Net, a novel family of network architectures that acts as a surrogate network to finetune a pretrained model with substantially reduced memory consumption. Dr2Net contains two types of residual connections, one maintaining the residual structure in the pretrained models, and the other making the network reversible. Due to its reversibility, intermediate activations, which can be reconstructed from output, are cleared from memory during training. We use two coefficients on either type of residual connections respectively, and introduce a dynamic training strategy that seamlessly transitions the pretrained model to a reversible network with much higher numerical precision. We evaluate Dr2Net on various pretrained models and various tasks, and show that it can reach comparable performance to conventional finetuning but with significantly less memory usage.

N:M sparsity has received increasing attention due to its remarkable performance and latency trade-off compared with structured and unstructured sparsity. However, existing N:M sparsity methods do not differentiate the relative importance of weights among blocks and leave important weights underappreciated. Besides, they directly apply N:M sparsity to the whole network, which will cause severe information loss. Thus, they are still sub-optimal. In this paper, we propose an efficient and effective Multi-Axis Query methodology, dubbed as MaxQ, to rectify these problems. During the training, MaxQ employs a dynamic approach to generate soft N:M masks, considering the weight importance across multiple axes. This method enhances the weights with more importance and ensures more effective updates. Meanwhile, a sparsity strategy that gradually increases the percentage of N:M weight blocks is applied, which allows the network to heal from the pruning-induced damage progressively. During the runtime, the N:M soft masks can be precomputed as constants and folded into weights without causing any distortion to the sparse pattern and incurring additional computational overhead. Comprehensive experiments demonstrate that MaxQ achieves consistent improvements across diverse CNN architectures in various computer vision tasks, including image classification, object detection and instance segmentation. For ResNet50 with 1:16 sparse pattern, MaxQ can achieve 74.6\% top-1 accuracy on ImageNet and improve by over 2.8\% over the state-of-the-art. Codes and checkpoints are available at \url{https://github.com/JingyangXiang/MaxQ}.

Quantization is of significance for compressing the over-parameterized deep neural models and deploying them on resource-limited devices. Fixed-precision quantization suffers from performance drop due to the limited numerical representation ability. Conversely, mixed-precision quantization (MPQ) is advocated to compress the model effectively by allocating heterogeneous bit-width for layers. MPQ is typically organized into a searching-retraining two-stage process. Previous works only focus on determining the optimal bit-width configuration in the first stage efficiently, while ignoring the considerable time costs in the second stage. However, retraining always consumes hundreds of GPU-hours on the cutting-edge GPUs, thus hindering deployment efficiency significantly. In this paper, we devise a one-shot training-searching paradigm for mixed-precision model compression. Specifically, in the first stage, all potential bit-width configurations are coupled and thus optimized simultaneously within a set of shared weights. However, our observations reveal a previously unseen and severe bit-width interference phenomenon among highly coupled weights during optimization, leading to considerable performance degradation under a high compression ratio. To tackle this problem, we first design a bit-width scheduler to dynamically freeze the most turbulent bit-width of layers during training, to ensure the rest bit-widths converged properly. Then, taking inspiration from information theory, we present an information distortion mitigation technique to align the behaviour of the bad-performing bit-widths to the well-performing ones. In the second stage, an inference-only greedy search scheme is devised to evaluate the goodness of configurations without introducing any additional training costs. Extensive experiments on three representative models and three datasets demonstrate the effectiveness of the proposed method.

Deep learning models, particularly those based on transformers, often employ numerous stacked structures, which possess identical architectures and perform similar functions. While effective, this stacking paradigm leads to a substantial increase in the number of parameters, posing challenges for practical applications. In today's landscape of increasingly large models, stacking depth can even reach dozens, further exacerbating this issue. To mitigate this problem, we introduce LORS (LOw-rank Residual Structure). LORS allows stacked modules to share the majority of parameters, requiring a much smaller number of unique ones per module to match or even surpass the performance of using entirely distinct ones, thereby significantly reducing parameter usage. We validate our method by applying it to the stacked decoders of a query-based object detector, and conduct extensive experiments on the widely used MS COCO dataset. Experimental results demonstrate the effectiveness of our method, as even with a 70\% reduction in the parameters of the decoder, our method still enables the model to achieve comparable or even better performance than its original.

We develop a novel vectorized image representation scheme accommodating both shape/geometry and texture in a decoupled way, particularly tailored for reconstruction and editing tasks of artistic/design images such as Emojis and Cliparts. In the heart of this representation is a set of sparsely and unevenly located 2D control points. On one hand, these points constitute a collection of parametric/vectorized geometric primitives (e.g., curves and closed shapes) describing the shape characteristics of the target image. On the other hand, local texture codes, in terms of implicit neural network parameters, are spatially distributed into each control point, yielding local coordinate-to-RGB mappings within the anchored region of each control point. In the meantime, a zero-shot learning algorithm is developed to decompose an arbitrary raster image into the above representation, for the sake of high-fidelity image vectorization with convenient editing ability. Extensive experiments on a series of image vectorization and editing tasks well demonstrate the high accuracy offered by our proposed method, with a significantly higher image compression ratio over prior art.

We introduce SynCLR, a novel approach for learning visual representations exclusively from synthetic images, without any real data. We synthesize a large dataset of image captions using LLMs, then use an off-the-shelf text-to-image model to generate multiple images corresponding to each synthetic caption. We perform visual representation learning on these synthetic images via contrastive learning, treating images sharing the same caption as positive pairs. The resulting representations demonstrate remarkable transferability, competing favorably with other general-purpose visual representation learners such as CLIP and DINO v2 in image classification tasks. Furthermore, in dense prediction tasks such as semantic segmentation, SynCLR outperforms previous self-supervised methods by a significant margin, e.g., improving over MAE and iBOT by 5.0 and 3.1 mIoU on ADE20k for ViT-B/16.

Multi-task learning for dense prediction has emerged as a pivotal area in computer vision, enabling simultaneous processing of diverse yet interrelated pixel-wise prediction tasks. However, the substantial computational demands of state-of-the-art (SoTA) models often limit their widespread deployment. This paper addresses this challenge by introducing network binarization to compress resource-intensive multi-task dense predictors. Specifically, our goal is to significantly accelerate multi-task dense prediction models via Binary Neural Networks (BNNs) while maintaining and even improving model performance at the same time. To reach this goal, we propose a Binary Multi-task Dense Predictor, Bi-MTDP, and several variants of \bimtdp, in which a multi-task dense predictor is constructed via specified binarized modules. Our systematical analysis of this predictor reveals that performance drop from binarization is primarily caused by severe information degradation. To address this issue, we introduce a deep information bottleneck layer that enforces representations for downstream tasks satisfying Gaussian distribution in forward propagation. Moreover, we introduce a knowledge distillation mechanism to correct the direction of information flow in backward propagation. Intriguingly, one variant of Bi-MTDP outperforms full-precision (FP) multi-task dense prediction SoTAs, ARTC (CNN-based) and InvPT (ViT-based). This result indicates that Bi-MTDP is not merely a naive trade-off between performance and efficiency, but is rather a benefit of the redundant information flow thanks to the multi-task architecture.

Recently, lightweight Vision Transformers (ViTs) demonstrate superior performance and lower latency, compared with lightweight Convolutional Neural Networks (CNNs), on resource-constrained mobile devices. Researchers have discovered many structural connections between lightweight ViTs and lightweight CNNs. However, the notable architectural disparities in the block structure, macro, and micro designs between them have not been adequately examined. In this study, we revisit the efficient design of lightweight CNNs from ViT perspective and emphasize their promising prospect for mobile devices. Specifically, we incrementally enhance the mobile-friendliness of a standard lightweight CNN, \ie, MobileNetV3, by integrating the efficient architectural designs of lightweight ViTs. This ends up with a new family of pure lightweight CNNs, namely RepViT. Extensive experiments show that RepViT outperforms existing state-of-the-art lightweight ViTs and exhibits favorable latency in various vision tasks. Notably, on ImageNet, RepViT achieves over 80\% top-1 accuracy with 1.0 ms latency on an iPhone 12, which is the first time for a lightweight model, to the best of our knowledge. Besides, when RepViT meets SAM, our RepViT-SAM can achieve nearly 10$\times$ faster inference than the advanced MobileSAM. Codes and models are available at https://github.com/THU-MIG/RepViT.

Post-training quantization (PTQ) converts a pre-trained full-precision (FP) model into a quantized model in a training-free manner. Determining suitable quantization parameters, such as scaling factors and weight rounding, is the primary strategy for mitigating the impact of quantization noise (calibration) and restoring the performance of the quantized models. However, the existing activation calibration methods have never considered information degradation between pre- (FP) and post-quantized activations. In this study, we introduce a well-defined distributional metric from information theory, mutual information, into PTQ calibration. We aim to calibrate the quantized activations by maximizing the mutual information between the pre- and post-quantized activations. To realize this goal, we establish a contrastive learning (CL) framework for the PTQ calibration, where the quantization parameters are optimized through a self-supervised proxy task. Specifically, by leveraging CL during the PTQ process, we can benefit from pulling the positive pairs of quantized and FP activations collected from the same input samples, while pushing negative pairs from different samples. Thanks to the ingeniously designed critic function, we avoid the unwanted but often-encountered collision solution in CL, especially in calibration scenarios where the amount of calibration data is limited. Additionally, we provide a theoretical guarantee that minimizing our designed loss is equivalent to maximizing the desired mutual information. Consequently, the quantized activations retain more information, which ultimately enhances the performance of the quantized network. Experimental results show that our method can effectively serve as an add-on module to existing SoTA PTQ methods.

Knowledge distillation (KD) has been applied to various tasks successfully, and mainstream methods typically boost the student model via spatial imitation losses. However, the consecutive downsamplings induced in the spatial domain of teacher model is a type of corruption, hindering the student from analyzing what specific information needs to be imitated, which results in accuracy degradation. To better understand the underlying pattern of corrupted feature maps, we shift our attention to the frequency domain. During frequency distillation, we encounter a new challenge: the low-frequency bands convey general but minimal context, while the high are more informative but also introduce noise. Not each pixel within the frequency bands contributes equally to the performance. To address the above problem: (1) We propose the Frequency Prompt plugged into the teacher model, absorbing the semantic frequency context during finetuning. (2) During the distillation period, a pixel-wise frequency mask is generated via Frequency Prompt, to localize those pixel of interests (PoIs) in various frequency bands. Additionally, we employ a position-aware relational frequency loss for dense prediction tasks, delivering a high-order spatial enhancement to the student model. We dub our frequency knowledge distillation method as FreeKD, which determines the optimal localization and extent for the frequency distillation. Extensive experiments demonstrate that FreeKD not only outperforms spatial-based distillation methods consistently on dense prediction tasks (e.g., FreeKD brings 3.8 AP gains for RepPoints-R50 on COCO2017 and 4.55 mIoU gains for PSPNet-R18 on Cityscapes), but also conveys more robustness to the student. Notably, we also validate the generalization of our approach on large-scale vision models (e.g., DINO and SAM).

Segment Anything Model (SAM) has achieved impressive performance in many computer vision tasks. However, as a large-scale model, the immense memory and computation costs hinder its practical deployment. In this paper, we propose a post-training quantization (PTQ) framework for Segment Anything Model, namely PTQ4SAM. First, we investigate the inherent bottleneck of SAM quantization attributed to the bimodal distribution in post-Key-Linear activations. We analyze its characteristics from both per-tensor and per-channel perspectives, and propose a Bimodal Integration strategy, which utilizes a mathematically equivalent sign operation to transform the bimodal distribution into a relatively easy-quantized normal distribution offline. Second, SAM encompasses diverse attention mechanisms (i.e., self-attention and two-way cross-attention), resulting in substantial variations in the post-Softmax distributions. Therefore, we introduce an Adaptive Granularity Quantization for Softmax through searching the optimal power-of-two base, which is hardware-friendly. Extensive experimental results across various vision tasks (instance segmentation, semantic segmentation and object detection), datasets and model variants show the superiority of PTQ4SAM. For example, when quantizing SAM-L to 6-bit, we achieve lossless accuracy for instance segmentation, about 0.5\% drop with theoretical 3.9$\times$ acceleration. The code is available at https://github.com/chengtao-lv/PTQ4SAM.

Contrastive Language-Image Pre-training (CLIP) has become a promising language-supervised visual pre-training framework. This paper aims to distill small CLIP models supervised by a large teacher CLIP model. We propose several distillation strategies, including relation, feature, gradient and contrastive paradigms, to examine the effectiveness of CLIP-Knowledge Distillation (KD). We show that a simple feature mimicry with Mean Squared Error loss works surprisingly well. Moreover, interactive contrastive learning across teacher and student encoders is also effective in performance improvement. We explain that the success of CLIP-KD can be attributed to maximizing the feature similarity between teacher and student. The unified method is applied to distill several student models trained on CC3M+12M. CLIP-KD improves student CLIP models consistently over zero-shot ImageNet classification and cross-modal retrieval benchmarks. When using ViT-L/14 pretrained on Laion-400M as the teacher, CLIP-KD achieves 57.5% and 55.4% zero-shot top-1 ImageNet accuracy over ViT-B/16 and ResNet-50, surpassing the original CLIP without KD by 20.5% and 20.1% margins, respectively. Our code is released on https://github.com/winycg/CLIP-KD.

Contrastive pretraining of image-text foundation models, such as CLIP, demonstrated excellent zero-shot performance and improved robustness on a wide range of downstream tasks. However, the utilization of large transformer-based text and image encoders in these models introduces significant memory and latency overhead, posing challenges for deployment on mobile devices. In this work, we introduce MobileCLIP, a new family of efficient image-text models optimized for runtime performance along with a novel, efficient, and highly effective training approach, referred to as multi-modal reinforced training. The proposed training approach leverages knowledge transfer from an image captioning model and an ensemble of strong CLIP encoders to improve the accuracy of efficient models, while avoiding training time compute overhead by storing the additional knowledge in a reinforced dataset. MobileCLIP sets a new state-of-the-art latency-accuracy tradeoff for zero-shot classification and retrieval tasks on several datasets. Our MobileCLIP-S2 variant is 2.3x faster while more accurate in terms of average zero-shot performance when compared to previous best ViT-B/16 based CLIP model. We also demonstrate the effectiveness of the proposed multi-modal reinforced training in isolation by training a CLIP model with standard ViT-B/16 image backbone and achieving +2.9% average performance improvement compared to the previous best on OpenCLIP 38 evaluation benchmarks. Moreover, we show that the proposed approach achieves 10x-1000x improved learning efficiency when compared with non-reinforced CLIP training.

Logit knowledge distillation attracts increasing attention due to its practicality in recent studies. However, it often suffers inferior performance compared to the feature knowledge distillation. In this paper, we argue that existing logit-based methods may be sub-optimal since they only leverage the global logit output that couples multiple semantic knowledge. This may transfer ambiguous knowledge to the student and mislead its learning. To this end, we propose a simple but effective method, i.e., Scale Decoupled Distillation (SDD), for logit knowledge distillation. SDD decouples the global logit output into multiple local logit outputs and establishes distillation pipelines for them. This helps the student to mine and inherit fine-grained and unambiguous logit knowledge. Moreover, the decoupled knowledge can be further divided into consistent and complementary logit knowledge that transfers the semantic information and sample ambiguity, respectively. By increasing the weight of complementary parts, SDD can guide the student to focus more on ambiguous samples, improving its discrimination ability. Extensive experiments on several benchmark datasets demonstrate the effectiveness of SDD for wide teacher-student pairs, especially in the fine-grained classification task.

We propose an efficient abnormal event detection model based on a lightweight masked auto-encoder (AE) applied at the video frame level. The novelty of the proposed model is threefold. First, we introduce an approach to weight tokens based on motion gradients, thus shifting the focus from the static background scene to the foreground objects. Second, we integrate a teacher decoder and a student decoder into our architecture, leveraging the discrepancy between the outputs given by the two decoders to improve anomaly detection. Third, we generate synthetic abnormal events to augment the training videos, and task the masked AE model to jointly reconstruct the original frames (without anomalies) and the corresponding pixel-level anomaly maps. Our design leads to an efficient and effective model, as demonstrated by the extensive experiments carried out on four benchmarks: Avenue, ShanghaiTech, UBnormal and UCSD Ped2. The empirical results show that our model achieves an excellent trade-off between speed and accuracy, obtaining competitive AUC scores, while processing 1655 FPS. Hence, our model is between 8 and 70 times faster than competing methods. We also conduct an ablation study to justify our design. Our code is freely available at: http//github.com/link.hidden.for.review.

Low-precision quantization is recognized for its efficacy in neural network optimization.Our analysis reveals that non-quantized elementwise operations which are prevalent in layers such as parameterized activation functions, batch normalization, and quantization scaling dominate the inference cost of low-precision models.These non-quantized elementwise operations are commonly overlooked in SOTA efficiency metrics such as Arithmetic Computation Effort (*ACE*).In this paper, we propose $ACE_{v2}$ - an extended version of *ACE* which offers a better alignment with the inference cost of quantized models and their energy consumption on ML hardware.Moreover, we introduce *PikeLPN*, a model that addresses these efficiency issues by applying quantization to both elementwise operations and multiply-accumulate operations. In particular, we present a novel quantization technique for batch normalization layers named *QuantNorm* which allows for quantizing the batch normalization parameters without compromising the model performance. Additionally, we propose applying *Double Quantization* where the quantization scaling parameters are quantized. Furthermore, we recognize and resolve the issue of distribution mismatch in Separable Convolution layers by introducing *Distribution-Heterogeneous Quantization* which enables quantizing them to low-precision.*PikeLPN* achieves Pareto-optimality in efficiency-accuracy trade-off with up to *3*$\times$ efficiency improvement compared to SOTA low-precision models.

Existing Knowledge Distillation (KD) methods typically focus on transferring knowledge from a large-capacity teacher to a low-capacity student model, achieving substantial success in unimodal knowledge transfer. However, existing methods can hardly be extended to Cross-Modal Knowledge Distillation (CMKD), where the knowledge is transferred from a teacher modality to a different student modality, with inference only on the distilled student modality. We empirically reveal that the modality gap, i.e., modality imbalance and soft label misalignment, incurs the ineffectiveness of traditional KD in CMKD. As a solution, we propose a novel \textbf{\underline{C}}ustomized \textbf{\underline{C}}rossmodal \textbf{\underline{K}}nowledge \textbf{\underline{D}}istillation (C$^2$KD). Specifically, to alleviate the modality gap, the pre-trained teacher performs bidirectional distillation with the student to provide customized knowledge. The On-the-Fly Selection Distillation(OFSD) strategy is applied to selectively filter out the samples with misaligned soft labels, where we distill cross-modal knowledge from non-target classes to avoid the modality imbalance issue. To further provide receptive cross-modal knowledge, proxy student and teacher, inheriting unimodal and cross-modal knowledge, is formulated to progressively transfer cross-modal knowledge through bidirectional distillation. Experimental results on audio-visual, image-text, and RGB-depth datasets demonstrate that our method can effectively transfer knowledge across modalities, achieving superior performance against traditional KD by a large margin.

DETR is a novel end-to-end transformer architecture object detector, which significantly outperforms classic detectors when scaling up. In this paper, we focus on the compression of DETR with knowledge distillation. While knowledge distillation has been well-studied in classic detectors, there is a lack of researches on how to make it work effectively on DETR. We first provide experimental and theoretical analysis to point out that the main challenge in DETR distillation is the lack of consistent distillation points. Distillation points refer to the corresponding inputs of the predictions for student to mimic, which have different formulations in CNN detector and DETR, and reliable distillation requires sufficient distillation points which are consistent between teacher and student. Based on this observation, we propose the first general knowledge distillation paradigm for DETR(KD-DETR) with consistent distillation points sampling, for both homogeneous and heterogeneous distillation. Specifically, we decouple detection and distillation tasks by introducing a set of specialized object queries to construct distillation points for DETR. We further propose a general-to-specific distillation points sampling strategy to explore the extensibility of KD-DETR. Extensive experiments validate the effectiveness and generalization of KD-DETR. For both single-scale DAB-DETR and multis-scale Deformable DETR and DINO, KD-DETR boost the performance of student model with improvements of 2.6\%-5.2\%. We further extend KD-DETR to heterogeneous distillation, and achieves 2.1\% improvement by distilling the knowledge from DINO to Faster R-CNN with ResNet-50, which is comparable with homogeneous distillation methods.

In this paper, we propose an accurate post-training quantization framework of diffusion models (APQ-DM) for efficient image generation. Conventional quantization frameworks learn shared quantization functions for tensor discretization regardless of the generation timesteps in diffusion models, while the activation distribution differs significantly across various timesteps. Meanwhile, the calibration images are acquired in random timesteps which fail to provide sufficient information for generalizable quantization function learning. Both issues cause sizable quantization errors with obvious image generation performance degradation. On the contrary, we design distribution-aware quantization functions for activation discretization in different timesteps and search the optimal timesteps for informative calibration image generation, so that our quantized diffusion model can reduce the discretization errors with negligible computational overhead. Specifically, we partition various timestep quantization functions into different groups according to the importance weights, which are optimized by differentiable search algorithms. We also extend structural risk minimization principle for informative calibration image generation to enhance the generalization ability in the deployment of quantized diffusion model. Extensive experimental results show that our method outperforms the state-of-the-art post-training quantization of diffusion model by a sizable margin with similar computational cost.

To achieve greater accuracy, hypergraph matching algorithms require exponential increases in computational resources. Recent kd-tree-based approximate nearest neighbor (ANN) methods, despite the sparsity of their compatibility tensor, still require exhaustive calculations for large-scale graph matching. This work utilizes CUR tensor decomposition and introduces a novel cascaded second and third-order hypergraph matching framework (CURSOR) for efficient hypergraph matching. A CUR-based second-order graph matching algorithm is used to provide a rough match, and then the core of CURSOR, a fiber-CUR-based tensor generation method, directly calculates entries of the compatibility tensor by leveraging the initial second-order match result. This significantly decreases the time complexity and tensor density. A probability relaxation labeling (PRL)-based matching algorithm, especially suitable for sparse tensors, is developed. Experiment results on large-scale synthetic datasets and widely-adopted benchmark sets demonstrate the superiority of CURSOR over existing methods. The tensor generation method in CURSOR can be integrated seamlessly into existing hypergraph matching methods to improve their performance and lower their computational costs.

Training a linear classifier or lightweight model on top of pretrained vision model outputs, so-called 'frozen features', leads to impressive performance on a number of downstream few-shot tasks. Currently, frozen features are not modified during training. On the other hand, when networks are trained directly on images, data augmentation is a standard recipe that improves performance with no substantial overhead. In this paper, we conduct an extensive pilot study on few-shot image classification that explores applying data augmentations in the frozen feature space, dubbed 'frozen feature augmentation (FroFA)', covering twenty augmentations in total. Our study demonstrates that adopting a deceptively simple pointwise FroFA, such as brightness, can improve few-shot performance consistently across three network architectures, three large pretraining datasets, and eight transfer datasets.

Structural model pruning is a prominent approach used for reducing the computational cost of Convolutional Neural Networks (CNNs) before their deployment on resource-constrained devices. Yet, the majority of proposed ideas require a pretrained model before pruning, which is costly to secure. In this paper, we propose a novel structural pruning approach to jointly learn the weights and structurally prune architectures of CNN models. The core element of our method is a Reinforcement Learning (RL) agent whose actions determine the pruning ratios of the CNN model's layers, and the resulting model's accuracy serves as its reward. We conduct the joint training and pruning by iteratively training the model's weights and the agent's policy, and we regularize the model's weights to align with the selected structure by the agent. The evolving model's weights result in a dynamic reward function for the agent, which prevents using prominent episodic RL methods with stationary environment assumption for our purpose. We address this challenge by designing a mechanism to model the complex changing dynamics of the reward function and provide a representation of it to the RL agent. To do so, we take a learnable embedding for each training epoch and employ a recurrent model to calculate a representation of the changing environment. We train the recurrent model and embeddings using a decoder model to reconstruct observed rewards. Such a design empowers our agent to effectively leverage episodic observations along with the environment representations to learn a proper policy to determine performant sub-networks of the CNN model. Our extensive experiments on CIFAR-10 and ImageNet using ResNets and MobileNets demonstrate the effectiveness of our method.

Deployment of Transformer models on edge devices is becoming increasingly challenging due to the exponentially growing inference cost that scales quadratically with the number of tokens in the input sequence. Token pruning is an emerging solution to address this challenge due to its ease of deployment on various Transformer backbones. However, most token pruning methods require computationally expensive fine-tuning, which is undesirable in many edge deployment cases. In this work, we propose Zero-TPrune, the first zero-shot method that considers both the importance and similarity of tokens in performing token pruning. It leverages the attention graph of pre-trained Transformer models to produce an importance distribution for tokens via our proposed Weighted Page Rank (WPR) algorithm. This distribution further guides token partitioning for efficient similarity-based pruning. Due to the elimination of the fine-tuning overhead, Zero-TPrune can prune large models at negligible computational cost, switch between different pruning configurations at no computational cost, and perform hyperparameter tuning efficiently. We evaluate the performance of Zero-TPrune on vision tasks by applying it to various vision Transformer backbones and testing them on ImageNet. Without any fine-tuning, Zero-TPrune reduces the FLOPs cost of DeiT-S by 34.7\% and improves its throughput by 45.3\% with only 0.4\% accuracy loss. Compared with state-of-the-art pruning methods that require fine-tuning, Zero-TPrune not only eliminates the need for fine-tuning after pruning but also does so with only 0.1\% accuracy loss. Compared with state-of-the-art fine-tuning-free pruning methods, Zero-TPrune reduces accuracy loss by up to 49\% with the same or higher throughput.

Diffusion models (DMs) have exhibited superior performance in generating high-quality and diverse images. However, this exceptional performance comes at the cost of expensive architectural design, particularly with the attention module heavily used in leading models.Existing works mainly adopt a retraining process under one specific compression ratio to enhance the efficiency which is computational expensive and less scalable. To this end, we introduce the Attention-driven Training-free Efficient Diffusion Model (AT-EDM), a framework that leverages attention maps to perform run-time pruning of redundant tokens during inference without retraining. Specifically, we develop a single denoising step pruning strategy to identify redundant tokens and a similarity-based recovery method to restore tokens for convolution operation.Additionally, a novel pruning schedule is proposed to adjust the pruning budget across different denoising timesteps for better generation quality.Extensive evaluations show that our method performs favorably against prior arts in terms of efficiency (e.g., 38.8\% FLOPs saving over Stable Diffusion XL) while maintaining nearly the same FID and CLIP scores with the full model.

Most existing dynamic or runtime channel pruning methods have to store all weights to achieve efficient inference, which brings extra storage costs. Static pruning methods can reduce storage costs directly, but their performance is limited by using a fixed sub-network to approximate the original model. Most existing pruning works suffer from these drawbacks because they were designed to only conduct either static or dynamic pruning. In this paper, we propose a novel method to solve both efficiency and storage challenges via simultaneously conducting dynamic and static channel pruning for convolutional neural networks. We propose a new bi-level optimization based model to naturally integrate the static and dynamic channel pruning. By doing so, our method enjoys benefits from both sides, and the disadvantages of dynamic and static pruning are reduced. After pruning, we permanently remove redundant parameters and then finetune the model with dynamic flexibility. Experimental results on CIFAR-10 and ImageNet datasets suggest that our method can achieve state-of-the-art performance compared to existing dynamic and static channel pruning methods.

Parameter-efficient fine-tuning for pre-trained Vision Transformers aims to adeptly tailor a model to downstream tasks by learning a minimal set of new adaptation parameters while preserving the frozen majority of pre-trained parameters. Striking a balance between retaining the generalizable representation capacity of the pre-trained model and acquiring task-specific features poses a key challenge. Currently, there is a lack of focus on guiding this delicate trade-off. In this study, we approach the problem from the perspective of Singular Value Decomposition (SVD) of pre-trained parameter matrices, providing insights into the tuning dynamics of existing methods. Building upon this understanding, we propose a Residual-based Low-Rank Rescaling (RLRR) fine-tuning strategy. This strategy not only enhances flexibility in parameter tuning but also ensures that new parameters do not deviate excessively from the pre-trained model through a residual design. Extensive experiments demonstrate that our method achieves competitive performance across various downstream image classification tasks, all while maintaining comparable new parameters. We believe this work takes a step forward in offering a unified perspective for interpreting existing methods and serves as motivation for the development of new approaches that move closer to effectively considering the crucial trade-off mentioned above. Our code is available at https://github.com/zstarN70/RLRR.git.

Segment Anything Model (SAM) has emerged as a powerful tool for numerous vision applications. A key component that drives the impressive performance for zero-shot transfer and high versatility is a super large Transformer model trained on the extensive high-quality SA-1B dataset. While beneficial, the huge computation cost of SAM model has limited its applications to wider real-world applications. To address this limitation, we propose EfficientSAMs, light-weight SAM models that exhibits decent performance with largely reduced complexity. Our idea is based on leveraging masked image pretraining, SAMI, which learns to reconstruct features from SAM image encoder for effective visual representation learning. Further, we take SAMI-pretrained light-weight image encoders and mask decoder to build EfficientSAMs, and finetune the models on SA-1B for segment anything task. We perform evaluations on multiple vision tasks including image classification, object detection, instance segmentation, and semantic object detection, and find that our proposed pretraining method, SAMI, consistently outperforms other masked image pretraining methods. On segment anything task such as zero-shot instance segmentation, our EfficientSAMs with SAMI-pretrained lightweight image encoders perform favorably with a significant gain (e.g., ~4 AP on COCO/LVIS) over other fast SAM models.

In recent years, there has been significant progress in the development of text-to-image generative models. Evaluating the quality of the generative models is one essential step in the development process. Unfortunately, the evaluation process could consume a significant amount of computational resources, making the required periodic evaluation of model performance (e.g., monitoring training progress) impractical. Therefore, we seek to improve the evaluation efficiency by selecting the representative subset of the text-image dataset.We systematically investigate the design choices, including the section criteria (textural features or image-based metrics) and the selection granularity (prompt-level or set-level). We find that the insights from prior work on subset selection for training data do not generalize to this problem, and we propose FlashEval, an iterative search algorithm tailored to evaluation data selection. We demonstrate the effectiveness of FlashEval on ranking diffusion models with various configurations including architectures, quantization levels, and sampler schedules on COCO and DiffusionDB datasets.Our searched 50-item subset could achieve comparable evaluation quality to the randomly sampled 500-item subset for COCO annotations on unseen models, achieving a 10x evaluation speedup.We will release the condensed subset of these commonly used datasets to help accelerate and facilitate diffusion algorithm design and evaluation,and open-source FlashEval as a tool for condensing future datasets.

Post-training quantization (PTQ) is an efficient model compression technique that quantizes a pretrained full-precision model using only a small calibration set of unlabeled samples without retraining. PTQ methods for convolutional neural networks (CNNs) provide quantization results comparable to full-precision counterparts. Directly applying them to vision transformers (ViTs), however, incurs severe performance degradation, mainly due to the differences in architectures between CNNs and ViTs. In particular, the distribution of activations for each channel vary drastically according to input instances, making PTQ methods for CNNs inappropriate for ViTs. To address this, we introduce instance-aware group quantization for ViTs (IGQ-ViT). To this end, we propose to split the channels of activation maps into multiple groups dynamically for each input instance, such that activations within each group share similar statistical properties. We also extend our scheme to quantize softmax attentions across tokens. In addition, the number of groups for each layer is adjusted to minimize the discrepancies between predictions from quantized and full-precision models, under a bit-operation (BOP) constraint. We show extensive experimental results on image classification, object detection, and instance segmentation, with various transformer architectures, demonstrating the effectiveness of our approach.

Recent advances in neural network pruning have shown how it is possible to reduce the computational costs and memory demands of deep learning models before training. We focus on this framework and propose a new pruning at initialization algorithm that leverages the Neural Tangent Kernel (NTK) theory to align the training dynamics of the sparse network with that of the dense one. Specifically, we show how the usually neglected data-dependent component in the NTK's spectrum can be taken into account by providing an analytical upper bound to the NTK's trace obtained by decomposing neural networks into individual paths. This leads to our Path eXclusion (PX), a foresight pruning method designed to preserve the parameters that mostly influence the NTK's trace. PX is able to find lottery tickets (i.e. good paths) even at high sparsity levels and largely reduces the need for additional training. When applied to pre-trained models it extracts subnetworks directly usable for several downstream tasks, resulting in performance comparable to those of the dense counterpart but with substantial cost and computational savings.

Multi-task learning has become increasingly popular in the machine learning field, but its practicality is hindered by the need for large, labeled datasets. Most multi-task learning methods depend on fully labeled datasets wherein each input example is accompanied by ground-truth labels for all target tasks. Unfortunately, curating such datasets can be prohibitively expensive and impractical, especially for dense prediction tasks which require per-pixel labels for each image. With this in mind, we propose Joint-Task Regularization (JTR), an intuitive technique which leverages cross-task relations to simultaneously regularize all tasks in a single joint-task latent space to improve learning when data is not fully labeled for all tasks. JTR stands out from existing approaches in that it regularizes all tasks jointly rather than separately in pairs---therefore, it achieves linear complexity relative to the number of tasks while previous methods scale quadratically. To demonstrate the validity of our approach, we extensively benchmark our method across a wide variety of partially labeled scenarios based on NYU-v2, Cityscapes, and Taskonomy.

Current techniques for deep neural network (DNN) pruning often involve intricate multi-step processes that require domain-specific expertise, making their widespread adoption challenging. To address the limitation, the Only-Train-Once (OTO) and OTOv2 are proposed to eliminate the need for additional fine-tuning steps by directly training and compressing a general DNN from scratch. Nevertheless, the static design of optimizers (in OTO) can lead to convergence issues of local optima. In this paper, we proposed the Auto-Train-Once (ATO), an innovative network pruning algorithm designed to automatically reduce the computational and storage costs of DNNs. During the model training phase, our approach not only trains the target model but also leverages a controller network (CN) as an architecture generator to guide the learning of model weights. Furthermore, we developed a novel stochastic gradient algorithm that enhances the coordination between model training and CN training, thereby improving pruning performance. We provide a comprehensive convergence analysis as well as extensive experiments, and the results show that our approach achieves state-of-the-art performance across various model architectures (including ResNet18, ResNet34, ResNet50, ResNet56, and MobileNetv2) on standard benchmark datasets (CIFAR-10, CIFAR-100, and ImageNet).

Although deep learning based object detection is of great significance for various applications, it faces challenges when deployed on edge devices due to the computation and energy limitations.Post-training quantization (PTQ) can improve inference efficiency through integer computing. However, they suffer from severe performance degradation when performing full quantization due to overlooking the unique characteristics of regression tasks in object detection. In this paper, we are the first to explore regression-friendly quantization and evaluate full quantization on various detectors. We reveal the intrinsic reason behind the difficulty of quantizing regressors with empirical and theoretical justifications, and introduce a novel Regression-specialized Post-Training Quantization (Reg-PTQ) scheme. It includes Filtered Global Loss Integration Calibration to combine the global loss with a two-step filtering mechanism, mitigating the adverse impact of false positive bounding boxes; and Learnable Logarithmic-Affine Quantizer tailored for the non-uniform distributed parameters in regression structures. Extensive experiments on prevalent detectors showcase the effectiveness of the well-designed Reg-PTQ. Notably, our Reg-PTQ achieves $7.6\times$ and $5.4\times$ reduction in computation and storage consumption under INT4 with little performance degradation, which indicates the immense potential of fully quantized detectors in real-world object detection applications.

While excellent in transfer learning, Vision-Language models (VLMs) come with high computational costs due to their large number of parameters. To address this issue, removing parameters via model pruning is a viable solution. However, existing techniques for VLMs are task-specific, and thus require pruning the network from scratch for each new task of interest. In this work, we explore a new direction: Task-Agnostic Vision-Language Pruning (TA-VLP). Given a pretrained VLM, the goal is to find a unique pruned counterpart transferable to multiple unknown downstream tasks. In this challenging setting, the transferable representations already encoded in the pretrained model are a key aspect to preserve. Thus, we propose Multimodal Flow Pruning (MULTIFLOW), a first, gradient-free, pruning framework for TA-VLP where: (i) the importance of a parameter is expressed in terms of its magnitude and its information flow, by incorporating the saliency of the neurons it connects; and (ii) pruning is driven by the emergent (multimodal) distribution of the VLM parameters after pretraining. We benchmark eight state-of-the-art pruning algorithms in the context of TA-VLP, experimenting with two VLMs, three vision-language tasks, and three pruning ratios. Our experimental results show that MULTIFLOW outperforms recent sophisticated, combinatorial competitors in the vast majority of the cases, paving the way towards addressing TA-VLP. The code is publicly available at https://github.com/FarinaMatteo/multiflow.

Adapting models pre-trained on large-scale datasets to a variety of downstream tasks is a common strategy in deep learning. Consequently, parameter-efficient fine-tuning methods have emerged as a promising way to adapt pre-trained models to different tasks while training only a minimal number of parameters. While most of these methods are designed for single-task adaptation, parameter-efficient training in Multi-Task Learning (MTL) architectures is still unexplored. In this paper, we introduce MTLoRA, a novel framework for parameter-efficient training for MTL models. MTLoRA employs Task-Agnostic and Task-Specific Low-Rank Adaptation modules, which effectively disentangle the parameter space in MTL fine-tuning, thereby enabling the model to adeptly handle both task specialization and interaction within MTL contexts. We applied MTLoRA to hierarchical-transformer-based MTL architectures, adapting them to multiple downstream dense prediction tasks. Our extensive experiments on the PASCAL dataset show that MTLoRA achieves higher accuracy in downstream tasks compared to fully fine-tuning the entire MTL model while training significantly fewer parameters. Furthermore, MTLoRA establishes a Pareto-optimal trade-off between the number of trainable parameters and the accuracy of the downstream tasks, outperforming current state-of-the-art parameter-efficient training methods in both accuracy and efficiency.

With the recent advances in vision transformers and large language models (LLMs), finetuning costly large models on downstream learning tasks poses significant challenges under limited computational resources. This paper presents a REsource and ComputAtion-efficient Pruning framework (RECAP) for the finetuning of transformer-based large models. RECAP by design bridges the gap between efficiency and performance through an iterative process cycling between pruning, finetuning, and updating stages to explore different chunks of the given large-scale model. At each iteration, we first prune the model with Taylor-approximation-based importance estimation and then only update a subset of the pruned model weights based on the Fisher-information criterion. In this way, RECAP achieves two synergistic and yet conflicting goals: reducing the GPU memory footprint while maintaining model performance, unlike most existing pruning methods that require the model to be finetuned beforehand for better preservation of model performance. We perform extensive experiments with a wide range of large transformer-based architectures on various computer vision and natural language understanding tasks. Compared to recent pruning techniques, we demonstrate that RECAP offers significant improvements in GPU memory efficiency, capable of reducing the footprint by up to 65\%.

Customizing robotic behaviors to be aligned with diverse human preferences is an underexplored challenge in the field of embodied AI.In this paper, we present Promptable Behaviors, a novel framework that facilitates efficient personalization of robotic agents to diverse human preferences in complex environments. We use multi-objective reinforcement learning to train a single policy adaptable to a broad spectrum of preferences. We introduce three distinct methods to infer human preferences by leveraging different types of interactions: (1) human demonstrations, (2) preference feedback on trajectory comparisons, and (3) language instructions. We evaluate the proposed method in personalized object-goal navigation and flee navigation tasks in ProcTHOR and RoboTHOR, demonstrating the ability to prompt agent behaviors to satisfy human preferences in various scenarios.

3D simulated environments play a critical role in Embodied AI, but their creation requires expertise and extensive manual effort, restricting their diversity and scope. To mitigate this limitation, we present Holodeck, a system that generates 3D environments to match a user-supplied prompt fully automatedly. Holodeck can generate diverse scenes, e.g., arcades, spas, and museums, adjust the designs for styles, and can capture the semantics of complex queries such as "apartment for a researcher with a cat” and “office of a professor who is a fan of Star Wars“. Holodeck leverages a large language model (i.e., GPT-4) for common sense knowledge about what the scene might look like and uses a large collection of 3D assets from Objaverse to populate the scene with diverse objects. To address the challenge of positioning objects correctly, we prompt GPT-4 to generate spatial relational constraints between objects and then optimize the layout to satisfy those constraints. Our large-scale human evaluation shows that annotators prefer Holodeck over manually designed procedural baselines in residential scenes and that Holodeck can produce high-quality outputs for diverse scene types. We also demonstrate an exciting application of Holodeck in Embodied AI, training agents to navigate in novel scenes like music rooms and daycares without human-constructed data, which is a significant step forward in developing general-purpose embodied agents.

Reinforcement learning (RL) with dense rewards and imitation learning (IL) with human-generated trajectories are the most widely used approaches for training modern embodied agents. RL requires extensive reward shaping and auxiliary losses and is often too slow and ineffective for long-horizon tasks. While IL with human supervision is effective, collecting human trajectories at scale is extremely expensive. In this work, we show that imitating shortest-path planners in simulation produces agents that, given a language instruction, can proficiently navigate, explore, and manipulate objects in both simulation and in the real world using only RGB sensors (no depth map or GPS coordinates). This surprising result is enabled by our end-to-end, transformer-based, SPOC architecture, powerful visual encoders paired with extensive image augmentation, and the dramatic scale and diversity of our training data: millions of frames of shortest-path-expert trajectories collected inside approximately 200,000 procedurally generated houses containing 40,000 unique 3D assets. Our models, data, training code, and newly proposed 10-task benchmarking suite CHORES is available at https://spoc-robot.github.io/.

We leverage Large Language Models (LLM) for zero-shot Semantic Audio Visual Navigation (SAVN). Existing methods utilize extensive training demonstrations for reinforcement learning, yet achieve relatively low success rates and lack generalizability. The intermittent nature of auditory signals further poses additional obstacles to inferring the goal information. To address this challenge, we present the Reflective and Imaginative Language Agent (RILA). By employing multi-modal models to process sensory data, we instruct an LLM-based planner to actively explore the environment. During the exploration, our agent adaptively evaluates and dismisses inaccurate perceptual descriptions. Additionally, we introduce an auxiliary LLM-based assistant to enhance global environmental comprehension by mapping room layouts and providing strategic insights. Through comprehensive experiments and analysis, we show that our method outperforms relevant baselines without training demonstrations from the environment and complementary semantic information.

With recent developments in Embodied Artificial Intelligence (EAI) research, there has been a growing demand for high-quality, large-scale interactive scene generation. While prior methods in scene synthesis have primarily emphasized the naturalness and realism of the generated scenes, the physical plausibility and interactivity of scenes have been primarily left untouched. To bridge this gap, we introduce PhyScene, a novel approach dedicated to generating interactive 3D scenes characterized by realistic layouts, articulated objects, and rich physical interactivity tailored for embodied agents. Based on a conditional diffusion model for capturing scene layouts, we devise novel physics- and interactivity-based guidance functions encompassing constraints from both object collision, room layout, as well as agent interactivity. Through extensive experiments, we demonstrate that PhyScene effectively leverages these guidance functions for physically interactable scene synthesis, outperforming existing state-of-the-art scene synthesis methods by a large margin. We believe that our generated scenes could be broadly beneficial for agents to learn diverse skills within interactive scenes and pave the way for new research in embodied AI.

Computer vision tasks typically involve describing what is visible in an image (e.g. classification, detection, segmentation, and captioning). We study a visual common sense task that requires understanding 'what is not visible'. Specifically, given an image (e.g. of a living room) and a name of an object ("cushion"), a vision system is asked to predict semantically-meaningful regions (masks or bounding boxes) in the image where that object could be placed or is likely be placed by humans (e.g. on the sofa). We call this task: Semantic Placement (SP) and believe that such common-sense visual understanding is critical for assitive robots (tidying a house), AR devices (automatically rendering an object in the user's space), and visually-grounded chatbots with common sense. Studying the invisible is hard. Datasets for image description are typically constructed by curating relevant images (e.g. via image search with object names) and asking humans to annotate the contents of the image; neither of those two steps are straightforward for objects not present in the image. We overcome this challenge by operating in the opposite direction: we start with an image of an object in context (which is easy to find online) and remove that object from the image via inpainting. This automated pipeline converts unstructured web data into a paired with/without object dataset. With this proposed data generation pipeline, we collect a novel dataset, containing ~1.3M images across 9 object categories. We then train a SP prediction model, called CLIP-UNet, on our dataset. The CLIP-UNet outperforms existing VLMs and baselines that combine semantic priors with object detectors, generalizes well to real-world and simulated images, exhibits semantics-aware reasoning for object placement, and enables downstream applications like tidying robots in indoor environments.

Learning 3D human-object interaction relation is pivotal to embodied AI and interaction modeling. Most existing methods approach the goal by learning to predict isolated interaction elements, e.g., human contact, object affordance, and human-object spatial relation, primarily from the perspective of either the human or the object. Which underexploit certain correlations between the interaction counterparts (human and object), and struggle to address the uncertainty in interactions. Actually, objects' functionalities potentially affect humans' interaction intentions, which reveals what the interaction is. Meanwhile, the interacting humans and objects exhibit matching geometric structures, which presents how to interact. In light of this, we propose harnessing these inherent correlations between interaction counterparts to mitigate the uncertainty and jointly anticipate the above interaction elements in 3D space. To achieve this, we present $\mathbf{LEMON}$ ($\mathbf{LE}$arning 3D hu$\mathbf{M}$an-$\mathbf{O}$bject i$\mathbf{N}$teraction relation), a unified model that mines interaction intentions of the counterparts and employs curvatures to guide the extraction of geometric correlations, combining them to anticipate the interaction elements. Besides, the $\mathbf{3D}$ $\mathbf{I}$nteraction $\mathbf{R}$elation dataset ($\mathbf{3DIR}$ ) is collected to serve as the test bed for training and evaluation. Extensive experiments demonstrate the superiority of LEMON over methods estimating each element in isolation. The code and dataset will be publicly released.

Recent advances in Iterative Vision-and-Language Navigation~(IVLN) introduce a more meaningful and practical paradigm of VLN by maintaining the agent's memory across tours of scenes. Although the long-term memory aligns better with the persistent nature of the VLN task, it poses more challenges on how to utilize the highly unstructured navigation memory with extremely sparse supervision. Towards this end, we propose OVER-NAV, which aims to go over and beyond the current arts of IVLN techniques. In particular, we propose to incorporate LLMs and open-vocabulary detectors to distill key information and establish correspondence between multi-modal signals. Such a mechanism introduces reliable cross-modal supervision and enables on-the-fly generalization to unseen scenes without the need of extra annotation and re-training. To fully exploit the interpreted navigation data, we further introduce a structured representation, coded Omnigraph, to effectively integrate multi-modal information along the tour. Accompanied with a novel omnigraph fusion mechanism, OVER-NAV is able to extract the most relevant knowledge from omnigraph for a more accurate navigating action. In addition, OVER-NAV seamlessly supports both discrete and continuous environments under a unified framework. We demonstrate the superiority of OVER-NAV in extensive experiments.

It is a long-lasting goal to design an embodied system that can solve long-horizon open-world tasks in human-like ways. However, existing approaches usually struggle with compound difficulties caused by the logic-aware decomposition and context-aware execution of these tasks. To this end, we introduce MP5, an open-ended multimodal embodied system built upon the challenging Minecraft simulator, which can decompose feasible sub-objectives, design sophisticated situation-aware plans, and perform embodied action control, with frequent communication with a goal-conditioned active perception scheme. Specifically, MP5 is developed on top of recent advances in Multimodal Large Language Models (MLLMs), and the system is modulated into functional modules that can be scheduled and collaborated to ultimately solve pre-defined context- and process-dependent tasks. Extensive experiments prove that MP5 can achieve a 22% success rate on difficult process-dependent tasks and a 91% success rate on tasks that heavily depend on the context. Moreover, MP5 exhibits a remarkable ability to address many open-ended tasks that are entirely novel.

Vision-language navigation (VLN) requires an agent to navigate through an 3D environment based on visual observations and natural language instructions. It is clear that the pivotal factor for successful navigation lies in the comprehensive scene understanding. Previous VLN agents employ monocular frameworks to extract 2D features of perspective views directly. Though straightforward, they struggle for capturing 3D geometry and semantics, leading to a partial and incomplete environment representation. To achieve a comprehensive 3D representation with fine-grained details, we introduce a Volumetric Environment Representation (VER), which voxelizes the physical world into structured 3D cells. For each cell, VER aggregates multi-view 2D features into such a unified 3D space via 2D-3D sampling. Through coarse-to-fine feature extraction and multi-task learning for VER, our agent predicts 3D occupancy, 3D room layout, and 3D bounding boxes jointly. Based on online collected VERs, our agent performs volume state estimation and builds episodic memory for predicting the next step. Experimental results show our environment representations from multi-task learning lead to evident performance gains on VLN. Our model achieves state-of-the-art performance across VLN benchmarks (R2R, REVERIE, and R4R).

As a new embodied vision task, Instance ImageGoal Navigation (IIN) aims to navigate to a specified object depicted by a goal image in an unexplored environment. The main challenge of this task lies in identifying the target object from different viewpoints while rejecting similar distractors. Existing ImageGoal Navigation methods usually adopt the simple Exploration-Exploitation framework and ignore the identification of specific instance during navigation. In this work, we propose to imitate the human behaviour of "getting closer to confirm" when distinguishing objects from a distance. Specifically, we design a new modular navigation framework named Instance-aware Exploration-Verification-Exploitation (IEVE) for instance-level image goal navigation. Our method allows for active switching among the exploration, verification, and exploitation actions, thereby facilitating the agent in making reasonable decisions under different situations. On the challenging HabitatMatterport 3D semantic (HM3D-SEM) dataset, our method surpasses previous state-of-the-art work, with a classical segmentation model (0.684 vs. 0.561 success) or a robust model (0.702 vs. 0.561 success). Our code will be made publicly available at https://anonymous.4open.science/r/IEVE-0CC3.

Garment manipulation (e.g., unfolding, folding and hanging clothes) is essential for future robots to accomplish home-assistant tasks, while highly challenging due to the diversity of garment configurations, geometries and deformations. Although able to manipulate similar shaped garments in a certain task, previous works mostly have to design different policies for different tasks, could not generalize to garments with diverse geometries, and often rely heavily on human-annotated data. In this paper, we leverage the property that, garments in a certain category have similar structures, and then learn the topological dense (point-level) visual correspondence among garments in the category level with different deformations in the self-supervised manner. The topological correspondence can be easily adapted to the functional correspondence to guide the manipulation policies for various downstream tasks, within only one or few-shot demonstrations. Experiments over garments in 3 different categories on 3 representative tasks in diverse scenarios, using one or two arms, taking one or more steps, inputting flat or messy garments, demonstrate the effectiveness of our proposed method. Project page: https://warshallrho.github.io/unigarmentmanip.

Active recognition enables robots to intelligently explore novel observations, thereby acquiring more information while circumventing undesired viewing conditions. Recent approaches favor learning policies from simulated or collected data, wherein appropriate actions are more frequently selected when the recognition is accurate. However, most recognition modules are developed under the closed-world assumption, which makes them ill-equipped to handle unexpected inputs, such as the absence of the target object in the current observation. To address this issue, we propose treating active recognition as a sequential evidence-gathering process, providing by-step uncertainty quantification and reliable prediction under the evidence combination theory. Additionally, the reward function developed in this paper effectively characterizes the merit of actions when operating in open-world environments. To evaluate the performance, we collect a dataset from an indoor simulator, encompassing various recognition challenges such as distance, occlusion levels, and visibility. Through a series of experiments on recognition and robustness analysis, we demonstrate the necessity of introducing uncertainties to active recognition and the superior performance of the proposed method.

This paper presents GenH2R, a framework for learning generalizable vision-based human-to-robot (H2R) handover skills. The goal is to equip robots with the ability to reliably receive objects with unseen geometry handed over by humans in various complex trajectories. We acquire such generalizability by learning H2R handover at scale with a comprehensive solution including procedural simulation assets creation, automated demonstration generation, and effective imitation learning. We leverage large-scale 3D model repositories, dexterous grasp generation methods, and curve-based 3D animation to create an H2R handover simulation environment named GenH2R-Sim, surpassing the number of scenes in existing simulators by three orders of magnitude. We further introduce a distillation-friendly demonstration generation method that automatically generates a million high-quality demonstrations suitable for learning. Finally, we present a 4D imitation learning method augmented by a future forecasting objective to distill demonstrations into a visuo-motor handover policy. Experimental evaluations in both simulators and the real world demonstrate significant improvements (at least +10% success rate) over baselines in all cases.

The Embodied AI community has recently made significant strides in visual navigation tasks, exploring targets from 3D coordinates, objects, language description, and images. However, these navigation models often handle only a single input modality as the target. With the progress achieved so far, it is time to move towards universal navigation models capable of handling various goal types, enabling more effective user interaction with robots. To facilitate this goal, we propose GOAT-Bench, a benchmark for the universal navigation task referred to as GO to AnyThing (GOAT). In this task, the agent is directed to navigate to a sequence of targets specified by the category name, language description, or instance image in an open-vocabulary fashion. We benchmark monolithic RL and modular methods on the GOAT task, analyzing their performance across modalities, the role of explicit and implicit scene memories, their robustness to noise in goal specifications, and the impact of memory in lifelong scenarios.

We contribute the Habitat Synthetic Scene Dataset, a dataset of 211 high-quality 3D scenes, and use it to test navigation agent generalization to realistic 3D environments. Our dataset represents real interiors and contains a diverse set of 18,656 models of real-world objects. We investigate the impact of synthetic 3D scene dataset scale and realism on the task of training embodied agents to find and navigate to objects (ObjectGoal navigation). By comparing to synthetic 3D scene datasets from prior work, we find that scale helps in generalization, but the benefits quickly saturate, making visual fidelity and correlation to real-world scenes more important. Our experiments show that agents trained on our smaller-scale dataset can match or outperform agents trained on much larger datasets. Surprisingly, we observe that agents trained on just 122 scenes from our dataset outperform agents trained on 10,000 scenes from the ProcTHOR-10K dataset in terms of zero-shot generalization in real-world scanned environments.

Active recognition, which allows intelligent agents to explore observations for better recognition performance, serves as a prerequisite for various embodied AI tasks, such as grasping, navigation and room arrangements. Given the evolving environment and the multitude of object classes, it is impractical to include all possible classes during the training stage. In this paper, we aim at advancing active open-vocabulary recognition, empowering embodied agents to actively perceive and classify arbitrary objects. However, directly adopting recent open-vocabulary classification models, like Contrastive Language Image Pretraining (CLIP), poses its unique challenges. Specifically, we observe that CLIP's performance is heavily affected by the viewpoint and occlusions, compromising its reliability in unconstrained embodied perception scenarios. Further, the sequential nature of observations in agent-environment interactions necessitates an effective method for integrating features that maintains discriminative strength for open-vocabulary classification. To address these issues, we introduce a novel agent for active open-vocabulary recognition. The proposed method leverages inter-frame and inter-concept similarities to navigate agent movements and to fuse features, without relying on class-specific knowledge. Compared to baseline CLIP model with 29.6\% accuracy on ShapeNet dataset, the proposed agent could achieve 53.3\% accuracy for open-vocabulary recognition, without any fine-tuning to the equipped CLIP model. Additional experiments conducted with the Habitat simulator further affirm the efficacy of our method.

Developing generalizable manipulation skills is a core challenge in embodied AI. This includes generalization across diverse task configurations, encompassing variations in object shape, density, friction coefficient, and external disturbances such as forces applied to the robot. Rapid Motor Adaptation (RMA) offers a promising solution to this challenge.It posits that essential hidden variables influencing an agent's task performance, such as object mass and shape, can be effectively inferred from the agent's action and proprioceptive history. Drawing inspiration from RMA in locomotion and in-hand rotation, we use depth perception to develop agents tailored for rapid motor adaptation in a variety of manipulation tasks. We evaluated our agents on four challenging tasks from the Maniskill2 benchmark, namely pick-and-place operations with hundreds of objects from the YCB and EGAD datasets, peg insertion with precise position and orientation, and operating a variety of faucets and handles. Empirical results demonstrate that our agents surpass state-of-the-art methods like automatic domain randomization and vision-based policies, obtaining better generalization performance and sample efficiency.

The Object Goal navigation (ObjectNav) task requires the agent to navigate to a specified target in an unseen environment. Since the environment layout is unknown, the agent needs to infer the unknown contextual objects from partially observations, thereby deducing the likely location of the target. Previous end-to-end RL methods learn the contextual relation with implicit representations while they lack notion of geometry. Alternatively, modular methods construct a local map for recording the observed geometric structure of unseen environment, however, lacking the reasoning of contextual relation limits the exploration efficiency. In this work, we propose the self-supervised generative map (SGM), a modular method that learns the explicit context relation via self-supervised learning. The SGM is trained to leverage both episodic observations and general knowledge to reconstruct the masked pixels of a cropped global map. During navigation, the agent maintains an incomplete local semantic map, meanwhile, the unknown regions of the local map are generated by the pre-trained SGM. Based on the scaled-up local map, the agent sets the predicted location of the target as the goal and moves towards it. Experiments on Gibson, MP3D and HM3D demonstrate the effectiveness and efficiency of our method.

Many reinforcement learning environments (e.g., Minecraft) provide only sparse rewards that indicate task completion or failure with binary values. The challenge in exploration efficiency in such environments makes it difficult for reinforcement-learning-based agents to learn complex tasks. To address this, this paper introduces an advanced learning system, named Auto MC-Reward, that leverages Large Language Models (LLMs) to automatically design dense reward functions, thereby enhancing the learning efficiency. Auto MC-Reward consists of three important components: Reward Designer, Reward Critic, and Trajectory Analyzer. Given the environment information and task descriptions, the Reward Designer first design the reward function by coding an executable Python function with predefined observation inputs. Then, our Reward Critic will be responsible for verifying the code, checking whether the code is self-consistent and free of syntax and semantic errors. Further, the Trajectory Analyzer summarizes possible failure causes and provides refinement suggestions according to collected trajectories. In the next round, Reward Designer will further refine and iterate the dense reward function based on feedback. Experiments demonstrate a significant improvement in the success rate and learning efficiency of our agents in complex tasks in Minecraft, such as obtaining diamond with the efficient ability to avoid lava, and efficiently explore trees and animals that are sparse in the plains biome.

While recent advances in neural radiance field enable realistic digitization for large-scale scenes, the image-capturing process is still time-consuming and labor-intensive. Previous works attempt to automate this process using the Next-Best-View (NBV) policy for active 3D reconstruction. However, the existing NBV policies heavily rely on hand-crafted criteria, limited action space, or per-scene optimized representations. These constraints limit their zero-shot generalizability. To overcome them, we propose \textbf{GenNBV}, an end-to-end generalizable NBV policy. Our policy adopts reinforcement learning (RL)-based framework and extends typical limited action space to 5D free space. It empowers our agent drone to scan from any viewpoint, and even interact with unseen geometries during training. In order to boost the zero-shot generalizability, we also propose a novel multi-source state embedding, including geometric, semantic, and action representations. We establish a benchmark using the Isaac Gym simulator with the Houses3K and OmniObject3D datasets to evaluate this NBV policy. Experiments demonstrate that our policy achieves a 98.26\% and 97.12\% coverage ratio on unseen building-scale objects from these datasets, respectively, outperforming prior solutions.

Visual navigation is to let the agent reach the target according to the continuous visual input. In most previous works, visual navigation is usually assumed to be done in a static and ideal environment: the target is always reachable with no need to alter the environment. However, the “messy” environments are more general and practical in our daily lives, where the agent may get blocked by obstacles. Thus Interactive Navigation (InterNav) is introduced to navigate to the objects in more realistic "messy" environments according to the object interaction. Prior work on InterNav learns short-term interaction through extensive trials with reinforcement learning. However, interaction does not guarantee efficient navigation, that is, planning obstacle interactions that make shorter paths and consume less effort is also crucial. In this paper, we introduce an effect-oriented affordance map to enable long-term interactive navigation, extending the existing map-based navigation framework to the domain of dynamic environment. We train a set of affordance functions predicting available interactions and the time cost of removing obstacles, which informatively support an interactive modular system to address interaction and long-term planning. Experiments on the ProcTHOR simulator demonstrate the capability of our affordance-driven system in long-term navigation in complex dynamic environments.

This paper presents a novel category agnostic model for visual rearrangement task, which can help an embodied agent to physically recover the shuffled scene configuration without any category concepts to the goal configuration. Previous methods usually follow a similar architecture, completing the rearrangement task by aligning the scene changes of the goal and shuffled configuration, according to the semantic scene graphs. However, constructing scene graphs requires the inference of category labels, which not only causes the accuracy drop of the entire task but also limits the application in real world scenario. In this paper, we delve deep into the essence of visual rearrangement task and focus on the two most essential issues, scene change detection and scene change matching. We utilize the movement and the protrusion of point cloud to accurately identify the scene changes and match these changes depending on the similarity of category agnostic appearance feature. Moreover, to assist the agent to explore the environment more efficiently and comprehensively, we propose a closer-aligned-retrace exploration policy, aiming to observe more details of the scene at a closer distance. We conduct extensive experiments on AI2THOR Rearrangement Challenge based on RoomR dataset and a new multi-room multi-instance dataset MrMiR collected by us. The experimental results demonstrate the effectiveness of our proposed method.

Diffusion models have demonstrated strong potential for robotic trajectory planning. However, generating coherent trajectories from high-level instructions remains challenging, especially for long-range composition tasks requiring multiple sequential skills. We propose SkillDiffuser, an end-to-end hierarchical planning framework integrating interpretable skill learning with conditional diffusion planning to address this problem. At the higher level, the skill abstraction module learns discrete, human-understandable skill representations from visual observations and language instructions. These learned skill embeddings are then used to condition the diffusion model to generate customized latent trajectories aligned with the skills. This allows generating diverse state trajectories that adhere to the learnable skills. By integrating skill learning with conditional trajectory generation, SkillDiffuser produces coherent behavior following abstract instructions across diverse tasks. Experiments on multi-task robotic manipulation benchmarks like Meta-World and LOReL demonstrate state-of-the-art performance and human-interpretable skill representations from SkillDiffuser. More visualization results and information could be found on https://skilldiffuser.github.io/.

As wearable cameras become more popular, an important question emerges: how to identify camera wearers within the perspective of conventional static cameras. The drastic difference between first-person (egocentric) and third-person (exocentric) camera views makes this a challenging task. We present PersonEnvironmentNet (PEN), a framework designed to integrate information from both the individuals in the two views and geometric cues inferred from the background environment. To facilitate research in this direction, we also present TF2023, a novel dataset comprising synchronized first-person and third-person views, along with masks of camera wearers and labels associating these masks with the respective first-person views. In addition, we propose a novel quantitative metric designed to measure a model's ability to comprehend the relationship between the two views. Our experiments reveal that PEN outperforms existing methods. The code and dataset are available at https://github.com/ziweizhao1993/PEN.

We present a modern formulation of Embodied Question Answering (EQA) as the task of understanding an environment well enough to answer questions about it in natural language. An agent can achieve such an understanding by either drawing upon episodic memory, exemplified by agents on smart glasses, or by actively exploring the environment, as in the case of mobile robots. We accompany our formulation with OpenEQA -- the first open-vocabulary benchmark dataset for EQA supporting both episodic memory and active exploration use cases. OpenEQA contains over 1600 high-quality human generated questions drawn from over 180 real-world environments. In addition to the dataset, we also provide an automatic LLM-powered evaluation protocol that has excellent correlation with human judgement. Using this dataset and evaluation protocol, we evaluate several state-of-the-art foundation models including GPT-4V, and find that they significantly lag behind human-level performance. Consequently, OpenEQA stands out as a straightforward, measurable, and practically relevant benchmark that poses a considerable challenge to current generation of foundation models. We hope this inspires and stimulates future research at the intersection of Embodied AI, conversational agents, and world models.

When adopting a deep learning model for embodied agents, it is required that the model structure be optimized for specific tasks and operational conditions. Such optimization can be static such as model compression or dynamic such as adaptive inference. Yet, these techniques have not been fully investigated for embodied control systems subject to time constraints, which necessitate sequential decision-making for multiple tasks, each with distinct inference latency limitations. In this paper, we present MoDeC, a time constraint-aware embodied control framework using the modular model adaptation. We formulate model adaptation to varying operational conditions on resource and time restrictions as dynamic routing on a modular network, incorporating these conditions as part of multi-task objectives. Our evaluation across several vision-based embodied environments demonstrates the robustness of MoDeC, showing that it outperforms other model adaptation methods in both performance and adherence to constraints in robotic manipulation and autonomous driving applications.

Understanding the world in first-person view is fundamental in Augmented Reality (AR). This immersive perspective brings dramatic visual changes and unique challenges compared to third-person views. Synthetic data has empowered third-person-view vision models, but its application to embodied egocentric perception tasks remains largely unexplored. A critical challenge lies in simulating natural human movements and behaviors that effectively steer the embodied cameras to capture a faithful egocentric representation of the 3D world. To address this challenge, we introduce EgoGen, a new synthetic data generator that can produce accurate and rich ground-truth training data for egocentric perception tasks. At the heart of EgoGen is a novel human motion synthesis model that directly leverages egocentric visual inputs of a virtual human to sense the 3D environment. Combined with collision-avoiding motion primitives and a two-stage reinforcement learning approach, our motion synthesis model offers a closed-loop solution where the embodied perception and movement of the virtual human are seamlessly coupled. Compared to previous works, our model eliminates the need for a pre-defined global path, and is directly applicable to dynamic environments. Combined with our easy-to-use and scalable data generation pipeline, we demonstrate EgoGen’s efficacy in three tasks: mapping and localization for head-mounted cameras, egocentric camera tracking, and human mesh recovery from egocentric views. EgoGen will be fully open-sourced, offering a practical solution for creating realistic egocentric training data and aiming to serve as a useful tool for egocentric computer vision research.

We propose RoHM, an approach for robust 3D human motion reconstruction from monocular RGB(-D) videos in the presence of noise and occlusions. Most previous approaches either train neural networks to directly regress motion in 3D or learn data-driven motion priors and combine them with optimization at test time. RoHM is a novel diffusion-based motion model that, conditioned on noisy and occluded input data, reconstructs complete, plausible motions in consistent global coordinates. Given the complexity of the problem -- requiring one to address different tasks (denoising and infilling) in different solution spaces (local and global motion) -- we decompose it into two sub-tasks and learn two models, one for global trajectory and one for local motion. To capture the correlations between the two, we then introduce a novel conditioning module, combining it with an iterative inference scheme. We apply RoHM to a variety of tasks -- from motion reconstruction and denoising to spatial and temporal infilling. Extensive experiments on three popular datasets show that our method outperforms state-of-the-art approaches qualitatively and quantitatively, while being faster at test time. The code is available at https://sanweiliti.github.io/ROHM/ROHM.html.

Event cameras respond primarily to edges---formed by strong gradients---and are thus particularly well-suited for line-based motion estimation. Recent work has shown that events generated by a single line each satisfy a polynomial constraint which describes a manifold in the space-time volume. Multiple such constraints can be solved simultaneously to recover the partial linear velocity and line parameters. In this work, we show that, with a suitable line parametrization, this system of constraints is actually linear in the unknowns, which allows us to design a novel linear solver. Unlike existing solvers, our linear solver (i) is fast and numerically stable since it does not rely on expensive root finding, (ii) can solve both minimal and overdetermined systems with more than 5 events, and (iii) admits the characterization of all degenerate cases and multiple solutions. The found line parameters are singularity-free and have a fixed scale, which eliminates the need for auxiliary constraints typically encountered in previous work. To recover the full linear camera velocity we fuse observations from multiple lines with a novel velocity averaging scheme that relies on a geometrically-motivated residual, and thus solves the problem more efficiently than previous schemes which minimize an algebraic residual. Extensive experiments in synthetic and real-world settings demonstrate that our method surpasses the previous work in numerical stability, and operates over 600 times faster.

We present the subspace-constrained Tyler's estimator (STE) designed for recovering a low-dimensional subspace within a dataset that may be highly corrupted with outliers. STE is a fusion of the Tyler's M-estimator (TME) and a variant of the fast median subspace, offering superior computational efficiency compared to TME. Our theoretical analysis suggests that, under a common inlier-outlier model, STE can effectively recover the underlying subspace, even when it contains a smaller fraction of inliers relative to other methods in the field of robust subspace recovery. We apply STE in the context of Structure from Motion (SfM) in two ways: for robust estimation of the fundamental matrix and for the removal of outlying cameras, enhancing the robustness and speed of the SfM pipeline. Numerical experiments confirm the state-of-the-art performance of our method in these applications. This research makes significant contributions to the field of robust subspace recovery, particularly in the context of computer vision and 3D reconstruction.

Finding shortest paths on product spaces is a popular approach to tackle numerous variants of matching problems, including the dynamic time warping method for matching signals, the matching of curves, or the matching of a curve to a 3D shape. While these approaches admit the computation of globally optimal solutions in polynomial time, their natural generalisation to 3D shape matching is widely known to be intractable. In this work we address this issue by proposing a novel path-based formalism for 3D shape matching. More specifically, we consider an alternative shape discretisation in which one of the 3D shapes (the source shape) is represented as a SpiderCurve, i.e. a long self-intersecting curve that traces the 3D shape surface. We then tackle the 3D shape matching problem as finding a shortest path in the product graph of the SpiderCurve and the target 3D shape. Our approach introduces a set of novel constraints that ensure a globally geometrically consistent matching. Overall, our formalism leads to an integer linear programming problem for which we experimentally show that it can efficiently be solved to global optimality. We demonstrate that our approach is competitive with recent state-of-the-art shape matching methods, while in addition guaranteeing geometric consistency.

Two primary input modalities prevail in image retrieval: sketch and text. While text is widely used for inter-category retrieval tasks, sketches have been established as the sole preferred modality for fine-grained image retrieval due to their ability to capture intricate visual details. In this paper, we question the reliance on sketches alone for fine-grained image retrieval by simultaneously exploring the fine-grained representation capabilities of both sketch and text, orchestrating a duet between the two. The end result enables precise retrievals previously unattainable, allowing users to pose ever-finer queries and incorporate attributes like colour and contextual cues from text. For this purpose, we introduce a novel compositionality framework, effectively combining sketches and text using pre-trained CLIP models, while eliminating the need for extensive fine-grained textual descriptions. Last but not least, our system extends to novel applications in composed image retrieval, domain attribute transfer, and fine-grained generation, providing solutions for various real-world scenarios.

Knowledge Distillation (KD) has been validated as an effective model compression technique for learning compact object detectors. Existing state-of-the-art KD methods for object detection are mostly based on feature imitation, which is generally observed to be better than prediction mimicking. In this paper, we show that the inconsistency of the optimization objectives between the ground-truth signals and distillation targets is the key reason for the inefficiency of prediction mimicking. To alleviate this issue, we present a general and effective distillation scheme, called CrossKD, which delivers the intermediate features of the student's detection head to the teacher's detection head. The resulting cross-head predictions are then forced to mimic the teacher's predictions. This manner relieves the student's head from receiving contradictory supervision signals from the annotations and the teacher's predictions, greatly improving the student's detection performance. On MS COCO, with only prediction mimicking losses applied, our CrossKD boosts the average precision of GFL ResNet-50 with 1x training schedule from 40.2 to 43.7, outperforming all existing KD methods. In addition, our method also works well when distilling detectors with heterogeneous backbones. Code will be made publicly available.

Visual-language foundation models, like CLIP, learn generalized representations that enable zero-shot open-set classification. Few-shot adaptation methods, based on prompt tuning, have been shown to further improve performance on downstream datasets. However, these methods do not fare well in the taxonomic open set (TOS) setting, where the classifier is asked to make prediction from label set across different levels of semantic granularity. Frequently, they infer incorrect label at coarser taxonomic class levels, even when the inference at the leaf level (original class labels) is correct. To address this problem, we propose a prompt tuning technique that calibrates the hierarchical consistency of model predictions. A set of metrics of hierarchical consistency, the Hierarchical Consistent Accuracy (HCA) and the Mean Treecut Accuracy (MTA), are first proposed to evaluate TOS model performance. A new Prompt Tuning for Hierarchical Consistency (ProTeCt) technique is then proposed to calibrate classification across label set granularities. Results show that ProTeCt can be combined with existing prompt tuning methods to significantly improve TOS classification without degrading the leaf level classification performance.

Domain adaptive object detection aims to adapt detection models to domains where annotated data is unavailable. Existing methods have been proposed to address the domain gap using the semi-supervised student-teacher framework. However, a fundamental issue arises from the class imbalance in the labelled training set, which can result in inaccurate pseudo-labels. The relationship between classes, especially where one class is a majority and the other minority, has a large impact on class bias. We propose Class-Aware Teacher (CAT) to address the class bias issue in the domain adaptation setting. In our work, we approximate the class relationships with our Inter-Class Relation module (ICRm) and exploit it to reduce the bias within the model. In this way, we are able to apply augmentations to highly related classes, both inter- and intra-domain, to boost the performance of minority classes while having minimal impact on majority classes. We further reduce the bias by implementing a class-relation weight to our classification loss. Experiments conducted on various datasets and ablation studies show that our method is able to address the class bias in the domain adaptation setting. On the Cityscapes → Foggy Cityscapes dataset, we attained a 52.5 mAP, a substantial improvement over the 51.2 mAP achieved by the state-of-the-art method.

The increasing prevalence of video clips has sparked growing interest in text-video retrieval. Recent advances focus on establishing a joint embedding space for text and video, relying on consistent embedding representations to compute similarity. However, the text content in existing datasets is generally short and concise, making it hard to fully describe the redundant semantics of a video. Correspondingly, a single text embedding may be less expressive to capture the video embedding and empower the retrieval. In this study, we propose a new stochastic text modeling method T-MASS, i.e., text is modeled as a stochastic embedding, to enrich text embedding with a flexible and resilient semantic range, yielding a text mass. To be specific, we introduce a similarity-aware radius module to adapt the scale of the text mass upon the given text-video pairs. Plus, we design and develop a support text regularization to further control the text mass during the training. The inference pipeline is also tailored to fully exploit the text mass for accurate retrieval. Empirical evidence suggests that T-MASS not only effectively attracts relevant text-video pairs while distancing irrelevant ones, but also enables the determination of precise text embeddings for relevant pairs. Our experimental results show a substantial improvement of T-MASS over baseline (3%~6.3% by R@1). Also, T-MASS achieves state-of-the-art performance on five benchmark datasets, including MSRVTT, LSMDC, DiDeMo, VATEX, and Charades. We will release code and models.

Realizing unified monocular 3D object detection, including both indoor and outdoor scenes, holds great importance in applications like robot navigation. However, involving various scenarios of data to train models poses challenges due to their significantly different characteristics, e.g., diverse geometry properties and heterogeneous domain distributions. To address these challenges, we build a detector based on the bird's-eye-view (BEV) detection paradigm, where the explicit feature projection is beneficial to addressing the geometry learning ambiguity when employing multiple scenarios of data to train detectors. Then, we split the classical BEV detection architecture into two stages and propose an uneven BEV grid design to handle the convergence instability caused by the aforementioned challenges. Moreover, we develop a sparse BEV feature projection strategy to reduce computational cost and a unified domain alignment method to handle heterogeneous domains. Combining these techniques, a unified detector UniMODE is derived, which surpasses the previous state-of-the-art on the challenging Omni3D dataset (a large-scale dataset including both indoor and outdoor scenes) by 4.9% $\rm AP_{3D}$, revealing the first successful generalization of a BEV detector to unified 3D object detection.

The challenge of open-vocabulary recognition lies in the model has no clue of new categories it is applied to. Existing works have proposed different methods to embed category cues into the model, e.g., through few-shot fine-tuning, providing category names or textual descriptions to Vision-Language Models. Fine-tuning is time-consuming and degrades the generalization capability. Textual descriptions could be ambiguous and fail to depict visual details. This paper tackles open-vocabulary recognition from a different perspective by referring to multi-modal clues composed of textual descriptions and exemplar images. Our method, named OVMR, adopts two innovative components to pursue a more robust category cues embedding. A multi-modal classifier is first generated by dynamically complementing textual descriptions with image exemplars. A preference-based refinement module is hence applied to fuse uni-modal and multi-modal classifiers, with the aim to alleviate issues of low-quality exemplar images or textual descriptions. The proposed OVMR is a plug-and-play module, and works well with exemplar images randomly crawled from the Internet.Extensive experiments have demonstrated the promising performance of OVMR, e.g., it outperforms existing methods across various scenarios and setups.

Action understanding matters for intelligent agents and has attracted long-term attention. It can be formed as the mapping from the action physical space to the semantic space. Typically, researchers built action datasets according to idiosyncratic choices to define classes and push the envelope of benchmarks respectively. Thus, datasets are incompatible with each other like "Isolated Islands" due to semantic gaps and various class granularities, e.g., do housework in dataset A and wash plate in dataset B. We argue that a more principled semantic space is an urgent need to concentrate the community efforts and enable us to use all datasets together to pursue generalizable action learning. To this end, we design a structured action semantic space in view of verb taxonomy hierarchy and covering massive actions. By aligning the classes of previous datasets to our semantic space, we gather (image/video/skeleton/MoCap) datasets into a unified database in a unified label system, i.e., bridging ``isolated islands'' into a "Pangea". Accordingly, we propose a novel model mapping from the physical space to semantic space to fully use Pangea. In extensive experiments, our new system shows significant superiority, especially in transfer learning. Our code and data will be made public at https://mvig-rhos.com/pangea.

We present a new open-vocabulary detection framework. Our framework uses both image-level labels and detailed detection annotations when available. Our framework proceeds in three steps. We first train a language-conditioned object detector on fully-supervised detection data. This detector gets to see the presence or absence of ground truth classes during training, and conditions prediction on the set of present classes. We use this detector to pseudo-label images with image-level labels. Our detector provides much more accurate pseudo-labels than prior approaches with its conditioning mechanism. Finally, we train an unconditioned open-vocabulary detector on the pseudo-annotated images. The resulting detector, named DECOLA, shows strong zero-shot performance in open-vocabulary LVIS benchmark as well as direct zero-shot transfer benchmarks on LVIS, COCO, Object365, and OpenImages. DECOLA outperforms the prior arts by 17.1 AP-rare and 9.4 mAP on zero-shot LVIS benchmark. DECOLA achieves state-of-the-art results in various model sizes, architectures, and datasets by only training on open-sourced data and academic-scale computing. Code is available at https://github.com/janghyuncho/DECOLA.

Lifelong person re-identification (LReID) suffers from the catastrophic forgetting problem when learning from non-stationary data. Existing exemplar-based and knowledge distillation-based LReID methods encounter data privacy and limited acquisition capacity respectively. In this paper, we instead introduce the prototype, which is under-investigated in LReID, to better balance knowledge forgetting and acquisition. Existing prototype-based works primarily focus on the classification task, where the prototypes are set as discrete points or statistical distributions. However, they either discard the distribution information or omit instance-level diversity which are crucial fine-grained clues for LReID. To address the above problems, we propose Distribution-aware Knowledge Prototyping (DKP) where the instance-level diversity of each sample is modeled to transfer comprehensive fine-grained knowledge for prototyping and facilitating LReID learning. Specifically, an Instance-level Distribution Modeling network is proposed to capture the local diversity of each instance. Then, the Distribution-oriented Prototype Generation algorithm transforms the instance-level diversity into identity-level distributions as prototypes, which is further explored by the designed Prototype-based Knowledge Transfer module to enhance the knowledge anti-forgetting and acquisition capacity of the LReID model. Extensive experiments verify that our method achieves superior plasticity and stability balancing and outperforms existing LReID methods by 8.1%/9.1% average mAP/R@1 improvement. The code is available at https://github.com/zhoujiahuan1991/CVPR2024-DKP

Lifelong Person Re-identification (L-ReID) aims to learn from sequentially collected data to match a person across different scenes. Once an L-ReID model is updated using new data, all historical images in the gallery are required to be re-calculated to obtain new features for testing, known as ``re-indexing''. However, it is infeasible when raw images in the gallery are unavailable due to data privacy concerns, resulting in incompatible retrieval between the query and the gallery features calculated by different models, which causes significant performance degradation. In this paper, we focus on a new task called Re-indexing Free Lifelong Person Re-identification (RFL-ReID), which requires achieving effective L-ReID without re-indexing raw images in the gallery. To this end, we propose a Continual Compatible Representation (C2R) method, which facilitates the query feature calculated by the continuously updated model to effectively retrieve the gallery feature calculated by the old model in a compatible manner. Specifically, we design a Continual Compatible Transfer (CCT) network to continuously transfer and consolidate the old gallery feature into the new feature space. Besides, a Balanced Compatible Distillation module is introduced to achieve compatibility by aligning the transferred feature space with the new feature space. Finally, a Balanced Anti-forgetting Distillation module is proposed to eliminate the accumulated forgetting of old knowledge during the continual compatible transfer. Extensive experiments on several benchmark L-ReID datasets demonstrate the effectiveness of our method against state-of-the-art methods for both RFL-ReID and L-ReID tasks. The source code of this paper is available at https://github.com/PKU-ICST-MIPL/C2R_CVPR2024.

Accurately detecting active objects undergoing state changes is essential for comprehending human interactions and facilitating decision-making. The existing methods for active object detection (AOD) primarily rely on visual appearance of the objects within input, such as changes in size, shape and relationship with hands. However, these visual changes can be subtle, posing challenges, particularly in scenarios with multiple distracting no-change instances of the same category. We observe that the state changes are often the result of an interaction being performed upon the object, thus propose to use informed priors about object related plausible interactions (including semantics and visual appearance) to provide more reliable cues for AOD. Specifically, we propose a knowledge aggregation procedure to integrate the aforementioned informed priors into oracle queries within the teacher decoder, offering more object affordance commonsense to locate the active object. To streamline the inference process and reduce extra knowledge inputs, we propose a knowledge distillation approach that encourages the student decoder to mimic the detection capabilities of the teacher decoder using the oracle query by replicating its predictions and attention. Our proposed framework achieves state-of-the-art performance on four datasets, namely Ego4D, Epic-Kitchens, MECCANO, and 100DOH, which demonstrates the effectiveness of our approach in improving AOD.

Open-vocabulary object detection (OvOD) has transformed detection into a language-guided task, empowering users to freely define their class vocabularies of interest during inference. However, our initial investigation indicates that existing OvOD detectors exhibit significant variability when dealing with vocabularies across various semantic granularities, posing a concern for real-world deployment. To this end, we introduce Semantic Hierarchy Nexus (SHiNe), a novel classifier that leverages semantic knowledge from class hierarchies. It is built offline in three steps: i) it retrieves relevant super-/sub-categories from a hierarchy for each target class; ii) it integrates these categories into hierarchy-aware sentences; iii) it fuses these sentence embeddings to generate the nexus classifier vector. Our evaluation on various detection benchmarks demonstrates that SHiNe enhances robustness across diverse vocabulary granularities, achieving up to +31.9% mAP50 with an existing hierarchy, while retaining improvements using hierarchies generated by large language models. Moreover, when applied to open-vocabulary classification on ImageNet-1k, SHiNe improves the CLIP zero-shot baseline by +2.8% accuracy. SHiNe is training-free and can be seamlessly integrated with any off-the-shelf OvOD detector, without incurring extra computational overhead during inference.

We present an approach to pose object recognition as next token prediction.The idea is to apply a language decoder that auto-regressively predicts the text tokens from image embeddings to form labels.To ground this prediction process in auto-regression, we customize a non-causal attention mask for the decoder, incorporating two key features: modeling tokens from different labels to be independent, and treating image tokens as a prefix.This masking mechanism inspires an efficient method $-$ one-shot sampling $-$ to simultaneously sample tokens of multiple labels in parallel and rank generated labels by their probabilities during inference.To further enhance the efficiency, we propose a simple strategy to construct a compact decoder by simply discarding the intermediate blocks of a pretrained language model.This approach yields a decoder that matches the full model's performance while being notably more efficient.The code is available at [github.com/nxtp](https://github.com/kaiyuyue/nxtp).

Open-vocabulary human-object interaction (HOI) detection, which is concerned with the problem of detecting novel HOIs guided by natural language, is crucial for understanding human-centric scenes. However, prior zero-shot HOI detectors often employ the same levels of feature maps to model HOIs with varying distances, leading to suboptimal performance in scenes containing human-object pairs with a wide range of distances.In addition, these detectors primarily rely on category names and overlook the rich contextual information that language can provide, which is essential for capturing open vocabulary concepts that are typically rare and not well-represented by category names alone.In this paper, we introduce a novel end-to-end open vocabulary HOI detection framework with conditional multi-level decoding and fine-grained semantic enhancement~(CMD-SE), harnessing the potential of Visual-Language Models (VLMs). Specifically, we propose to model human-object pairs with different distances with different levels of feature maps by incorporating a soft constraint during the bipartite matching process. Furthermore, by leveraging large language models (LLMs) such as GPT models, we exploit their extensive world knowledge to generate descriptions of human body part states for various interactions. Then we integrate the generalizable and fine-grained semantics of human body parts to improve interaction recognition.Experimental results on two datasets, SWIG-HOI and HICO-DET, demonstrate that our proposed method achieves state-of-the-art results in open vocabulary HOI detection.

Class-Incremental Learning (CIL) trains a model to continually recognize new classes from non-stationary data while retaining learned knowledge. A major challenge of CIL arises when applying to real-world data characterized by non-uniform distribution, which introduces a dual imbalance problem involving (i) disparities between stored exemplars of old tasks and new class data (inter-phase imbalance), and (ii) severe class imbalances within each individual task (intra-phase imbalance). We show that this dual imbalance issue causes skewed gradient updates with biased weights in FC layers, thus inducing over/under-fitting and catastrophic forgetting in CIL. Our method addresses it by reweighting the gradients towards balanced optimization and unbiased classifier learning. Additionally, we observe imbalanced forgetting where paradoxically the instance-rich classes suffer higher performance degradation during CIL due to a larger amount of training data becoming unavailable in subsequent learning phases. To tackle this, we further introduce a distribution-aware knowledge distillation loss to mitigate forgetting by aligning output logits proportionally with the distribution of lost training data. We validate our method on CIFAR-100, ImageNetSubset, and Food101 across various evaluation protocols and demonstrate consistent improvements compared to existing works, showing great potential to apply CIL in real-world scenarios with enhanced robustness and effectiveness.

Open vocabulary object detection (OVD) aims at seeking an optimal object detector capable of recognizing objects from both base and novel categories. Recent advances leverage knowledge distillation to transfer insightful knowledge from pre-trained large-scale vision-language models to the task of object detection, significantly generalizing the powerful capabilities of the detector to identify more unknown object categories. However, these methods face significant challenges in background interpretation and model overfitting and thus often result in the loss of crucial background knowledge, giving rise to sub-optimal inference performance of the detector. To mitigate these issues, we present a novel OVD framework termed LBP to propose learning background prompts to harness explored implicit background knowledge, thus enhancing the detection performance w.r.t. base and novel categories. Specifically, we devise three modules: Background Category-specific Prompt, Background Object Discovery, and Inference Probability Rectification, to empower the detector to discover, represent, and leverage implicit object knowledge explored from background proposals. Evaluation on two benchmark datasets, OV-COCO and OV-LVIS, demonstrates the superiority of our proposed method over existing state-of-the-art approaches in handling the OVD tasks.

We present Multi-View Attentive Contextualization (MvACon), a simple yet effective method for improving 2D-to-3D feature lifting in query-based multi-view 3D (MV3D) object detection. Despite remarkable progress witnessed in the field of query-based MV3D object detection, prior art often suffers from either the lack of exploiting high-resolution 2D features in dense attention-based lifting, due to high computational costs, or from insufficiently dense grounding of 3D queries to multi-scale 2D features in sparse attention-based lifting. Our proposed MvACon hits the two birds with one stone using a representationally dense yet computationally sparse attentive feature contextualization scheme that is agnostic to specific 2D-to-3D feature lifting approaches. Specifically, we propose a plug-and-play module which computes global cluster-based contextualized features as complementary context for the 2D-to-3D feature lifting.In experiments, MvACon is tested on the nuScenes benchmark, using both the BEVFormer and its recent 3D deformable attention (DFA3D) variant, as well as the PETR, showing consistent detection performance improvement, especially in enhancing performance in location, orientation, and velocity prediction. We qualitatively and quantitatively show that global cluster-based contexts effectively encode dense scene-level contexts for MV3D object detection.The promising results of our MvACon reinforces the slang in computer vision --- ``(contextualized) feature matters".

Self-supervised feature reconstruction methods have shown promising advances in industrial image anomaly detection and localization. Despite this progress, these methods still face challenges in synthesizing realistic and diverse anomaly samples, as well as addressing the feature redundancy and pre-training bias of pre-trained feature. In this work, we introduce RealNet, a feature reconstruction network with realistic synthetic anomaly and adaptive feature selection. It is incorporated with three key innovations: First, we propose Strength-controllable Diffusion Anomaly Synthesis (SDAS), a diffusion process-based synthesis strategy capable of generating samples with varying anomaly strengths that mimic the distribution of real anomalous samples. Second, we develop Anomaly-aware Features Selection (AFS), a method for selecting representative and discriminative pre-trained feature subsets to improve anomaly detection performance while controlling computational costs. Third, we introduce Reconstruction Residuals Selection (RRS), a strategy that adaptively selects discriminative residuals for comprehensive identification of anomalous regions across multiple levels of granularity. We assess RealNet on four benchmark datasets, and our results demonstrate significant improvements in both Image AUROC and Pixel AUROC compared to the current state-of-the-art methods. The code, data, and models are available at https://github.com/cnulab/RealNet.

The lightweight "local-match-global" matching introduced by SRe2L successfully creates a distilled dataset with comprehensive information on the full 224$\times$224 ImageNet-1k. However, this one-sided approach is limited to a particular backbone, layer, and statistics, which limits the improvement of the generalization of a distilled dataset. We suggest that sufficient and various "local-match-global" matching are more precise and effective than a single one and has the ability to create a distilled dataset with richer information and better generalization. We call this perspective "generalized matching" and propose Generalized Various Backbone and Statistical Matching (G-VBSM) in this work, which aims to create a synthetic dataset with densities, ensuring consistency with the complete dataset across various backbones, layers, and statistics. As experimentally demonstrated, G-VBSM is the first algorithm to obtain strong performance across both small-scale and large-scale datasets. Specifically, G-VBSM achieves a performance of 38.7% on CIFAR-100 with 128-width ConvNet, 47.6% on Tiny-ImageNet with ResNet18, and 31.4% on the full 224$\times$224 ImageNet-1k with ResNet18, under images per class (IPC) 10, 50, and 10, respectively. These results surpass all SOTA methods by margins of 3.9%, 6.5%, and 10.1%, respectively.

This paper investigates the effective utilization of unlabeled data for large-area cross-view geo-localization (CVGL), encompassing both unsupervised and semi-supervised settings. Common approaches to CVGL rely on ground-satellite image pairs and employ label-driven supervised training. However, the cost of collecting precise cross-view image pairs hinders the deployment of CVGL in real-life scenarios. Without the pairs, CVGL will be more challenging to handle the significant imaging and spatial gaps between ground and satellite images. To this end, we propose an unsupervised framework including a cross-view projection to guide the model for retrieving initial pseudo-labels and a fast re-ranking mechanism to refine the pseudo-labels by leveraging the fact that ``the perfectly paired ground-satellite image is located in a unique and identical scene". The framework exhibits competitive performance compared with supervised works on three open-source benchmarks. Our code and models will be released on https://github.com/liguopeng0923/UCVGL.

Single point-supervised object detection is gaining attention due to its cost-effectiveness. However, existing approaches focus on generating horizontal bounding boxes (HBBs) while ignoring oriented bounding boxes (OBBs) commonly used for objects in aerial images. This paper proposes PointOBB, the first single Point-based OBB generation method, for oriented object detection. PointOBB operates through the collaborative utilization of three distinctive views: an original view, a resized view, and a rotated/flipped (rot/flp) view. Upon the original view, we leverage the resized and rot/flp views to build a scale augmentation module and an angle acquisition module, respectively. In the former module, a Scale-Sensitive Consistency (SSC) loss is designed to enhance the deep network's ability to perceive the object scale. For accurate object angle predictions, the latter module incorporates self-supervised learning to predict angles, which is associated with a scale-guided Dense-to-Sparse (DS) matching strategy for aggregating dense angles corresponding to sparse objects. The resized and rot/flp views are switched using a progressive multi-view switching strategy during training to achieve coupled optimization of scale and angle. Experimental results on the DIOR-R and DOTA-v1.0 datasets demonstrate that PointOBB achieves promising performance, and significantly outperforms potential point-supervised baselines.

Open Vocabulary Object Detection (OVD) aims to detect objects from novel classes described by text inputs based on the generalization ability of trained classes. Existing methods mainly focus on transferring knowledge from large Vision and Language models (VLM) to detectors based on knowledge distillation. However, these approaches show weak ability in adaptation to diverse classes and alignment between the image-level pre-training and region-level detection, hindering successful knowledge transfer. Motivated by the prompt tuning, we propose scene-adaptive and region-aware multi-modal prompts to address these issues by effectively adapting class-aware knowledge from VLM to the detector at the region level. Specifically, to enhance the adaptability to diverse classes, we design a scene-adaptive prompt generator from a scene perspective to consider both the commonality and diversity of the class distributions, and formulate a novel selection mechanism to facilitate the acquisition of common knowledge across all classes and specific insights relevant to each scene. Meanwhile, to bridge the gap between the pre-trained model and the detector, we present a region-aware multi-modal alignment module, which employs the region prompt to incorporate the positional information for feature distillation and integrates textual prompts to align visual and linguistic representations. Extensive experimental results demonstrate that the proposed method significantly outperforms the state-of-the-art models on the OV-COCO and OV-LVIS datasets, surpassing the current method by 3.0% mAP and 4.6% $\text{AP}_r$.

Active domain adaptation (ADA) aims to maximally boost the model adaptation on a new target domain by actively selecting limited target data to annotate. Such a setting neglects the more practical scenario where training data are collected from multiple sources, which motivates us to target a new and challenging setting of knowledge transfer that extends ADA from a single source domain to multiple source domains, called Multi-source Active Domain Adaptation (MADA). Not surprisingly, we find that most of the traditional ADA methods cannot directly work in such a setting, resulting in significant performance degradation compared with the single-source domain setting. This is mainly due to the excessive domain gap introduced from all the source domains and their uncertainty-aware sample selection can easily be miscalibrated on data under the multi-domain shifts.Considering this, we propose a $\underline{\textbf{D}}$ynamic int$\underline{\textbf{e}}$gra$\underline{\textbf{te}}$d un$\underline{\textbf{c}}$er$\underline{\textbf{t}}$a$\underline{\textbf{i}}$nty $\underline{\textbf{v}}$aluation fram$\underline{\textbf{e}}$work~($\textbf{Detective}$) that comprehensively consider the domain shift between multi-source domains and target domain to detect the informative target samples. Specifically, the Detective leverages a dynamic DA model that learns how to adapt the model’s parameters and fit the union of multi-source domains, which enables an approximate single-source domain modeling by the dynamic model. Then we comprehensively measure the domain uncertainty and predictive uncertainty on the target domain to detect the informative target samples by evidential deep learning, mitigating the uncertainty miscalibration. We further seek a contextual diversity-aware calculator to enhance the diversity of the selected samples. Experiments show that the solution outperforms existing methods by a considerable margin on three domain adaptation benchmarks.

Open-world detection poses significant challenges, as it requires the detection of any object using either object class labels or free-form texts. Existing related works often use large-scale manual annotated caption datasets for training, which are extremely expensive to collect. Instead, we propose to transfer knowledge from vision-language models (VLMs) to enrich the open-vocabulary descriptions automatically. Specifically, we bootstrap dense synthetic captions using pre-trained VLMs to provide rich descriptions on different regions in images, and incorporate these captions to train a novel detector that generalizes to novel concepts. To mitigate the noise caused by hallucination in synthetic captions, we also propose a novel hyperbolic vision-language learning approach to impose a hierarchy between visual and caption embeddings. We call our detector ``HyperLearner''. We conduct extensive experiments on a wide variety of open-world detection benchmarks (COCO, LVIS, Object Detection in the Wild, RefCOCO) and our results show that our model consistently outperforms existing state-of-the-art methods, such as GLIP, GLIPv2 and Grounding DINO, when using the same backbone.

Over the past decade, most methods in visual place recognition (VPR) have used neural networks to produce feature representations. These networks typically produce a global representation of a place image using only this image itself and neglect the cross-image variations (e.g. viewpoint and illumination), which limits their robustness in challenging scenes. In this paper, we propose a robust global representation method with cross-image correlation awareness for VPR, named CricaVPR. Our method uses the self-attention mechanism to correlate multiple images within a batch. These images can be taken in the same place with different conditions or viewpoints, or even captured from different places. Therefore, our method can utilize the cross-image variations as a cue to guide the representation learning, which ensures more robust features are produced. To further facilitate the robustness, we propose a multi-scale convolution-enhanced adaptation method to adapt pre-trained visual foundation models to the VPR task, which introduces the multi-scale local information to further enhance the cross-image correlation-aware representation. Experimental results show that our method outperforms state-of-the-art methods by a large margin with significantly less training time. Our method achieves 94.5% R@1 on Pitts30k using 512-dim global features. The code will be publicly available.

With the rapidly increasing demand for oriented object detection (OOD), recent research involving weakly-supervised detectors for learning rotated box (RBox) from the horizontal box (HBox) has attracted more and more attention. In this paper, we explore a more challenging yet label-efficient setting, namely single point-supervised OOD, and present our approach called Point2RBox. Specifically, we propose to leverage two principles: 1) Synthetic pattern knowledge combination: By sampling around each labeled point on the image, we spread the object feature to synthetic visual patterns with known boxes to provide the knowledge for box regression. 2) Transform self-supervision: With a transformed input image (e.g. scaled/rotated), the output RBoxes are trained to follow the same transformation so that the network can perceive the relative size/rotation between objects. The detector is further enhanced by a few devised techniques to cope with peripheral issues, e.g. the anchor/layer assignment as the size of the object is not available in our point supervision setting. To our best knowledge, Point2RBox is the first end-to-end solution for point-supervised OOD. In particular, our method uses a lightweight paradigm, yet it achieves a competitive performance among point-supervised alternatives, 41.05%/27.62%/80.01% on DOTA/DIOR/HRSC datasets.

While recent Transformer-based approaches have shown impressive performances on event-based object detection tasks, their high computational costs still diminish the low power consumption advantage of event cameras. Image-based works attempt to reduce these costs by introducing sparse Transformers. However, they display inadequate sparsity and adaptability when applied to event-based object detection, since these approaches cannot balance the fine granularity of token-level sparsification and the efficiency of window-based Transformers, leading to reduced performance and efficiency. Furthermore, they lack scene-specific sparsity optimization, resulting in information loss and a lower recall rate. To overcome these limitations, we propose the Scene Adaptive Sparse Transformer (SAST). SAST enables window-token co-sparsification, significantly enhancing fault tolerance and reducing computational overhead. Leveraging the innovative scoring and selection modules, along with the Masked Sparse Window Self-Attention, SAST showcases remarkable scene-aware adaptability: It focuses only on important objects and dynamically optimizes sparsity level according to scene complexity, maintaining a remarkable balance between performance and computational cost. The evaluation results show that SAST outperforms all other dense and sparse networks in both performance and efficiency on two large-scale event-based object detection datasets (1Mpx and Gen1).

Composed Image Retrieval (CIR) is a task that retrieves images similar to a query, based on a provided textual modification. Current techniques rely on supervised learning for CIR models using labeled triplets of the . These specific triplets are not as commonly available as simple image-text pairs, limiting the widespread use of CIR and its scalability. On the other hand, zero-shot CIR can be relatively easily trained with image-caption pairs without considering the image-to-image relation, but this approach tends to yield lower accuracy. We propose a new semi-supervised CIR approach where we search for a reference and its related target images in auxiliary data and learn our large language model-based Visual Delta Generator (VDG) to generate text describing the visual difference (i.e., visual delta) between the two. VDG, equipped with fluent language knowledge and being model agnostic, can generate pseudo triplets to boost the performance of CIR models. Our approach significantly improves the existing supervised learning approaches and achieves state-of-the-art results on the CIR benchmarks.

Although effective deepfake detection models have been developed in recent years, recent studies have revealed that these models can result in unfair performance disparities among demographic groups, such as race and gender. This can lead to particular groups facing unfair targeting or exclusion from detection, potentially allowing misclassified deepfakes to manipulate public opinion and undermine trust in the model.The existing method for addressing this problem is providing a fair loss function. It shows good fairness performance for intra-domain evaluation but does not maintain fairness for cross-domain testing. This highlights the significance of fairness generalization in the fight against deepfakes. In this work, we propose the first method to address the fairness generalization problem in deepfake detection by simultaneously considering features, loss, and optimization aspects. Our method employs disentanglement learning to extract demographic and domain-agnostic forgery features, fusing them to encourage fair learning across a flattened loss landscape. Extensive experiments on prominent deepfake datasets demonstrate our method's effectiveness, surpassing state-of-the-art approaches in preserving fairness during cross-domain deepfake detection. The code is available at https://anonymous.4open.science/r/fairness_deefape-E2ED.

This paper, for the first time, explores text-to-image diffusion models for Zero-Shot Sketch-based Image Retrieval (ZS-SBIR). We highlight a pivotal discovery: the capacity of text-to-image diffusion models to seamlessly bridge the gap between sketches and photos. This proficiency is underpinned by their robust cross-modal capabilities and shape bias, findings that are substantiated through our pilot studies. In order to harness pre-trained diffusion models effectively, we introduce a straightforward yet powerful strategy focused on two key aspects: selecting optimal feature layers and utilising visual and textual prompts. For the former, we identify which layers are most enriched with information and are best suited for the specific retrieval requirements (category-level or fine-grained). Then we employ visual and textual prompts to guide the model's feature extraction process, enabling it to generate more discriminative and contextually relevant cross-modal representations. Extensive experiments on several benchmark datasets validate significant performance improvements.

The vision-language model has brought great improvement to few-shot industrial anomaly detection, which usually needs to design of hundreds of prompts through prompt engineering. For automated scenarios, we first use conventional prompt learning with many-class paradigm as the baseline to automatically learn prompts but found that it can not work well in one-class anomaly detection. To address the above problem, this paper proposes a one-class prompt learning method for few-shot anomaly detection, termed PromptAD. First, we propose semantic concatenation which can transpose normal prompts into anomaly prompts by concatenating normal prompts with anomaly suffixes, thus constructing a large number of negative samples used to guide prompt learning in one-class setting. Furthermore, to mitigate the training challenge caused by the absence of anomaly images, we introduce the concept of explicit anomaly margin, which is used to explicitly control the margin between normal prompt features and anomaly prompt features through a hyper-parameter. For image-level/pixel-level anomaly detection, PromptAD achieves first place in 11/12 few-shot settings on MVTec and VisA.

Despite encouraging results from recent developments in transfer learning for adapting pre-trained model to downstream tasks, the performance of model probing is still lagging behind the state-of-the-art parameter efficient tuning methods. Our investigation reveals that existing model probing methods perform well for the easy case when the source domain (where models are pre-trained) and the adapted domain are similar, but fail for the difficult case when the two domains are significantly different. Simply incorporating features extracted from multiple layers and increasing complexity of the probing model can mitigate the gap in the difficult case, but degrades the performance in the easy case. To address this challenge, we propose structured model probing ({\bf SMP}) that is able to deliver good performance for both cases through {\bf structured regularization}. The regularization performs feature selection leveraging model structure as a prior, and controls the complexity of the probing model through the weights of selected structures. This enables us to construct a simple adaptation model, with a small number of selected features and a linear prediction model, for the easy case; and to automatically increase the complexity of adaptation model, with a large number of selected features and a non-linear model, for the difficult case. Our extensive empirical studies show that SMP significantly outperforms the state-of-the-art methods for parameter efficient tuning, and at the same time, still maintains the advantage of computational efficiency for probing-based methods.

In this paper, we propose a novel abstraction-aware sketch-based image retrieval framework capable of handling sketch abstraction at varied levels. Prior works had mainly focused on tackling sub-factors such as drawing style and order, we instead attempt to model abstraction as a whole, and propose feature-level and retrieval granularity-level designs so that the system builds into its DNA the necessary means to interpret abstraction. On learning abstraction-aware features, we for the first-time harness the rich semantic embedding of pre-trained StyleGAN model, together with a novel abstraction-level mapper that deciphers the level of abstraction and dynamically selects appropriate dimensions in the feature matrix correspondingly, to construct a feature matrix embedding that can be freely traversed to accommodate different levels of abstraction. For granularity-level abstraction understanding, we dictate that the retrieval model should not treat all abstraction-levels equally and introduce a differentiable surrogate $\texttt{Acc.@q}$ loss to inject that understanding into the system. Different to the gold-standard triplet loss, our $\texttt{Acc.@q}$ loss uniquely allows a sketch to narrow/broaden its focus in terms of how stringent the evaluation should be -- the more abstract a sketch, the less stringent (higher $q$). Extensive experiments depict our method to outperform existing state-of-the-arts in standard SBIR tasks along with challenging scenarios like early retrieval, forensic sketch-photo matching, and style-invariant retrieval.

Unsupervised visible-infrared person re-identification (US-VI-ReID) centers on learning a cross-modality retrieval model without labels, reducing the reliance on expensive cross-modality manual annotation. Previous US-VI-ReID works gravitate toward learning cross-modality information with the deep features extracted from the ultimate layer. Nevertheless, interfered by the multiple discrepancies, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization.The shallow feature from the shallow layers contains nuanced detail information, which is critical for effective cross-modality learning but is disregarded regrettably by the existing methods. To address the above issues, we design a Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer with shallow-deep contrastive learning, incorporating Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. Specifically, CNL unveils the intrinsic homogeneous and heterogeneous collaboration which are harnessed for neighbor alignment, enhancing the robustness in a dynamic manner. Furthermore, CRA associates the cross-modality labels with the ranking association between shallow and deep features, furnishing valuable supervision for cross-modality learning. Extensive experiments validate the superiority of our method, even outperforming certain supervised counterparts.

Generalized Category Discovery(GCD) aims to identify a mix of known and novel categories within unlabeled data sets, providing a more realistic setting for image recognition.Essentially, GCD needs to remember existing patterns thoroughly to recognize novel categories.Recent state-of-the-art method SimGCD transfers the knowledge from known-class data to the learning of novel classes through debiased learning. However, some patterns are catastrophically forgot during adaptation and thus lead to poor performance in novel categories classification.To address this issue, we propose a novel learning approach, LegoGCD, which is seamlessly integrated into previous methods to enhance the discrimination of novel classes while maintaining performance on previously encountered known classes.Specifically, we design two types of techniques termed as Local Entropy Regularization(LER) and Dual-views Kullback–Leibler divergence constraint(DKL).The LER optimizes the distribution of potential known class samples in unlabeled data, thus ensuring the preservation of knowledge related to known categories while learning novel classes.Meanwhile, DKL introduces Kullback–Leibler divergence to encourage the model to produce a similar prediction distribution of two view samples from the same image.In this way, it successfully avoids mismatched prediction and generates more reliable potential known class samples simultaneously. Extensive experiments validate that the proposed LegoGCD effectively addresses the known category forgetting issue across all datasets, e.g., delivering a 7.74% and 2.51% accuracy boost on known and novel classes in CUB, respectively. Our code is released anonymously at: https://anonymous.4open.science/r/LegoGCD-E6BF/.

Generalized Category Discovery (GCD) is a pragmatic and challenging open-world task, which endeavors to cluster unlabeled samples from both novel and old classes, leveraging some labeled data of old classes. Given that knowledge learned from old classes is not fully transferable to new classes, and that novel categories are fully unlabeled, GCD inherently faces intractable problems, including imbalanced classification performance and inconsistent confidence between old and new classes, especially in the low-labeling regime. Hence, some annotations of new classes are deemed necessary. However, labeling new classes is extremely costly. To address this issue, we take the spirit of active learning and propose a new setting called Active Generalized Category Discovery (AGCD). The goal is to improve the performance of GCD by actively selecting a limited amount of valuable samples for labeling from the oracle. To solve this problem, we devise an adaptive sampling strategy, which jointly considers novelty, informativeness and diversity to adaptively select novel samples with proper uncertainty. However, owing to the varied orderings of label indices caused by the clustering of novel classes, the queried labels are not directly applicable to subsequent training. To overcome this issue, we further propose a stable label mapping algorithm that transforms ground truth labels to the label space of the classifier, thereby ensuring consistent training across different active selection stages. Our method achieves state-of-the-art performance on both generic and fine-grained datasets. Our code is available at https://github.com/mashijie1028/ActiveGCD

The You Only Look Once (YOLO) series of detectors have established themselves as efficient and practical tools. However, their reliance on predefined and trained object categories limits their applicability in open scenarios. Addressing this limitation, we introduce YOLO-World, an innovative approach that enhances YOLO with open-vocabulary detection capabilities through vision-language modeling and pre-training on large-scale datasets. Specifically, we propose a new Re-parameterizable Vision-Language Path Aggregation Network (RepVL-PAN) and region-text contrastive loss to facilitate the interaction between visual and linguistic information. Our method excels in detecting a wide range of objects in a zero-shot manner with high efficiency. On the challenging LVIS dataset, YOLO-World achieves 35.4 AP with 52.0 FPS on V100, which outperforms many state-of-the-art methods in terms of both accuracy and speed. Furthermore, the fine-tuned YOLO-World achieves remarkable performance on several downstream tasks, including object detection and open-vocabulary instance segmentation. The code will be made available.

Considerable efforts have been devoted to Oriented Object Detection (OOD). However, one lasting issue regarding the discontinuity in Oriented Bounding Box (OBB) representation remains unresolved, which is an inherent bottleneck for extant OOD methods. This paper endeavors to completely solve this issue in a theoretically guaranteed manner and puts an end to the ad-hoc efforts in this direction. Prior studies typically can only address one of the two cases of discontinuity: rotation and aspect ratio, and often inadvertently introduce decoding discontinuity, e.g. Decoding Incompleteness (DI) and Decoding Ambiguity (DA) as discussed in literature. Specifically, we propose a novel representation method called Continuous OBB (COBB), which can be readily integrated into existing detectors e.g. Faster-RCNN as a plugin. It can theoretically ensure continuity in bounding box regression which to our best knowledge, has not been achieved in literature for rectangle-based object representation. For fairness and transparency of experiments, we have developed a modularized benchmark based on the open-source deep learning framework Jittor's detection toolbox JDet for OOD evaluation. On the popular DOTA dataset, by integrating Faster-RCNN as the same baseline model, our new method outperforms the peer method Gliding Vertex by 1.13% mAP50 (relative improvement 1.54%), and 2.46% mAP75 (relative improvement 5.91%), without any tricks.

We delve into pseudo-labeling for semi-supervised monocular 3D object detection (SSM3OD) and discover two primary issues: a misalignment between the prediction quality of 3D and 2D attributes and the tendency of depth supervision derived from pseudo-labels to be noisy, leading to significant optimization conflicts with other reliable forms of supervision. To tackle these issues, we introduce a novel decoupled pseudo-labeling (DPL) approach for SSM3OD. Our approach features a Decoupled Pseudo-label Generation (DPG) module, designed to efficiently generate pseudo-labels by separately processing 2D and 3D attributes. This module incorporates aunique homography-based method for identifying dependable pseudo-labels in Bird’s Eye View (BEV) space, specifically for 3D attributes. Additionally, we present a Depth Gradient Projection (DGP) module to mitigate optimization conflicts caused by noisy depth supervision of pseudo-labels, effectively decoupling the depth gradient and removing conflicting gradients. This dual decoupling strategy—at both the pseudo-label generation and gradient levels—significantly improves the utilization of pseudo-labels in SSM3OD. Our comprehensive experiments on the KITTI benchmark demonstrate the superiority of our method over existing approaches.

Object detection with event cameras benefits from the sensor's low latency and high dynamic range. However, it is costly to fully label event streams for supervised training due to their high temporal resolution. To reduce this cost, we present LEOD, the first method for label-efficient event-based detection. Our approach unifies weakly- and semi-supervised object detection with a self-training mechanism. We first utilize a detector pre-trained on limited labels to produce pseudo ground truth on unlabeled events. Then, the detector is re-trained with both real and generated labels. Leveraging the temporal consistency of events, we run bi-directional inference and apply tracking-based post-processing to enhance the quality of pseudo labels. To stabilize training against label noise, we further design a soft anchor assignment strategy. We introduce new experimental protocols to evaluate the task of label-efficient event-based detection on Gen1 and 1Mpx datasets. LEOD consistently outperforms supervised baselines across various labeling ratios. For example, on Gen1, it improves mAP by 8.6% and 7.8% for RVT-S trained with 1% and 2% labels. On 1Mpx, RVT-S with 10% labels even surpasses its fully-supervised counterpart using 100% labels. LEOD maintains its effectiveness even when all labeled data are available, reaching new state-of-the-art results. Finally, we show that our method readily scales to improve larger detectors as well. Code is released at https://github.com/Wuziyi616/LEOD.

In this paper, we present a novel sequence generation-based framework for lane detection, called Lane2Seq. It unifies various lane detection formats by casting lane detection as a sequence generation task. This is different from previous lane detection methods, which depend on well-designed task-specific head networks and corresponding loss functions. Lane2Seq only adopts a plain transformer-based encoder-decoder architecture with a simple cross-entropy loss. Additionally, we propose a new multi-format model tuning based on reinforcement learning to incorporate the task-specific knowledge into Lane2Seq. Experimental results demonstrate that such a simple sequence generation paradigm not only unifies lane detection but also achieves competitive performance on benchmarks. For example, Lane2Seq gets 97.95\% and 97.42\% F1 score on Tusimple and LLAMAS datasets, establishing a new state-of-the-art result for two benchmarks.

In this paper, we develop MP-HOI, a powerful Multi-modal Prompt-based HOI detector designed toleverage both textual descriptions for open-set generalization and visual exemplars for handling high ambiguity in descriptions, realizing HOI detection in the open world. Specifically, it integrates visual prompts into existing language-guided-only HOI detectors to handle situations where textual descriptions face difficulties in generalization and to address complex scenarios with high interaction ambiguity. To facilitate MP-HOI training, we build a large-scale HOI dataset named Magic-HOI, which gathers six existing datasets into a unified label space, forming over 186K images with 2.4K objects, 1.2K actions, and 20K HOI interactions. Furthermore, to tackle the long-tail issue within the Magic-HOI dataset, we introduce an automated pipeline for generating realistically annotated HOI images and present SynHOI, a high-quality synthetic HOI dataset containing 100K images. Leveraging these two datasets, MP-HOI optimizes the HOI task as a similarity learning process between multi-modal prompts and objects/interactions via a unified contrastive loss, to learn generalizable and transferable objects/interactions representations from large-scale data. MP-HOI could serve as a generalist HOI detector, surpassing the HOI vocabulary of existing expert models by more than 30 times. Concurrently, our results demonstrate that MP-HOI exhibits remarkable zero-shot capability in real-world scenarios and consistently achieves a new state-of-the-art performance across various benchmarks. Ourproject homepage is available at https://MP-HOI.github.io/.

The YOLO series has become the most popular framework for real-time object detection due to its reasonable trade-off between speed and accuracy. However, we observe that the speed and accuracy of YOLOs are negatively affected by the NMS. Recently, end-to-end Transformer-based detectors (DETRs) have provided an alternative to eliminating NMS. Nevertheless, the high computational cost limits their practicality and hinders them from fully exploiting the advantage of excluding NMS. In this paper, we propose the Real-Time DEtection TRansformer (RT-DETR), the first real-time end-to-end object detector to our best knowledge that addresses the above dilemma. We build RT-DETR in two steps, drawing on the advanced DETR: first we focus on maintaining accuracy while improving speed, followed by maintaining speed while improving accuracy. Specifically, we design an efficient hybrid encoder to expeditiously process multi-scale features by decoupling intra-scale interaction and cross-scale fusion to improve speed. Then, we propose the uncertainty-minimal query selection to provide high-quality initial queries to the decoder, thereby improving accuracy. In addition, RT-DETR supports flexible speed tuning by adjusting the number of decoder layers to adapt to various scenarios without retraining. Our RT-DETR-R50 / R101 achieves 53.1% / 54.3% AP on COCO and 108 / 74 FPS on T4 GPU, outperforming previously advanced YOLOs in both speed and accuracy. Furthermore, RT-DETR-R50 outperforms DINO-R50 by 2.2% AP in accuracy and about 21 times in FPS. After pre-training with Objects365, RT-DETR-R50 / R101 achieves 55.3% / 56.2% AP. Code is available in the supplementary material.

Open-vocabulary object detection aims to detect novel categories that are independent from the base categories used during training. Most modern methods adhere to the paradigm of learning vision-language space from a large-scale multi-modal corpus and subsequently transferring the acquired knowledge to off-the-shelf detectors like Faster-RCNN. However, information attenuation or destruction may occur during the process of knowledge transfer due to the domain gap, hampering the generalization ability on novel categories. To mitigate this predicament, in this paper, we present a novel framework named BIND, standing for Bulit-IN Detector, to eliminate the need for module replacement or knowledge transfer to off-the-shelf detectors. Specifically, we design a two-stage training framework with an Encoder-Decoder structure. In the first stage, an image-text dual encoder is trained to learn region-word alignment from a corpus of image-text pairs. In the second stage, a DETR-style decoder is trained to perform detection on annotated object detection datasets. In contrast to conventional manually designed non-adaptive anchors, which generate numerous redundant proposals, we develop an anchor proposal network that generates anchor proposals with high likelihood based on candidates adaptively, thereby substantially improving detection efficiency. Experimental results on two public benchmarks, COCO and LVIS, demonstrate that our method stands as a state-of-the-art approach for open-vocabulary object detection. The code and models will be publicly available.

Existing counting tasks are limited to the class level, which don’t account for fine-grained details within the class. In real applications, it often requires in-context or referring human input for counting target objects. Take urban analysis as an example, fine-grained information such as traffic flow in different directions, pedestrians and vehicles waiting or moving at different sides of the junction, is more beneficial. Current settings of both class-specific and class-agnostic counting treat objects of the same class indifferently, which pose limitations in real use cases. To this end, we propose a new task named Referring Expression Counting (REC) which aims to count objects with different attributes within the same class. To evaluate the REC task, we create a novel dataset named REC-8K which contains 8011 images and 17122 referring expressions. Experiments on REC-8K show that our proposed method achieves state-of-the-art performance compared with several text-based counting methods and an open-set object detection model. We also outperform prior models on the class agnostic counting (CAC) benchmark [36] for the zero-shot setting, and perform on par with the few-shot methods. Code and dataset is available at https://github.com/sydai/referring-expression-counting.

The pretraining-finetuning paradigm has gained popularity in various computer vision tasks. In this paradigm, the emergence of active finetuning arises due to the abundance of large-scale data and costly annotation requirements. Active finetuning involves selecting a subset of data from an unlabeled pool for annotation, facilitating subsequent finetuning. However, the use of a limited number of training samples can lead to a biased distribution, potentially resulting in model overfitting. In this paper, we propose a new method called ActiveDC for the active finetuning tasks. Firstly, we select samples for annotation by optimizing the distribution similarity between the subset to be selected and the entire unlabeled pool in continuous space. Secondly, we calibrate the distribution of the selected samples by exploiting implicit category information in the unlabeled pool. The feature visualization provides an intuitive sense of the effectiveness of our approach to distribution calibration. We conducted extensive experiments on three image classification datasets with different sampling ratios. The results indicate that ActiveDC consistently outperforms the baseline performance in all image classification tasks. The improvement is particularly significant when the sampling ratio is low, with performance gains of up to 10%. Our code will be publicly available.

The evolution of Diffusion Models has dramatically improved image generation quality, making it increasingly difficult to differentiate between real and generated images. This development, while impressive, also raises significant privacy and security concerns. In response to this, we propose a novel Latent REconstruction error guided feature REfinement method ($\textbf{LaRE}^2$) for detecting the diffusion-generated images. We come up with the Latent Reconstruction Error (LaRE), the first reconstruction-error based feature in the latent space for generated image detection. LaRE surpasses existing methods in terms of feature extraction efficiency while preserving crucial cues required to differentiate between the real and the fake. To exploit LaRE, we propose an Error-Guided feature REfinement module (EGRE), which can refine the image feature guided by LaRE to enhance the discriminativeness of the feature. Our EGRE utilizes an align-then-refine mechanism, which effectively refines the image feature for generated-image detection from both spatial and channel perspectives. Extensive experiments on the large-scale GenImage benchmark demonstrate the superiority of our $\textbf{LaRE}^2$, which surpasses the best SoTA method by up to $\textbf{11.9\\%}$/$\textbf{12.1\\%}$ average ACC/AP across 8 different image generators. LaRE also surpasses existing methods in terms of feature extraction cost, delivering an impressive speed enhancement of $\textbf{8 times}$.

This paper studies the problem of semi-supervised 2D-3D retrieval, which aims to align both labeled and unlabeled 2D and 3D data into the same embedding space. The problem is challenging due to the complicated heterogeneous relationships between 2D and 3D data. Moreover, label scarcity in real-world applications hinders from generating discriminative representations. In this paper, we propose a semi-supervised approach named Fine-grained Prototypcical Voting with Heterogeneous Mixup (FIVE), which maps both 2D and 3D data into a common embedding space for cross-modal retrieval. Specifically, we generate fine-grained prototypes to model inter-class variation for both 2D and 3D data. Then, considering each unlabeled sample as a query, we retrieve relevant prototypes to vote for reliable and robust pseudo-labels, which serve as guidance for discriminative learning under label scarcity. Furthermore, to bridge the semantic gap between two modalities, we mix cross-modal pairs with similar semantics in the embedding space and then perform similarity learning for cross-modal discrepancy reduction in a soft manner. The whole FIVE is optimized with the consideration of sharpness to mitigate the impact of potential label noise. Extensive experiments on benchmark datasets validate the superiority of FIVE compared with a range of baselines in different settings. On average, FIVE outperforms the second-best approach by 4.74% on 3D MNIST, 12.94% on ModelNet10, and 22.10% on ModelNet40.

DETR accomplishes end-to-end object detection through iteratively generating multiple object candidates based on image features and promoting one candidate for each ground-truth object. The traditional training procedure using one-to-one supervision in the original DETR lacks direct supervision for the object detection candidates.We aim at improving the DETR training efficiency by explicitly supervising the candidate generation procedure through mixing one-to-one supervision and one-to-many supervision. Our approach, namely MS-DETR, is simple, and places one-to-many supervision to the object queries of the primary decoder that is used for inference. In comparison to existing DETR variants with one-to-many supervision, such as Group DETR and Hybrid DETR, our approach does not need additional decoder branches or object queries. The object queries of the primary decoder in our approach directly benefit from one-to-many supervision and thus are superior in object candidate prediction.Experimental results show that our approach outperforms related DETR variants, such as DN-DETR, Hybrid DETR, and Group DETR, and the combination with related DETR variants further improves the performance.

Compositional Zero-Shot Learning (CZSL) aims to recognize unseen attribute-object pairs based on a limited set of observed examples. Current CZSL methodologies, despite their advancements, tend to neglect the distinct specificity levels present in attributes. For instance, given images of sliced strawberries, they may fail to prioritize 'Sliced-Strawberry' over a generic 'Red-Strawberry', despite the former being more informative. They also suffer from ballooning search space when shifting from Close-World (CW) to Open-World (OW) CZSL. To address the issues, we introduce the Context-based and Diversity-driven Specificity learning framework for CZSL (CDS-CZSL). Our framework evaluates the specificity of attributes by considering the diversity of objects they apply to and their related context. This novel approach allows for more accurate predictions by emphasizing specific attribute-object pairs and improves composition filtering in OW-CZSL. We conduct experiments in both CW and OW scenarios, and our model achieves state-of-the-art results across three datasets.

Establishing precise semantic correspondence across object instances in different images is a fundamental and challenging task in computer vision. In this task, difficulty arises often due to three challenges: confusing regions with similar appearance, inconsistent object scale, and indistinguishable nearby pixels. Recognizing these challenges, our paper proposes a novel semantic matching pipeline named LPMFlow toward extracting fine-grained semantics and geometry layouts for building pixel-level semantic correspondences. LPMFlow consists of three modules, each addressing one of the aforementioned challenges. The layout-aware representation learning module uniformly encodes source and target tokens to distinguish pixels or regions with similar appearances but different geometry semantics. The progressive feature superresolution module outputs four sets of 4D correlation tensors to generate accurate semantic flow between objects in different scales. Finally, the matching flow integration and refinement module is exploited to fuse matching flow in different scales to give the final flow predictions. The whole pipeline can be trained end-to-end, with a balance of computational cost and correspondence details. Extensive experiments based on benchmarks such as SPair-71K, PF-PASCAL, and PF-WILLOW have proved that the proposed method can well tackle the three challenges and outperform the previous methods, especially in more stringent settings. Code is available at https://github.com/YXSUNMADMAX/LPMFlow.

Dataset distillation has emerged as a promising approach in deep learning, enabling efficient training with small synthetic datasets derived from larger real ones. Particularly, distribution matching-based distillation methods attract attention thanks to its effectiveness and low computational cost. However, these methods face two primary limitations: the dispersed feature distribution within the same class in synthetic datasets, reducing class discrimination, and an exclusive focus on mean feature consistency, lacking precision and comprehensiveness. To address these challenges, we introduce two novel constraints: a class centralization constraint and a covariance matching constraint. The class centralization constraint aims to enhance class discrimination by more closely clustering samples within classes. The covariance matching constraint seeks to achieve more accurate feature distribution matching between real and synthetic datasets through local feature covariance matrices, particularly beneficial when sample sizes are much smaller than the number of features. Experiments demonstrate notable improvements with these constraints, yielding performance boosts of up to 6.6% on CIFAR10, 2.9% on SVHN, 2.5% on CIFAR100, and 2.5% on TinyImageNet, compared to the state-of-the-art relevant methods. In addition, our method maintains robust performance in cross-architecture settings, with a maximum performance drop of 1.7% on four architectures.

Class-agnostic object counting aims to count all objects in an image with respect to example boxes or class names, \emph{a.k.a} few-shot and zero-shot counting. In this paper, we propose a generalized framework for both few-shot and zero-shot object counting based on detection. Our framework combines the superior advantages of two foundation models without compromising their zero-shot capability: (\textbf{i}) SAM to segment all possible objects as mask proposals, and (\textbf{ii}) CLIP to classify proposals to obtain accurate object counts. However, this strategy meets the obstacles of efficiency overhead and the small crowded objects that cannot be localized and distinguished. To address these issues, our framework, termed PseCo, follows three steps: point, segment, and count. Specifically, we first propose a class-agnostic object localization to provide accurate but least point prompts for SAM, which consequently not only reduces computation costs but also avoids missing small objects. Furthermore, we propose a generalized object classification that leverages CLIP image/text embeddings as the classifier, following a hierarchical knowledge distillation to obtain discriminative classifications among hierarchical mask proposals. Extensive experimental results on FSC-147, COCO, and LVIS demonstrate that PseCo achieves state-of-the-art performance in both few-shot/zero-shot object counting/detection. Code: https://github.com/Hzzone/PseCo

In the context of pose-invariant object recognition and retrieval, we demonstrate that it is possible to achieve significant improvements in performance if both the category-based and the object-identity-based embeddings are learned simultaneously during training. In hindsight, that sounds intuitive because learning about the categories is more fundamental than learning about the individual objects that correspond to those categories. However, to the best of what we know, no prior work in pose-invariant learning has demonstrated this effect. This paper presents an attention-based dual-encoder architecture with specially designed loss functions that optimize the inter- and intra-class distances simultaneously in two different embedding spaces, one for the category embeddings and the other for the object level embeddings. The loss functions we have proposed are pose-invariant ranking losses that are designed to minimize the intra-class distances and maximize the inter-class distances in the dual representation spaces. We demonstrate the power of our approach with three challenging multi-view datasets, ModelNet-40, ObjectPI, and FG3D. With our dual approach, for single-view object recognition, we outperform the previous best by 20.0% on ModelNet40, 2.0% on ObjectPI, and 46.5% on FG3D. On the other hand, for single-view object retrieval, we outperform the previous best by 33.7% on ModelNet40, 18.8% on ObjectPI, and 56.9% on FG3D.

Deep neural networks for learning Symmetric Positive Definite (SPD) matrices are gaining increasing attention in machine learning. Despite the significant progress, most existing SPD networks use traditional Euclidean classifiers on an approximated space rather than intrinsic classifiers that accurately capture the geometry of SPD manifolds. Inspired by Hyperbolic Neural Networks (HNNs), we propose Riemannian Multinomial Logistics Regression (RMLR) for the classification layers in SPD networks. We introduce a unified framework for building Riemannian classifiers under the metrics pulled back from the Euclidean space, and showcase our framework under the parameterized Log-Euclidean Metric (LEM) and Log-Cholesky Metric (LCM). Besides, our framework offers a novel intrinsic explanation for the most popular LogEig classifier in existing SPD networks. The effectiveness of our method is demonstrated in three applications: radar recognition, human action recognition, and electroencephalography (EEG) classification. The code is available at https://github.com/GitZH-Chen/SPDMLR.git.

In deep metric learning for visual recognition, the calibration of distance thresholds is crucial for achieving the desired model performance in the true positive rates (TPR) or true negative rates (TNR). However, calibrating this threshold presents significant challenges in open-world scenarios, where the test classes are entirely disjoint from those encountered during training. We define the problem of finding distance thresholds for a trained embedding model to achieve target performance metrics over unseen open-world test classes as open-world threshold calibration. Existing posthoc threshold calibration methods, reliant on inductive inference and requiring a calibration dataset with a similar distance distribution as the test data, often prove ineffective in open-world scenarios. To address this, we introduce OpenGCN, a Graph Neural Network-based transductive threshold calibration method with enhanced adaptability and robustness in open-world scenarios. OpenGCN learns to predict pairwise connectivity for the unlabeled test instances embedded in a graph to determine its TPR and TNR at various distance thresholds, allowing for transductive inference of the distance thresholds which also incorporates test information. Extensive experiments across open-world visual recognition benchmarks validate OpenGCN's superiority over existing posthoc calibration methods for open-world threshold calibration.

We investigate whether region-based representations are effective for recognition. Regions were once a mainstay in recognition approaches, but pixel and patch-based features are now used almost exclusively. We show that recent class-agnostic segmenters like SAM can be effectively combined with strong unsupervised representations like DINOv2 and used for a wide variety of tasks, including semantic segmentation, object-based image retrieval, and multi-image analysis. Once the masks and features are extracted, these representations, even with linear decoders, enable competitive performance, making them well suited to applications that require custom queries. The compactness of the representation also makes it well-suited to video analysis and other problems requiring inference across many images.

Single-modal object re-identification (ReID) faces great challenges in maintaining robustness within complex visual scenarios. In contrast, multi-modal object ReID utilizes complementary information from diverse modalities, showing great potentials for practical applications. However, previous methods may be easily affected by irrelevant backgrounds and usually ignore the modality gaps. To address above issues, we propose a novel learning framework named \textbf{EDITOR} to select diverse tokens from vision Transformers for multi-modal object ReID. We begin with a shared vision Transformer to extract tokenized features from different input modalities. Then, we introduce a Spatial-Frequency Token Selection (SFTS) module to adaptively select object-centric tokens with both spatial and frequency information. Afterwards, we employ a Hierarchical Masked Aggregation (HMA) module to facilitate feature interactions within and across modalities. Finally, to further reduce the effect of backgrounds, we propose a Background Consistency Constraint (BCC) and an Object-Centric Feature Refinement (OCFR). They are formulated as two new loss functions, which improve the feature discrimination with background suppression. As a result, our framework can generate more discriminative features for multi-modal object ReID. Extensive experiments on three multi-modal ReID benchmarks verify the effectiveness of our methods. The code is available at https://github.com/924973292/EDITOR.

Text-to-image person re-identification (ReID) retrieves pedestrian images according to textual descriptions. Manually annotating textual descriptions is time-consuming, restricting the scale of existing datasets and therefore the generalization ability of ReID models. As a result, we study the transferable text-to-image ReID problem, where we train a model on our proposed large-scale database and directly deploy it to various datasets for evaluation. We obtain substantial training data via Multi-modal Large Language Models (MLLMs). Moreover, we identify and address two key challenges in utilizing the obtained textual descriptions. First, an MLLM tends to generate descriptions with similar structures, causing the model to overfit specific sentence patterns. Thus, we propose a novel method that uses MLLMs to caption images according to various templates. These templates are obtained using a multi-turn dialogue with a Large Language Model (LLM). Therefore, we can build a large-scale dataset with diverse textual descriptions. Second, an MLLM may produce incorrect descriptions. Hence, we introduce a novel method that automatically identifies words in a description that do not correspond with the image. This method is based on the similarity between one text and all patch token embeddings in the image. Then, we mask these words with a larger probability in the subsequent training epoch, alleviating the impact of noisy textual descriptions. The experimental results demonstrate that our methods significantly boost the direct transfer text-to-image ReID performance. Benefiting from the pre-trained model weights, we also achieve state-of-the-art performance in the traditional evaluation settings. The code and dataset for this study will be released.

For text-to-video retrieval (T2VR), which aims to retrieve unlabeled videos by ad-hoc textual queries, CLIP-based methods currently lead the way. Compared to CLIP4Clip which is efficient and compact, state-of-the-art models tend to compute video-text similarity through fine-grained cross-modal feature interaction and matching, putting their scalability for large-scale T2VR applications into doubt. We propose $\textbf{TeachCLIP}$, enabling a CLIP4Clip based student network to learn from more advanced yet computationally intensive models. In order to create a learning channel to convey fine-grained cross-modal knowledge from a heavy model to the student, we add to CLIP4Clip a simple Attentional frame-Feature Aggregation (AFA) block, which by design adds no extra storage / computation overhead at the retrieval stage. Frame-text relevance scores calculated by the teacher network are used as soft labels to $\textbf{supervise}$ the attentive weights produced by AFA. Extensive experiments on multiple public datasets justify the viability of the proposed method. TeachCLIP has the same efficiency and compactness as CLIP4Clip, yet has near-SOTA effectiveness.

In this work, we offer a novel perspective on detecting out-of-distribution (OOD) samples and propose a sample-aware model selection algorithm by exploring the model zoo.Our algorithm automatically selects pre-trained models for each test input, effectively identifying OOD samples and classifying them. If no model is selected, the test input is classified as an in-distribution (ID) sample.We provide theoretical analysis that demonstrates our approach maintains the true positive rate of ID samples, and accurately identifies OOD samples with a high probability, given a sufficiently large model zoo. We conducted extensive experiments, which showed that our method leverages the complementarity among individual model detectors to consistently improve the effectiveness of OOD sample identification. Compared to baseline methods, our approach improved the relative performance by 65.40\% and 26.96\% on the CIFAR10 and ImageNet benchmarks, respectively.

Comprehensive modeling of the surrounding 3D world is crucial for the success of autonomous driving. However, existing perception tasks like object detection, road structure segmentation, depth & elevation estimation, and open-set object localization each only focus on a small facet of the holistic 3D scene understanding task. This divide-and-conquer strategy simplifies the algorithm development process but comes at the cost of losing an end-to-end unified solution to the problem. In this work, we address this limitation by studying camera-based 3D panoptic segmentation, aiming to achieve a unified occupancy representation for camera-only 3D scene understanding. To achieve this, we introduce a novel method called PanoOcc, which utilizes voxel queries to aggregate spatiotemporal information from multi-frame and multi-view images in a coarse-to-fine scheme, integrating feature learning and scene representation into a unified occupancy representation. We have conducted extensive ablation studies to validate the effectiveness and efficiency of the proposed method. Our approach achieves new state-of-the-art results for camera-based semantic segmentation and panoptic segmentation on the nuScenes dataset. Furthermore, our method can be easily extended to dense occupancy prediction and has demonstrated promising performance on the Occ3D benchmark. The code will be made available at https://github.com/Robertwyq/PanoOcc.

Salient object detection (SOD) and camouflaged object detection (COD) are related yet distinct binary mapping tasks. These tasks involve multiple modalities, sharing commonalities and unique cues. Existing research often employs intricate task-specific specialist models, potentially leading to redundancy and suboptimal results. We introduce VSCode, a generalist model with novel 2D prompt learning, to jointly address four SOD tasks and three COD tasks. We utilize VST as the foundation model and introduce 2D prompts within the encoder-decoder architecture to learn domain and task-specific knowledge on two separate dimensions. A prompt discrimination loss helps disentangle peculiarities to benefit model optimization. VSCode outperforms state-of-the-art methods across six tasks on 26 datasets and exhibits zero-shot generalization to unseen tasks by combining 2D prompts, such as RGB-D COD.

Existing methods for asymmetric image retrieval employ a rigid pairwise similarity constraint between the query network and the larger gallery network. However, these one-to-one constraint approaches often fail to maintain retrieval order consistency, especially when the query network has limited representational capacity. To overcome this problem, we introduce the Decoupled Differential Distillation (D3still) framework. This framework shifts from absolute one-to-one supervision to optimizing the relational differences in pairwise similarities produced by the query and gallery networks, thereby preserving a consistent retrieval order across both networks. Our method involves computing a pairwise similarity differential matrix within the gallery domain, which is then decomposed into three components: feature representation knowledge, inconsistent pairwise similarity differential knowledge, and consistent pairwise similarity differential knowledge. This strategic decomposition aligns the retrieval ranking of the query network with the gallery network effectively. Extensive experiments on various benchmark datasets reveal that D3still surpasses state-of-the-art methods in asymmetric image retrieval. Code is available at https://github.com/SCY-X/D3still.

Event cameras, characterized by high temporal resolution, high dynamic range, low power consumption, and high pixel bandwidth, offer unique capabilities for object detection in specialized contexts. Despite these advantages, the inherent sparsity and asynchrony of event data pose challenges to existing object detection algorithms. Spiking Neural Networks (SNNs), inspired by the way the human brain codes and processes information, offer a potential solution to these difficulties. However, their performance in object detection using event cameras is limited in current implementations. In this paper, we propose the Spiking Fusion Object Detector (SFOD), a simple and efficient approach to SNN-based object detection. Specifically, we design a Spiking Fusion Module, achieving the first-time fusion of feature maps from different scales in SNNs applied to event cameras. Additionally, through integrating our analysis and experiments conducted during the pretraining of the backbone network on the NCAR dataset, we delve deeply into the impact of spiking decoding strategies and loss functions on model performance. Thereby, we establish state-of-the-art classification results based on SNNs, achieving 93.7\% accuracy on the NCAR dataset. Experimental results on the GEN1 detection dataset demonstrate that the SFOD achieves a state-of-the-art mAP of 32.1\%, outperforming existing SNN-based approaches. Our research not only underscores the potential of SNNs in object detection with event cameras but also propels the advancement of SNNs. Code is available at https://github.com/yimeng-fan/SFOD.

Concealed Object Detection (COD) aims to identify objects visually embedded in their background. Existing COD datasets and methods predominantly focus on animals or humans, ignoring the agricultural domain, which often contains numerous, small, and concealed crops with severe occlusions.In this paper, we introduce Concealed Crop Detection (CCD), which extends classic COD to agricultural domains.Experimental study shows that unimodal data provides insufficient information for CCD.To address this gap, we first collect a large-scale RGB-D dataset, ACOD-12K, containing high-resolution crop images and depth maps.Then, we propose a foundational framework named Recurrent Iterative Segmentation Network (RISNet).To tackle the challenge of dense objects, we employ multi-scale receptive fields to capture objects of varying sizes, thus enhancing the detection performance for dense objects.By fusing depth features, our method can acquire spatial information about concealed objects to mitigate disturbances caused by intricate backgrounds and occlusions. Furthermore, our model adopts a multi-stage iterative approach, using predictions from each stage as gate attention to reinforce position information, thereby improving the detection accuracy for small objects. Extensive experimental results demonstrate that our RISNet achieves new state-of-the-art performance on both newly proposed CCD and classic COD tasks.All resources will be available at https://github.com/Kki2Eve/RISNet.

This paper introduces a novel approach to learning instance segmentation using extreme points, i.e., the topmost, leftmost, bottommost, and rightmost points, of each object. These points are readily available in the modern bounding box annotation process while offering strong clues for precise segmentation, and thus allows to improve performance at the same annotation cost with box-supervised methods.Our work considers extreme points as a part of the true instance mask and propagates them to identify potential foreground and background points, which are all together used for training a pseudo label generator. Then pseudo labels given by the generator are in turn used for supervised learning of our final model. On three public benchmarks, our method significantly outperforms existing box-supervised methods, further narrowing the gap with its fully supervised counterpart. In particular, our model generates high-quality masks when a target object is separated into multiple parts, where previous box-supervised methods often fail.

Text-to-image (T2I) generative models, in recent times, have given rise to various applications owing to their ability to generate high-fidelity, photo-realistic images. However, the question of how to utilize T2I models in fundamental image classification remains open. Currently, a common method to enhance image classification involves expanding the training set with synthetic datasets generated by T2I models. In this study, we explore the drawbacks and effectiveness of two main expansion methods, namely, distillation-based and data augmentation methods. Our findings indicate that these methods face challenges in generating both faithful and diverse images for domain-specific concepts. To address this issue, we propose a novel inter-class data augmentation method, Diff-Mix. Diff-Mix expands the dataset by conducting image translation in an inter-class manner, significantly improving the diversity of synthetic data. We observe an improved trade-off between faithfulness and diversity with Diff-Mix, resulting in a significant performance gain across various image classification settings, including few-shot classification, conventional classification, and long-tail classification, particularly for domain-specific datasets.

Recent advancements have shown the potential of leveraging both point clouds and images to localize anomalies. Nevertheless, their applicability in industrial manufacturing is often constrained by significant drawbacks, such as the use of memory banks, which leads to a substantial increase in terms of memory footprint and inference times. We propose a novel light and fast framework that learns to map features from one modality to the other on nominal samples and detect anomalies by pinpointing inconsistencies between observed and mapped features. Extensive experiments show that our approach achieves state-of-the-art detection and segmentation performance in both the standard and few-shot settings on the MVTec 3D-AD dataset while achieving faster inference and occupying less memory than previous multimodal AD methods. Furthermore, we propose a layer pruning technique to improve memory and time efficiency with a marginal sacrifice in performance.

Image-based mirror detection has recently undergone rapid research due to its significance in applications such as robotic navigation, semantic segmentation and scene reconstruction. Recently, VMD-Net was proposed as the first video mirror detection technique, by modeling dual correspondences between the inside and outside of the mirror both spatially and temporally. However, this approach is not reliable, as correspondences can occur completely inside or outside of the mirrors. In addition, the proposed dataset VMD-D contains many small mirrors, limiting its applicability to real-world scenarios. To address these problems, we developed a more challenging dataset that includes mirrors of various shapes and sizes at different locations of the frames, providing a better reflection of real-world scenarios. Next, we observed that the motions between the inside and outside of the mirror are often inconsistent. For instance, when moving in front of a mirror, the motion inside the mirror is often much smaller than the motion outside due to increased depth perception. With these observations, we propose modeling inconsistent motion cues to detect mirrors, and a new network with two novel modules. The Motion Attention Module (MAM) explicitly models inconsistent motions around mirrors via optical flow, and the Motion-Guided Edge Detection Module (MEDM) uses motions to guide mirror edge feature learning. Experimental results on our proposed dataset show that our method outperforms state-of-the-arts. The code and dataset are available at https://github.com/AlexAnthonyWarren/MG-VMD.

3D visual grounding aims to localize 3D objects described by free-form language sentences. Following the detection-then-matching paradigm, existing methods mainly focus on embedding object attributes in unimodal feature extraction and multimodal feature fusion, to enhance the discriminability of the proposal feature for accurate grounding. However, most of them ignore the explicit interaction of multiple attributes, causing a bias in unimodal representation and misalignment in multimodal fusion. In this paper, we propose a multi-attribute aware Transformer for 3D visual grounding, learning the multi-attribute interactions to refine the intra-modal and inter-modal grounding cues. Specifically, we first develop an attribute causal analysis module to quantify the causal effect of different attributes for the final prediction, which provides powerful supervision to correct the misleading attributes and adaptively capture other discriminative features. Then, we design an exchanging-based multimodal fusion module, which dynamically replaces tokens with low attribute attention between modalities before directly integrating low-dimensional global features. This ensures an attribute-level multimodal information fusion and helps align the language and vision details more efficiently for fine-grained multimodal features. Extensive experiments show that our method can achieve state-of-the-art performance on ScanRefer and SR3D/NR3D datasets.

Automatic anomaly detection based on visual cues holds practical significance in various domains, such as manufacturing and product quality assessment. This paper introduces a new conditional anomaly detection problem, which involves identifying anomalies in a query image by comparing it to a reference shape. To address this challenge, we have created a large dataset, BrokenChairs-180K, consisting of around 180K images, with diverse anomalies, geometries, and textures paired with 8,143 reference 3D shapes. To tackle this task, we have proposed a novel transformer-based approach that explicitly learns the correspondence between the query image and reference 3D shape via feature alignment and leverages a customized attention mechanism for anomaly detection. Our approach has been rigorously evaluated through comprehensive experiments, serving as a benchmark for future research in this domain.

The rapidly growing scale of data in practice poses demands on the efficiency of retrieval models. However, for fine-grained image retrieval task, there are inherent contradictions in the design of hashing based efficient models. Firstly, the limited information embedding capacity of low-dimensional binary hash codes, coupled with the detailed information required to describe fine-grained categories, results in a contradiction in feature learning. Secondly, there is also a contradiction between the complexity of fine-grained feature extraction models and retrieval efficiency. To address these issues, in this paper, we propose the characteristics matching based hash codes generation method. Coupled with the cross-layer semantic information transfer module and the multi-region feature embedding module, the proposed method can generate hash codes that effectively capture fine-grained differences among samples while ensuring efficient inference. Extensive experiments on widely-used datasets demonstrate that our method can significantly outperform state-of-the-art methods.

This paper views the DETR's non-duplicate detection ability as a competition result among object queries. Around each object, there are usually multiple queries, within which only a single one can win the chance to become the final detection. Such a competition is hard: while some competing queries initially have very close prediction scores, their leading query has to dramatically enlarge its score superiority after several decoder layers. To help the leading query stands out, this paper proposes EASE-DETR, which eases the competition by introducing bias that favours the leading one. EASE-DETR is very simple: in every intermediate decoder layer, we identify the ''leading / trailing'' relationship between any two queries, and encode this binary relationship into the following decoder layer to amplify the superiority of the leading one. More concretely, the leading query is to be protected from mutual query suppression in the self-attention layer and encouraged to absorb more object features in the cross-attention layer, therefore accelerating to win. Experimental results show that EASE-DETR brings consistent and remarkable improvement to various DETRs.

The goal of Universal Cross-Domain Retrieval (UCDR) is to achieve robust performance in generalized test scenarios, wherein data may belong to strictly unknown domains and categories during training. Recently, pre-trained models with prompt tuning have shown strong generalization capabilities and attained noteworthy achievements in various downstream tasks, such as few-shot learning and video-text retrieval. However, applying them directly to UCDR may not be sufficient to handle both domain shift (i.e., adapting to unfamiliar domains) and semantic shift (i.e., transferring to unknown categories). To this end, we propose Prompting-to-Simulate (ProS), the first method to apply prompt tuning for UCDR. ProS employs a two-step process to simulate Content-aware Dynamic Prompts (CaDP) which can impact models to produce generalized features for UCDR. Concretely, in Prompt Units Learning stage, we introduce two Prompt Units to individually capture domain and semantic knowledge in a mask-and-align way. Then, in Context-aware Simulator Learning stage, we train a Content-aware Prompt Simulator under a simulated test scenario to produce the corresponding CaDP. Extensive experiments conducted on three benchmark datasets show that our method achieves new state-of-the-art performance without bringing excessive parameters. Code is available at https://github.com/fangkaipeng/ProS.

Open world object detection aims to identify objects of unseen categories and incrementally recognize them once their annotations are provided. In distinction to the traditional paradigm that is limited to predefined categories, this setting promises a continual and generalizable way of estimating objectness using class-agnostic information. However, achieving such decorrelation between objectness and class information proves challenging. Without explicit consideration, existing methods usually exhibit low recall on unknown objects and can misclassify them into known classes. To address this problem, we exploit three levels of orthogonality in the detection process: First, the objectness and classification heads are disentangled by operating on separate sets of features that are orthogonal to each other in a devised polar coordinate system. Secondly, a prediction decorrelation loss is introduced to guide the detector towards more general and class-independent prediction. Furthermore, we propose a calibration scheme that helps maintain orthogonality throughout the training process to mitigate catastrophic interference and facilitate incremental learning of previously unseen objects. Our method is comprehensively evaluated on open world and incremental object detection benchmarks, demonstrating its effectiveness in detecting both known and unknown objects. Code and models are available at https://github.com/feifeiobama/OrthogonalDet.

In this paper, we address web-scale visual entity recognition, specifically the task of mapping a given query image to one of the 6 million existing entities in Wikipedia. One way of approaching a problem of such scale is using dual encoder models (e.g. CLIP), where all the entity names and query images are embedded into a unified space, paving the way for an approximate kNN search. Alternatively, it is also possible to re-purpose a captioning model to directly generate the entity names for a given image. In contrast, we introduce a novel Generative Entity Recognition (GER) framework, which given an input image learns to auto-regressively decode a semantic and discriminative “code” identifying the target entity. Our experiments demonstrate the efficacy of this GER paradigm, showcasing state-of-the-art performance on the challenging OVEN benchmark. GER surpasses strong captioning, dual-encoder, visual matching and hierarchical classification baselines, affirming its advantage in tackling the complexities of web-scale recognition.

Deep Convolutional Neural Networks (DCNNs) have achieved remarkable performance in synthetic aperture radar (SAR) object detection, but this comes at the cost of tremendous computational resources, partly due to extracting redundant features within a single convolutional layer. Recent works either delve into model compression methods or focus on the carefully-designed lightweight models, both of which result in performance degradation. In this paper, we propose an efficient convolution module for SAR object detection, called SFS-Conv, which increases feature diversity within each convolutional layer through a shunt-perceive-select strategy. Specifically, we shunt input feature maps into space and frequency aspects. The former perceives the context of various objects by dynamically adjusting receptive field, while the latter captures abundant frequency variations and textural features via fractional Gabor transformer. To adaptively fuse features from space and frequency aspects, a parameter-free feature selection module is proposed to ensure that the most representative and distinctive information are preserved. With SFS-Conv, we build a lightweight SAR object detection network, called SFS-CNet. Experimental results show that SFS-CNet outperforms state-of-the-art (SoTA) models on a series of SAR object detection benchmarks, while simultaneously reducing both the model size and computational cost.

Aiming to enhance the utilization of metric space by the parametric softmax classifier, recent studies suggest replacing it with a non-parametric alternative. Although a non-parametric classifier may provide better metric space utilization, it introduces the challenge of capturing inter-class relationships. A shared characteristic among prior non-parametric classifiers is the static assignment of labels to prototypes during the training, \ie, each prototype consistently represents a class throughout the training course. Orthogonal to previous works, we present a simple yet effective method to optimize the category assigned to each prototype (label-to-prototype assignment) during the training. To this aim, we formalize the problem as a two-step optimization objective over network parameters and label-to-prototype assignment mapping. We solve this optimization using a sequential combination of gradient descent and Bipartide matching. We demonstrate the benefits of the proposed approach by conducting experiments on balanced and long-tail classification problems using different backbone network architectures. In particular, our method outperforms its competitors by 1.22\% accuracy on CIFAR-100, and 2.15\% on ImageNet-200 using a metric space dimension half of the size of its competitors. \href{https://github.com/msed-Ebrahimi/DL2PA_CVPR24}{Code}

Extensive advancements have been made in person ReID through the mining of semantic information. Nevertheless, existing methods that utilize semantic-parts from a single image modality do not explicitly achieve this goal. Whiteness the impressive capabilities in multimodal understanding of Vision Language Foundation Model CLIP, a recent two-stage CLIP-based method employs automated prompt engineering to obtain specific textual labels for classifying pedestrians. However, we note that the predefined soft prompts may be inadequate in expressing the entire visual context and struggle to generalize to unseen classes. This paper presents an end-to-end Prompt-driven Semantic Guidance (PromptSG) framework that harnesses the rich semantics inherent in CLIP. Specifically, we guide the model to attend to regions that are semantically faithful to the prompt. To provide the personalized language descriptions for specific individuals, we propose learning pseudo tokens that represent specific visual context. This design not only facilitates learning fine-grained attribute information but also can inherently leverage language prompts during inference. Without requiring additional labeling efforts, our PromptSG achieves state-of-the-art by over 10\% on MSMT17 and nearly 5\% on the Market-1501 benchmark.

Monocular 3D object detection poses a significant challenge in 3D scene understanding due to its inherently ill-posed nature in monocular depth estimation. Existing methods heavily rely on supervised learning using abundant 3D labels, typically obtained through expensive and labor-intensive annotation on LiDAR point clouds. To tackle this problem, we propose a novel weakly supervised 3D object detection framework named VSRD (Volumetric Silhouette Rendering for Detection) to train 3D object detectors without any 3D supervision but only weak 2D supervision. VSRD consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. In the auto-labeling stage, we represent the surface of each instance as a signed distance field (SDF) and render its silhouette as an instance mask through our proposed instance-aware volumetric silhouette rendering. To directly optimize the 3D bounding boxes through rendering, we decompose the SDF of each instance into the SDF of a cuboid and the residual distance field (RDF) that represents the residual from the cuboid. This mechanism enables us to optimize the 3D bounding boxes in an end-to-end manner by comparing the rendered instance masks with the ground truth instance masks. The optimized 3D bounding boxes serve as effective training data for 3D object detection. We conduct extensive experiments on the KITTI-360 dataset, demonstrating that our method outperforms the existing weakly supervised 3D object detection methods. The code is available at https://github.com/skmhrk1209/VSRD.

Visual scenes are naturally organized in a hierarchy, where a coarse semantic is recursively comprised of several fine details.Exploring such a visual hierarchy is crucial to recognize the complex relations of visual elements, leading to a comprehensive scene understanding.In this paper, we propose a Visual Hierarchy Mapper (Hi-Mapper), a novel approach for enhancing the structured understanding of the pre-trained Deep Neural Networks (DNNs).Hi-Mapper investigates the hierarchical organization of the visual scene by 1) pre-defining a hierarchy tree through the encapsulation of probability densities;and 2) learning the hierarchical relations in hyperbolic space with a novel hierarchical contrastive loss.The pre-defined hierarchy tree recursively interacts with the visual features of the pre-trained DNNs through hierarchy decomposition and encoding procedures, thereby effectively identifying the visual hierarchy and enhancing the recognition of an entire scene.Extensive experiments demonstrate that Hi-Mapper significantly enhances the representation capability of DNNs, leading to an improved performance on various tasks, including image classification and dense prediction tasks.The code is available at \url{https://github.com/kwonjunn01/Hi-Mapper}.

How important is it for training and evaluation sets to not have class overlap in image retrieval? We revisit Google Landmarks v2 clean, the most popular training set, by identifying and removing class overlap with Revisited Oxford and Paris, the most popular training set. By comparing the original and the new RGLDv2-clean on a benchmark of reproduced state-of-the-art methods, our findings are striking. Not only is there a dramatic drop in performance, but it is inconsistent across methods, changing the ranking.What does it take to focus on objects or interest and ignore background clutter when indexing? Do we need to train an object detector and the representation separately? Do we need location supervision? We introduce Single-stage Detect-to-Retrieve (CiDeR), an end-to-end, single-stage pipeline to detect objects of interest and extract a global image representation. We outperform previous state-of-the-art on both existing training sets and the new RGLDv2-clean.

Recent progress in video anomaly detection suggests that the features of appearance and motion play crucial roles in distinguishing abnormal patterns from normal ones. However, we note that the effect of spatial scales of anomalies is ignored. The fact that many abnormal events occur in limited localized regions and severe background noise interferes with the learning of anomalous changes. Meanwhile, most existing methods are limited by coarse-grained modeling approaches, which are inadequate for learning highly discriminative features to discriminate subtle differences between small-scale anomalies and normal patterns. To this end, this paper address multi-scale video anomaly detection by multi-grained spatio-temporal representation learning. We utilize video continuity to design three proxy tasks to perform feature learning at both coarse-grained and fine-grained levels, i.e., continuity judgment, discontinuity localization, and missing frame estimation. In particular, we formulate missing frame estimation as a contrastive learning task in feature space instead of a reconstruction task in RGB space to learn highly discriminative features. Experiments show that our proposed method outperforms state-of-the-art methods on four datasets, especially in scenes with small-scale anomalies.

This paper introduces a novel approach for high-quality deepfake detection called Localized Artifact Attention Network (LAA-Net). Existing methods for high-quality deepfake detection are mainly based on a supervised binary classifier coupled with an implicit attention mechanism. As a result, they do not generalize well to unseen manipulations. To handle this issue, two main contributions are made. First, an explicit attention mechanism within a multi-task learning framework is proposed. By combining heatmap-based and self-consistency attention strategies, LAA-Net is forced to focus on a few small artifact-prone vulnerable regions. Second, an Enhanced Feature Pyramid Network (E-FPN) is proposed as a simple and effective mechanism for spreading discriminative low-level features into the final feature output, with the advantage of limiting redundancy. Experiments performed on several benchmarks show the superiority of our approach in terms of Area Under the Curve (AUC) and Average Precision (AP). The code is available at https://github.com/10Ring/LAA-Net.

Oriented object detection has been developed rapidly in the past few years, where rotation equivariance is crucial for detectors to predict rotated boxes. It is expected that the prediction can maintain the corresponding rotation when objects rotate, but severe mutation in angular prediction is sometimes observed when objects rotate near the boundary angle, which is well-known boundary discontinuity problem. The problem has been long believed to be caused by the sharp loss increase at the angular boundary, and widely used joint-optim IoU-like methods deal with this problem by loss-smoothing. However, we experimentally find that even state-of-the-art IoU-like methods actually fail to solve the problem. On further analysis, we find that the key to solution lies in encoding mode of the smoothing function rather than in joint or independent optimization. In existing IoU-like methods, the model essentially attempts to fit the angular relationship between box and object, where the break point at angular boundary makes the predictions highly unstable. To deal with this issue, we propose a dual-optimization paradigm for angles. We decouple reversibility and joint-optim from single smoothing function into two distinct entities, which for the first time achieves the objectives of both correcting angular boundary and blending angle with other parameters. Extensive experiments on multiple datasets show that boundary discontinuity problem is well-addressed. Moreover, typical IoU-like methods are improved to the same level without obvious performance gap. The code is available at https://github.com/hangxu-cv/cvpr24acm.

With the transformative impact of the Transformer, DETR pioneered the application of the encoder-decoder architecture to object detection. A collection of follow-up research, e.g., Deformable DETR, aims to enhance DETR while adhering to the encoder-decoder design. In this work, we revisit the DETR series through the lens of Faster R-CNN. We find that the DETR resonates with the underlying principles of Faster R-CNN's RPN-refiner design but benefits from end-to-end detection owing to the incorporation of Hungarian matching. We systematically adapt the Faster R-CNN towards the Deformable DETR, by integrating or repurposing each component of Deformable DETR, and note that Deformable DETR's improved performance over Faster R-CNN is attributed to the adoption of advanced modules such as a superior proposal refiner (e.g., deformable attention rather than RoI Align). When viewing the DETR through the RPN-refiner paradigm, we delve into various proposal refinement techniques such as deformable attention, cross attention, and dynamic convolution. These proposal refiners cooperate well with each other; thus, we synergistically combine them to establish a Hybrid Proposal Refiner (HPR). Our HPR is versatile and can be incorporated into various DETR detectors. For instance, by integrating HPR to a strong DETR detector, we achieve an AP of 54.9 on the COCO benchmark, utilizing a ResNet-50 backbone and a 36-epoch training schedule. The code and models will be accessible to the public.

Open-vocabulary object detection (OVD) has been studied with Vision-Language Models (VLMs) to detect novel objects beyond the pre-trained categories. Previous approaches improve the generalization ability to expand the knowledge of the detector, using 'positive' pseudo-labels with additional 'class' names, e.g., sock, iPod, and alligator. To extend the previous methods in two aspects, we propose Retrieval-Augmented Losses and visual Features (RALF). Our method retrieves related 'negative' classes and augments loss functions. Also, visual features are augmented with 'verbalized concepts' of classes, e.g., worn on the feet, handheld music player, and sharp teeth. Specifically, RALF consists of two modules: Retrieval Augmented Losses (RAL) and Retrieval-Augmented visual Features (RAF). RAL constitutes two losses reflecting the semantic similarity with negative vocabularies. In addition, RAF augments visual features with the verbalized concepts from a large language model (LLM). Our experiments demonstrate the effectiveness of RALF on COCO and LVIS benchmark datasets. We achieve improvement up to 3.4 box $AP_{50}^{\text{N}}$ on novel categories of the COCO dataset and 3.6 mask $AP_{\text{r}}$ gains on the LVIS dataset. Code is available at https://github.com/mlvlab/RALF.

Camera-based person re-identification (ReID) systems have been widely applied in the field of public security. However, cameras often lack the perception of 3D morphological information of human and are susceptible to various limitations, such as inadequate illumination, complex background, and personal privacy. In this paper, we propose a LiDAR-based ReID framework, ReID3D, that utilizes pre-training strategy to retrieve features of 3D body shape and introduces Graph-based Complementary Enhancement Encoder for extracting comprehensive features. Due to the lack of LiDAR datasets, we build LReID, the first LiDAR-based person ReID dataset, which is collected in several outdoor scenes with variations in natural conditions. Additionally, we introduce LReID-sync, a simulated pedestrian dataset designed for pre-training encoders with tasks of point cloud completion and shape parameter learning. Extensive experiments on LReID show that ReID3D achieves exceptional performance with a rank-1 accuracy of 94.0, highlighting the significant potential of LiDAR in addressing person ReID tasks. To the best of our knowledge, we are the first to propose a solution for LiDAR-based ReID. The code and dataset are available at https://github.com/GWxuan/ReID3D.

In this paper, we make the first attempt at achieving the cross-modal (i.e., image-to-events) adaptation for event-based object recognition without accessing any labeled source image data owning to privacy and commercial issues. Tackling this novel problem is non-trivial due to the novelty of event cameras and the distinct modality gap between images and events. In particular, as only the source model is available, a hurdle is how to extract the knowledge from the source model by only using the unlabeled target event data while achieving knowledge transfer. To this end, we propose a novel framework, dubbed EventDance for this unsupervised source-free cross-modal adaptation problem. Importantly, inspired by event-to-video reconstruction methods, we propose a reconstruction-based modality bridging (RMB) module, which reconstructs intensity frames from events in a self-supervised manner. This makes it possible to build up a surrogate domain in the image modality to extract the knowledge (i.e., labels) from the source model. We then propose a multi-representation knowledge adaptation (MKA) module that transfers the knowledge to target models learning events with multiple representation types for fully exploring the spatiotemporal information of events. The two modules connecting the source and target models are mutually updated so as to achieve the best performance. Experiments on three benchmark datasets with two adaption settings show that EventDance is on par with prior methods utilizing the source data.

In Re-identification (ReID), recent advancements yield noteworthy progress in both unimodal and cross-modal retrieval tasks. However, the challenge persists in developing a unified framework that could effectively handle varying multimodal data, including RGB, infrared, sketches, and textual information. Additionally, the emergence of large-scale models shows promising performance in various vision tasks but the foundation model in ReID is still blank. In response to these challenges, a novel multimodal learning paradigm for ReID is introduced, referred to as All-in-One (AIO), which harnesses a frozen pre-trained big model as an encoder, enabling effective multimodal retrieval without additional fine-tuning. The diverse multimodal data in AIO are seamlessly tokenized into a unified space, allowing the modality-shared frozen encoder to extract identity-consistent features comprehensively across all modalities. Furthermore, a meticulously crafted ensemble of cross-modality heads is designed to guide the learning trajectory. AIO is the first framework to perform all-in-one ReID, encompassing four commonly used modalities. Experiments on cross-modal and multimodal ReID reveal that AIO not only adeptly handles various modal data but also excels in challenging contexts, showcasing exceptional performance in zero-shot and domain generalization scenarios. Code will be available at: https://github.com/lihe404/AIO.

The design of deep network architectures and training methods in computer vision has been well-explored. However, in almost all cases the images have been used as provided, with little exploration of pre-processing steps beyond normalization and data augmentation. Virtually all images posted on the web or captured by devices are processed for viewing by humans. Is the pipeline used for humans also best for use by computers and deep networks?The human visual system uses logarithmic sensors; differences and sums correspond to ratios and products. Features in log space will be invariant to intensity changes and robust to color balance changes. Log RGB space also reveals structure that is corrupted by typical pre-processing.We explore using linear and log RGB data for training standard backbone architectures on an image classification task using data derived directly from RAW images to guarantee its integrity. We found that networks trained on log RGB data exhibit improved performance on an unmodified test set and invariance to intensity and color balance modifications without additional training or data augmentation.Furthermore, we found that the gains from using high quality log data could also be partially or fully realized from data in 8-bit sRGB-JPG format by inverting the sRGB transform and taking the log. These results imply existing databases may benefit from this type of pre-processing.While working with log data, we found it was critical to retain the integrity of the log relationships and that networks using log data train best with meta-parameters different than those used for sRGB or linear data.Finally, we introduce a new10-category 10k RAW image data set [RAW10] for image classification and other purposes to enable further the exploration of log RGB as an input format for deep networks in computer vision.

Out-of-distribution (OOD) detection methods often exploit auxiliary outliers to train model identifying OOD samples, especially discovering challenging outliers from auxiliary outliers dataset to improve OOD detection. However, they may still face limitations in effectively distinguishing between the most challenging OOD samples that are much like in-distribution (ID) data, i.e., ID-like samples. To this end, we propose a novel OOD detection framework that discovers ID-like outliers using CLIP from the vicinity space of the ID samples, thus helping to identify these most challenging OOD samples. Then a prompt learning framework is proposed that utilizes the identified ID-like outliers to further leverage the capabilities of CLIP for OOD detection. Benefiting from the powerful CLIP, we only need a small number of ID samples to learn the prompts of the model without exposing other auxiliary outlier datasets. By focusing on the most challenging ID-like OOD samples and elegantly exploiting the capabilities of CLIP, our method achieves superior few-shot learning performance on various real-world image datasets (e.g., in 4-shot OOD detection on the ImageNet-1k dataset, our method reduces the average FPR95 by 12.16\% and improves the average AUROC by 2.76\%, compared to state-of-the-art methods).

Recently, infrared small target detection (IRSTD) has been dominated by deep-learning-based methods. However, these methods mainly focus on the design of complex model structures to extract discriminative features, leaving the loss functions for IRSTD under-explored. For example, the widely used Intersection over Union (IoU) and Dice losses lack sensitivity to the scales and locations of targets, limiting the detection performance of detectors. In this paper, we focus on boosting detection performance with a more effective loss but a simpler model structure. Specifically, we first propose a novel Scale and Location Sensitive (SLS) loss to handle the limitations of existing losses: 1) for scale sensitivity, we compute a weight for the IoU loss based on target scales to help the detector distinguish targets with different scales: 2) for location sensitivity, we introduce a penalty term based on the center points of targets to help the detector localize targets more precisely. Then, we design a simple Multi-Scale Head to the plain U-Net (MSHNet). By applying SLS loss to each scale of the predictions, our MSHNet outperforms existing state-of-the-art methods by a large margin. In addition, the detection performance of existing detectors can be further improved when trained with our SLS loss, demonstrating the effectiveness and generalization of our SLS loss.

In this paper, we revisit techniques for uncertainty estimation within deep neural networks and consolidate a suite of techniques to enhance their reliability. Our investigation reveals that an integrated application of diverse techniques--spanning model regularization, classifier and optimization--substantially improves the accuracy of uncertainty predictions in image classification tasks. The synergistic effect of these techniques culminates in our novel SURE approach. We rigorously evaluate SURE against the benchmark of failure prediction, a critical testbed for uncertainty estimation efficacy. Our results showcase a consistently better performance than models that individually deploy each technique, across various datasets and model architectures. When applied to real-world challenges, such as data corruption, label noise, and long-tailed class distribution, SURE exhibits remarkable robustness, delivering results that are superior or on par with current state-of-the-art specialized methods. Particularly on Animal-10N and Food-101N for learning with noisy labels, SURE achieves state-of-the-art performance without any task-specific adjustments. This work not only sets a new benchmark for robust uncertainty estimation but also paves the way for its application in diverse, real-world scenarios where reliability is paramount. Our code is available at \url{https://yutingli0606.github.io/SURE/}.

Anomaly detection is a challenging computer vision task in industrial scenario. Advancements in deep learning constantly revolutionize vision-based anomaly detection methods, and considerable progress has been made in both supervised and self-supervised anomaly detection. The commonly-used pipeline is to optimize the model by constraining the feature embeddings using a distance-based loss function. However, these methods work in Euclidean space, and they cannot well exploit the data lied in non-Euclidean space. In this paper, we are the first to explore anomaly detection task in hyperbolic space that is a representative of non-Euclidean space, and propose a hyperbolic anomaly detection (HypAD) method. Specifically, we first extract image features and then map them from Euclidean space to hyperbolic space, where the hyperbolic distance metric is employed to optimize the proposed HypAD. Extensive experiments on the benchmarking datasets including MVTec AD and VisA show that our HypAD approach obtains the state-of-the-art performance, demonstrating the effectiveness of our HypAD and the promise of investigating anomaly detection in hyperbolic space.

Human intelligence can retrieve any person according to both visual and language descriptions. However, the current computer vision community studies specific person re-identification (ReID) tasks in different scenarios separately, which limits the applications in the real world. This paper strives to resolve this problem by proposing a new instruct-ReID task that requires the model to retrieve images according to the given image or language instructions. Our instruct-ReID is a more general ReID setting, where existing \textbf{6} ReID tasks can be viewed as special cases by designing different instructions. We propose a large-scale OmniReID benchmark and an adaptive triplet loss as a baseline method to facilitate research in this new setting. Experimental results show that the proposed multi-purpose ReID model, trained on our OmniReID benchmark without fine-tuning, can improve +0.5%, +0.6%, +7.7% mAP on Market1501, MSMT17, CUHK03 for traditional ReID, +6.4%, +7.1%, +11.2% mAP on PRCC, VC-Clothes, LTCC for clothes-changing ReID, +11.7% mAP on COCAS+ real2 for clothes template based clothes-changing ReID when using only RGB images, +24.9% mAP on COCAS+ real2 for our newly defined language-instructed ReID, +4.3% on LLCM for visible-infrared ReID, +2.6% on CUHK-PEDES for text-to-image ReID. The datasets, the model, and code shall be released upon acceptance.

Person re-identification (re-ID) is a challenging task that aims to learn discriminative features for person retrieval. In person re-ID, Jaccard distance is a widely used distance metric, especially in re-ranking and clustering scenarios. However, we discover that camera variation has a significant negative impact on the reliability of Jaccard distance. In particular, Jaccard distance calculates the distance based on the overlap of relevant neighbors. Due to camera variation, intra-camera samples dominate the relevant neighbors, which reduces the reliability of the neighbors by introducing intra-camera negative samples and excluding inter-camera positive samples. To overcome this problem, we propose a novel camera-aware Jaccard (CA-Jaccard) distance that leverages camera information to enhance the reliability of Jaccard distance. Specifically, we design camera-aware k-reciprocal nearest neighbors (CKRNNs) to find k-reciprocal nearest neighbors on the intra-camera and inter-camera ranking lists, which improves the reliability of relevant neighbors and guarantees the contribution of inter-camera samples in the overlap. Moreover, we propose a camera-aware local query expansion (CLQE) to mine reliable samples in relevant neighbors by exploiting camera variation as a strong constraint and assign these samples higher weights in overlap, further improving the reliability. Our CA-Jaccard distance is simple yet effective and can serve as a general distance metric for person re-ID methods with high reliability and low computational cost. Extensive experiments demonstrate the effectiveness of our method. Code is available at https://github.com/chen960/CA-Jaccard/.

The zero-shot performance of existing vision-language models (VLMs) such as CLIP is limited by the availability of large-scale, aligned image and text datasets in specific domains. In this work, we leverage two complementary sources of information---descriptions of categories generated by large language models (LLMs) and abundant, fine-grained image classification datasets---to improve the zero-shot classification performance of VLMs across fine-grained domains. On the technical side, we develop methods to train VLMs with this ``bag-level" image-text supervision. We find that simply using these attributes at test-time does not improve performance, but our training strategy, for example, on the iNaturalist dataset, leads to an average improvement of 4-5\% in zero-shot classification accuracy for novel categories of birds and flowers. Similar improvements are observed in domains where a subset of the categories was used to fine-tune the model. By prompting LLMs in various ways, we generate descriptions that capture visual appearance, habitat, and geographical regions and pair them with existing attributes such as the taxonomic structure of the categories. We systematically evaluate their ability to improve zero-shot categorization in natural domains. Our findings suggest that geographic priors can be just as effective and are complementary to visual appearance. Our method also outperforms prior work on prompt-based tuning of VLMs. We plan to release the benchmark, consisting of 7 datasets, which will contribute to future research in zero-shot recognition.

From content moderation to wildlife conservation, the number of applications that require models to recognize nuanced or subjective visual concepts is growing. Traditionally, developing classifiers for such concepts requires substantial manual effort measured in hours, days, or even months to identify and annotate data needed for training. Even with recently proposed Agile Modeling techniques, which enable rapid bootstrapping of image classifiers, users are still required to spend 30 minutes or more of monotonous, repetitive data labeling just to train a single classifier. Drawing on Fiske’s Cognitive Miser theory, we propose a new framework that alleviates manual effort by replacing human labeling with natural language interactions – reducing the total effort required to define a concept by an order of magnitude: from labeling 2,000 images to only 100 plus some natural language interactions. Our framework leverages the recent advances in foundation models, both large language models and vision-language models, to carve out the concept space through conversation and by automatically labeling training data points. Most importantly, our framework eliminates the need for crowd-sourced annotations. Moreover, our framework ultimately produces light-weight classification models that are deployable in cost-sensitive scenarios. Across 15 subjective concepts and across 2 public image classification datasets, our trained models outperform traditional Agile Modeling as well as state-of-the-art zero-shot classification models like ALIGN, CLIP, CuPL, and large visual question answering models like PaLI-X.

Computer vision in unconstrained outdoor scenarios must tackle challenging high dynamic range (HDR) scenes and rapidly changing illumination conditions. Existing methods address this problem with multi-capture HDR sensors and a hardware image signal processor (ISP) that produces a single fused image as input to a downstream neural network. The output of the HDR sensor is a set of low dynamic range (LDR) exposures, and the fusion in the ISP is performed in image space and typically optimized for human perception on a display. Preferring tonemapped content with smooth transition regions over detail (and noise) in the resulting image, this image fusion does typically not preserve all information from the LDR exposures that may be essential for downstream computer vision tasks. In this work, we depart from conventional HDR image fusion and propose a learned task-driven fusion in the feature domain. Instead of using a single companded image, we introduce a novel local cross-attention fusion mechanism that exploits semantic features from all exposures – learned in an end-to-end fashion with supervision from downstream detection losses. The proposed method outperforms all tested conventional HDR exposure fusion and auto-exposure methods in challenging automotive HDR scenarios.

DETR-like methods have significantly increased detection performance in an end-to-end manner. The mainstream two-stage frameworks of them perform dense self-attention and select a fraction of queries for sparse cross-attention, which is proven effective for improving performance but also introduces a heavy computational burden and high dependence on stable query selection. This paper demonstrates that suboptimal two-stage selection strategies result in scale bias and redundancy due to the mismatch between selected queries and objects in two-stage initialization. To address these issues, we propose hierarchical salience filtering refinement, which performs transformer encoding only on filtered discriminative queries, for a better trade-off between computational efficiency and precision. The filtering process overcomes scale bias through a novel scale-independent salience supervision. To compensate for the semantic misalignment among queries, we introduce elaborate query refinement modules for stable two-stage initialization. Based on above improvements, the proposed Salience DETR achieves significant improvements of +4.0% AP, +0.2% AP, +4.4% AP on three challenging task-specific detection datasets, as well as 49.2% AP on COCO 2017 with less FLOPs. The code is available at https://github.com/xiuqhou/Salience-DETR.

Existing prompt learning methods have demonstrated certain capabilities in Out-of-Distribution (OOD) detection, but their lack of perception of OOD images in the target dataset can lead to mismatches between OOD images and In-Distribution (ID) categories, leading to a high false positive rate. To address this issue, we introduce a novel OOD detection method, named `NegPrompt', which is designed to learn a set of negative prompts, each representing a negative connotation of a given class label, to delineate the boundaries between ID and OOD images. It learns such negative prompts with ID data only, eliminating its reliance on external data.Further, current methods assume the availability of samples of all ID classes, rendering them ineffective in open-vocabulary learning scenarios where the inference stage can contain novel ID classes not present in the training data. In contrast, our learned negative prompts are transferable to novel class labels, Experiments on various ImageNet-based benchmarks demonstrate that NegPrompt surpasses state-of-the-art prompt-learning-based OOD detection methods and maintains a consistent lead in hard OOD detection in closed- and open-vocabulary classification scenarios.

Place recognition is crucial for unmanned vehicles in terms of localization and mapping. Recent years have witnessed numerous explorations in the field, where 2D cameras and 3D LiDARs are mostly employed. Despite their admirable performance, they may encounter challenges in adverse weather such as rain and fog. Hopefully, 4D millimeter-wave Radar emerges as a promising alternative, as its longer wavelength makes it virtually immune to interference from tiny particles of fog and rain. Therefore, in this work, we propose a novel 4D Radar place recognition model, TransLoc4D, based on sparse convolution and Transformer structures. Specifically, a MinkLoc4D backbone is first proposed to leverage the geometric, intensity, and velocity information from 4D Radar scans. While mainstream 3D LiDAR solutions merely capture geometric structures of point clouds, MinkLoc4D explores the intensity and velocity properties of 4D Radar scans and demonstrates their effectiveness. After feature extraction, a Transformer layer is introduced to enhance local features, where linear self-attention captures the long-range dependency of point cloud, alleviating its sparsity and noise. To validate TransLoc4D, we construct two datasets and set up benchmarks for 4D place recognition. Experiments show TransLoc4D is feasible and can robustly deal with dynamic and adverse environments.

Single-domain generalization aims to learn a model from single source domain data to achieve generalized performance on other unseen target domains. Existing works primarily focus on improving the generalization ability of static networks. However, static networks are unable to dynamically adapt to the diverse variations in different image scenes, leading to limited generalization capability. Different scenes exhibit varying levels of complexity, and the complexity of images further varies significantly in cross-domain scenarios. In this paper, we propose a dynamic object-centric perception network based on prompt learning, aiming to adapt to the variations in image complexity. Specifically, we propose an object-centric gating module based on prompt learning to focus attention on the object-centric features guided by the various scene prompts. Then, with the object-centric gating masks, the dynamic selective module dynamically selects highly correlated feature regions in both spatial and channel dimensions enabling the model to adaptively perceive object-centric relevant features, thereby enhancing the generalization capability. Extensive experiments were conducted on single-domain generalization tasks in image classification and object detection. The experimental results demonstrate that our approach outperforms state-of-the-art methods, which validates the effectiveness and generally of our proposed method.

Open-set supervised anomaly detection (OSAD) - a recently emerging anomaly detection area - aims at utilizing a few samples of anomaly classes seen during training to detect unseen anomalies (i.e., samples from open-set anomaly classes), while effectively identifying the seen anomalies. Benefiting from the prior knowledge illustrated by the seen anomalies, current OSAD methods can often largely reduce false positive errors. However, these methods are trained in a closed-set setting and treat the anomaly examples as from a homogeneous distribution, rendering them less effective in generalizing to unseen anomalies that can be drawn from any distribution. This paper proposes to learn heterogeneous anomaly distributions using the limited anomaly examples to address this issue. To this end, we introduce a novel approach, namely Anomaly Heterogeneity Learning (AHL), that simulates a diverse set of heterogeneous anomaly distributions and then utilizes them to learn a unified heterogeneous abnormality model in surrogate open-set environments. Further, AHL is a generic framework that existing OSAD models can plug and play for enhancing their abnormality modeling. Extensive experiments on nine real-world anomaly detection datasets show that AHL can 1) substantially enhance different state-of-the-art OSAD models in detecting seen and unseen anomalies, and 2) effectively generalize to unseen anomalies in new domains. Code is available at https://github.com/mala-lab/AHL.

We propose a unified approach to simultaneously addressing the conventional setting of binary deepfake classification and a more challenging scenario of uncovering what facial components have been forged as well as the exact order of the manipulations. To solve the former task, we consider multiple instance learning (MIL) that takes each image as a bag and its patches as instances. A positive bag corresponds to a forged image that includes at least one manipulated patch (i.e., a pixel in the feature map). The formulation enables us to estimate the probability of an input image being a fake one and to establish the corresponding contrastive MIL loss. On the other hand, tackling the component-wise deepfake problem can be reduced to solving multi-label prediction, but the requirement to recover the manipulation order further complicates the learning task into a multi-label ranking problem. We resolve this difficulty by designing a tailor-made loss term to enforce that the rank order of the predicted multi-label probabilities respects the ground-truth order of the sequential modifications of a deepfake image. For experiments and comparisons with other relevant techniques, we provide extensive results and ablation studies to demonstrate that the proposed method is an overall more comprehensive solution to deepfake detection.

Softassign is a pivotal method in graph matching and other learning tasks. Many softassign-based algorithms exhibit performance sensitivity to a parameter in the softassign. However, tuning the parameter is challenging and almost done empirically. This paper proposes an adaptive softassign method for graph matching by analyzing the relationship between the objective score and the parameter. This method can automatically tune the parameter based on a given error bound to guarantee accuracy. The Hadamard-Equipped Sinkhorn formulas introduced in this study significantly enhance the efficiency and stability of the adaptive softassign. Moreover, these formulas can also be used in optimal transport problems. The resulting adaptive softassign graph matching algorithm enjoys significantly higher accuracy than previous state-of-the-art large graph matching algorithms while maintaining comparable efficiency.

Unsupervised fine-grained image hashing aims to learn compact binary hash codes in unsupervised settings, addressing challenges posed by large-scale datasets and dependence on supervision. In this paper, we first identify a granularity gap between generic and fine-grained datasets for unsupervised hashing methods, highlighting the inadequacy of conventional self-supervised learning for fine-grained visual objects. To bridge this gap, we propose the Asymmetric Augmented Self-Supervised Learning (A$^2$-SSL) method, comprising three modules. The asymmetric augmented SSL module employs suitable augmentation strategies for positive/negative views, preventing fine-grained category confusion inherent in conventional SSL. Part-oriented dense contrastive learning utilizes the Fisher Vector framework to capture and model fine-grained object parts, enhancing unsupervised representations through part-level dense contrastive learning. Self-consistent hash code learning introduces a reconstruction task aligned with the self-consistency principle, guiding the model to emphasize comprehensive features, particularly fine-grained patterns. Experimental results on five benchmark datasets demonstrate the superiority of A$^2$-SSL over existing methods, affirming its efficacy in unsupervised fine-grained image hashing.

The task of Visual Place Recognition (VPR) aims to match a query image against references from an extensive database of images from different places, relying solely on visual cues. State-of-the-art pipelines focus on the aggregation of features extracted from a deep backbone, in order to form a global descriptor for each image. In this context, we introduce SALAD (Sinkhorn Algorithm for Locally Aggregated Descriptors), which reformulates NetVLAD's soft-assignment of local features to clusters as an optimal transport problem. In SALAD, we consider both feature-to-cluster and cluster-to-feature relations and we also introduce a 'dustbin' cluster, designed to selectively discard features deemed non-informative, enhancing the overall descriptor quality. Additionally, we leverage and fine-tune DINOv2 as a backbone, which provides enhanced description power for the local features, and dramatically reduces the required training time. As a result, our single-stage method not only surpasses single-stage baselines in public VPR datasets, but also surpasses two-stage methods that add a re-ranking with significantly higher cost. Models and code will be made public upon acceptance and are provided as supplemental material.

Identifying the chemical structure from a graphical representation, or image, of a molecule is a challenging pattern recognition task that would greatly benefit drug development. Yet, existing methods for chemical structure recognition do not typically generalize well, and show diminished effectiveness when confronted with domains where data is sparse, or costly to generate, such as hand-drawn molecule images. To address this limitation, we propose a new chemical structure recognition tool that delivers state-of-the-art performance and can adapt to new domains with a limited number of data samples and supervision. Unlike previous approaches, our method provides atom-level localization, and can therefore segment the image into the different atoms and bonds. Our model is the first model to perform OCSR with atom-level entity detection with only SMILES supervision. Through rigorous and extensive benchmarking, we demonstrate the preeminence of our chemical structure recognition approach in terms of data efficiency, accuracy, and atom-level entity prediction.

Ultra-fine-grained visual categorization (Ultra-FGVC) aims at distinguishing highly similar sub-categories within fine-grained objects, such as different soybean cultivars. Compared to traditional fine-grained visual categorization, Ultra-FGVC encounters more hurdles due to the small inter-class and large intra-class variation. Given these challenges, relying on human annotation for Ultra-FGVC is impractical. To this end, our work introduces a novel task termed Ultra-Fine-Grained Novel Class Discovery (UFG-NCD), which leverages partially annotated data to identify new categories of unlabeled images for Ultra-FGVC. To tackle this problem, we devise a Region-Aligned Proxy Learning (RAPL) framework, which comprises a Channel-wise Region Alignment (CRA) module and a Semi-Supervised Proxy Learning (SemiPL) strategy. The CRA module is designed to extract and utilize discriminative features from local regions, facilitating knowledge transfer from labeled to unlabeled classes. Furthermore, SemiPL strengthens representation learning and knowledge transfer with proxy-guided supervised learning and proxy-guided contrastive learning. Such techniques leverage class distribution information in the embedding space, improving the mining of subtle differences between labeled and unlabeled ultra-fine-grained classes. Extensive experiments demonstrate that RAPL significantly outperforms baselines across various datasets, indicating its effectiveness in handling the challenges of UFG-NCD. Code is available at https://github.com/SSDUT-Caiyq/UFG-NCD.

In various domains such as surveillance and smart retail, pedestrian retrieval, centering on person re-identification (Re-ID), plays a pivotal role. Existing Re-ID methodologies often overlook subtle internal attribute variations, which are crucial for accurately identifying individuals with changing appearances. In response, our paper introduces the Attribute-Guided Pedestrian Retrieval (AGPR) task, focusing on integrating specified attributes with query images to refine retrieval results. Although there has been progress in attribute-driven image retrieval, there remains a notable gap in effectively blending robust Re-ID models with intra-class attribute variations. To bridge this gap, we present the Attribute-Guided Transformer-based Pedestrian Retrieval (ATPR) framework. ATPR adeptly merges global ID recognition with local attribute learning, ensuring a cohesive linkage between the two. Furthermore, to effectively handle the complexity of attribute interconnectivity, ATPR organizes attributes into distinct groups and applies both inter-group correlation and intra-group decorrelation regularizations. Our extensive experiments on a newly established benchmark using the RAP dataset demonstrate the effectiveness of ATPR within the AGPR paradigm.

The surge in multi-modal data has propelled cross-modal matching to the forefront of research interest. However, the challenge lies in the laborious and expensive process of curating a large and accurately matched multi-modal dataset. Commonly sourced from the Internet, these datasets often suffer from a significant presence of mismatched data, impairing the performance of matching models. To address this problem, we introduce a novel regularization approach named Equivariant Similarity Consistency (ESC), which can facilitate robust clean and noisy data separation and improve the training for cross-modal matching. Intuitively, our method posits that the semantic variations caused by image changes should be proportional to those caused by text changes for any two matched samples. Accordingly, we first calculate the ESC by comparing image and text semantic variations between a set of elaborated anchor points and other undivided training data. Then, pairs with high ESC are filtered out as noisy correspondence pairs. We implement our method by combining the ESC with a traditional hinge-based triplet loss. Extensive experiments on three widely used datasets, including Flickr30K, MS-COCO, and Conceptual Captions, verify the effectiveness of our method.

The recently proposed SparseFormer architecture provides an alternative approach to visual understanding by utilizing a significantly lower number of visual tokens via adjusting RoIs, greatly reducing computational costs while still achieving promising performance. However, training SparseFormers from scratch is still expensive, and scaling up the number of parameters can be challenging. In this paper, we propose to bootstrap SparseFormers from ViT-based vision foundation models in a simple and efficient way. Since the majority of SparseFormer blocks are the standard transformer ones, we can inherit weights from large-scale pre-trained vision transformers and freeze them as much as possible. Therefore, we only need to train the SparseFormer-specific lightweight focusing transformer to adjust token RoIs and fine-tune a few early pre-trained blocks to align the final token representation. In such a way, we can bootstrap SparseFormer architectures from various large-scale pre-trained models (e.g., IN-21K pre-trained AugRegs or CLIPs) using a rather smaller amount of training samples (e.g., IN-1K) and without labels or captions within just a few hours. As a result, the bootstrapped unimodal SparseFormer (from AugReg-ViT-L/16-384) can reach $84.9\%$ accuracy on IN-1K with only $49$ tokens, and the multimodal SparseFormer from CLIPs also demonstrates notable zero-shot performance with highly reduced computational cost without seeing any caption during the bootstrapping procedure. In addition, CLIP-bootstrapped SparseFormers, which align the output space with language without seeing a word, can serve as efficient vision encoders in multimodal large language models. Code and models are available at https://github.com/showlab/sparseformer}{https://github.com/showlab/sparseformer

The consistency training (CT)-based semi-supervised learning (SSL) bites state-of-the-art performance on SSL-based image classification. However, the existing CT-based SSL methods do not highlight the non-Euclidean characteristics and class-wise varieties of embedding spaces in an SSL model, thus they cannot fully utilize the effectiveness of CT. Thus, we propose a metric tensor-based consistency regularization, exploiting the class-variant geometrical structure of embeddings on the high-dimensional feature space. The proposed method not only minimizes the prediction discrepancy between different views of a given image but also estimates the intrinsic geometric curvature of embedding spaces by employing the global and local metric tensors. The global metric tensor is used to globally estimate the class-invariant embeddings from the whole data distribution while the local metric tensor is exploited to estimate the class-variant embeddings of each cluster. The two metric tensors are optimized by the consistency regularization based on the weak and strong augmentation strategy. The proposed method provides the highest classification accuracy on average compared to the existing state-of-the-art SSL methods on conventional datasets.

In this work, we tackle the problem of domain generalization for object detection, specifically focusing on the scenario where only a single source domain is available. We propose an effective approach that involves two key steps: diversifying the source domain and aligning detections based on class prediction confidence and localization.Firstly, we demonstrate that by carefully selecting a set of augmentations, a base detector can outperform existing methods for single domain generalization by a good margin. This highlights the importance of domain diversification in improving the performance of object detectors.Secondly, we introduce a method to align detections from multiple views, considering both classification and localization outputs. This alignment procedure leads to better generalized and well-calibrated object detector models, which are crucial for accurate decision-making in safety-critical applications.Our approach is detector-agnostic and can be seamlessly applied to both single-stage and two-stage detectors.To validate the effectiveness of our proposed methods, we conduct extensive experiments and ablations on challenging domain-shift scenarios. The results consistently demonstrate the superiority of our approach compared to existing methods.We will publicly release our code and models to facilitate further research in this area.

In Visual Place Recognition (VPR) the pose of a query image is estimated by comparing the image to a map of reference images with known reference poses. As is typical for image retrieval problems, a feature extractor maps the query and reference images to a feature space, where a nearest neighbor search is then performed. However, till recently little attention has been given to quantifying the confidence that a retrieved reference image is a correct match. Highly certain but incorrect retrieval can lead to catastrophic failure of VPR-based localization pipelines. This work compares for the first time the main approaches for estimating the image-matching uncertainty, including the traditional retrieval-based uncertainty estimation, more recent data-driven aleatoric uncertainty estimation, and the compute-intensive geometric verification by local feature matching. We further formulate a simple baseline method, ``SUE'', which unlike the other methods considers the freely-available poses of the reference images in the map. Our experiments reveal that a simple L2-distance between the query and reference descriptors is already a better estimate of image-matching uncertainty than current data-driven approaches. SUE outperforms the other efficient uncertainty estimation methods, and its uncertainty estimates complement the computationally expensive geometric verification approach. Future works for uncertainty estimation in VPR should consider the baselines discussed in this work.

Automating visual inspection in industrial production lines is essential for increasing product quality across various industries. Anomaly detection (AD) methods serve as robust tools for this purpose. However, existing public datasets primarily consist of images without anomalies, limiting the practical application of AD methods in production settings. To address this challenge, we present (1) the Valeo Anomaly Dataset (VAD), a novel real-world industrial dataset comprising 5000 images, including 2000 instances of challenging real defects across more than 20 subclasses. Acknowledging that traditional AD methods struggle with this dataset, we introduce (2) Segmentation-based Anomaly Detector (SegAD). First, SegAD leverages anomaly maps as well as segmentation maps to compute local statistics. Next, SegAD uses these statistics and an optional supervised classifier score as input features for a Boosted Random Forest (BRF) classifier, yielding the final anomaly score. Our SegAD achieves state-of-the-art performance on both VAD (+2.1% AUROC) and the VisA dataset (+0.4% AUROC). The code and the models are publicly available.

Computer vision models normally witness degraded performance when deployed in real-world scenarios, due to unexpected changes in inputs that were not accounted for during training. Data augmentation is commonly used to address this issue, as it aims to increase data variety and reduce the distribution gap between training and test data. However, common visual augmentations might not guarantee extensive robustness of computer vision models. In this paper, we propose Auxiliary Fourier-basis Augmentation (AFA), a complementary technique targeting augmentation in the frequency domain and filling the robustness gap left by visual augmentations. We demonstrate the utility of augmentation via Fourier-basis additive noise in a straightforward and efficient adversarial setting. Our results show that AFA benefits the robustness of models against common corruptions, OOD generalization, and consistency of performance of models against increasing perturbations, with negligible deficit to the standard performance of models. It can be seamlessly integrated with other augmentation techniques to further boost performance. Codes and models are available at \href{https://github.com/nis-research/afa-augment}{https://github.com/nis-research/afa-augment}.

Due to the depth degradation effect in residual connections, many efficient Vision Transformers models that rely on stacking layers for information exchange often fail to form sufficient information mixing, leading to unnatural visual perception. To address this issue, in this paper, we propose Aggregated Attention, a biomimetic design-based token mixer that simulates biological foveal vision and continuous eye movement while enabling each token on the feature map to have a global perception. Furthermore, we incorporate learnable tokens that interact with conventional queries and keys, which further diversifies the generation of affinity matrices beyond merely relying on the similarity between queries and keys. Our approach does not rely on stacking for information exchange, thus effectively avoiding depth degradation and achieving natural visual perception. Additionally, we propose Convolutional GLU, a channel mixer that bridges the gap between GLU and SE mechanism, which empowers each token to have channel attention based on its nearest neighbor image features, enhancing local modeling capability and model robustness. We combine aggregated attention and convolutional GLU to create a new visual backbone called TransNeXt. Extensive experiments demonstrate that our TransNeXt achieves state-of-the-art performance across multiple model sizes. At a resolution of $224^2$, TransNeXt-Tiny attains an ImageNet accuracy of 84.0\%, surpassing ConvNeXt-B with 69\% fewer parameters. Our TransNeXt-Base achieves an ImageNet accuracy of 86.2\% and an ImageNet-A accuracy of 61.6\% at a resolution of $384^2$, a COCO object detection mAP of 57.1, and an ADE20K semantic segmentation mIoU of 54.7.

Annotating datasets for object detection is an expensive and time-consuming endeavor. To minimize this burden, active learning (AL) techniques are employed to select the most informative samples for annotation within a constrained “annotation budget”. Traditional AL strategies typically rely on model uncertainty or sample diversity for query sampling, while more advanced methods have focused on developing AL-specific object detector architectures to enhance performance. However, these specialized approaches are not readily adaptable to different object detectors due to the significant engineering effort required for integration. To overcome this challenge, we introduce Plug and Play Active Learning (PPAL), a simple and effective AL strategy for object detection. PPAL is a two-stage method comprising uncertainty-based and diversity-based sampling phases. In the first stage, our Difficulty Calibrated Uncertainty Sampling leverage a category-wise difficulty coefficient that combines both classification and localisation difficulties to re-weight instance uncertainties, from which we sample a candidate pool for the subsequent diversity-based sampling. In the second stage, we propose Category Conditioned Matching Similarity to better compute the similarities of multi-instance images as ensembles of their instance similarities, which is used by the k-Means++ algorithm to sample the final AL queries. PPAL makes no change to model architectures or detector training pipelines; hence it can be easily generalized to different object detectors. We benchmark PPAL on the MS-COCO and Pascal VOC datasets using different detector architectures and show that our method outperforms prior work by a large margin. Code is available at https://github.com/ChenhongyiYang/PPAL

In visual place recognition, accurately identifying and matching images of locations under varying environmental conditions and viewpoints remains a significant challenge. In this paper, we introduce a new technique, called Bag-of-Queries (BoQ), which learns a set of global queries, designed to capture universal place-specific attributes. Unlike existing techniques that employ self-attention and generate the queries directly from the input, BoQ employ distinct learnable global queries, which probe the input features via cross-attention, ensuring consistent information aggregation. In addition, this technique provides an interpretable attention mechanism and integrates with both CNN and Vision Transformer backbones. The performance of BoQ is demonstrated through extensive experiments on 14 large-scale benchmarks. It consistently outperforms current state-of-the-art techniques including NetVLAD, MixVPR and EigenPlaces. Moreover, despite being a global retrieval technique (one-stage), BoQ surpasses two-stage retrieval methods, such as Patch-NetVLAD, TransVPR and R2Former, all while being orders of magnitude faster and more efficient. The code and model weights are publicly available at https://github.com/amaralibey/Bag-of-Queries.

Open-set recognition (OSR) methods aim to identify whether or not a test example belongs to a category observed during training. Depending on how visually similar a test example is to the training categories, the OSR task can be easy or extremely challenging. However, the vast majority of previous work has studied OSR in the presence of large, coarse-grained semantic shifts. In contrast, many real-world problems are inherently fine-grained, which means that test examples may be highly visually similar to the training categories. Motivated by this observation, we investigate three aspects of OSR: label granularity, similarity between the open- and closed-sets, and the role of hierarchical supervision during training. To study these dimensions, we curate new open-set splits of a large fine-grained visual categorization dataset. Our analysis results in several interesting findings, including: (i) the best OSR method to use is heavily dependent on the degree of semantic shift present, and (ii) hierarchical representation learning can improve coarse-grained OSR, but has little effect on fine-grained OSR performance. To further enhance fine-grained OSR performance, we propose a hierarchy-adversarial learning method to discourage hierarchical structure in the representation space, which results in a perhaps counter-intuitive behaviour, and a relative improvement in fine-grained OSR of up to 2% in AUROC and 7% in AUPR over standard training. Code and data are available: langnico.github.io/fine-grained-osr.

Human-Object Interaction (HOI) detection plays a crucial role in visual scene comprehension. In recent advancements, two-stage detectors have taken a prominent position. However, they are encumbered by two primary challenges. First, the misalignment between feature representation and relation reasoning gives rise to a deficiency in discriminative features crucial for interaction detection. Second, due to sparse annotation, the second-stage interaction head generates numerous candidate $<$human, object$>$ pairs, with only a small fraction receiving supervision. Towards these issues, we propose a hybrid learning method based on pose-aware HOI feature refinement. Specifically, we devise pose-aware feature refinement that encodes spatial features by considering human body pose characteristics. It can direct attention towards key regions, ultimately offering a wealth of fine-grained features imperative for HOI detection. Further, we introduce a hybrid learning method that combines HOI triplets with probabilistic soft labels supervision, which is regenerated from decoupled verb-object pairs. This method explores the implicit connections between the interactions, enhancing model generalization without requiring additional data. Our method establishes state-of-the-art performance on HICO-DET benchmark and excels notably in detecting rare HOIs.

This paper explores the problem of Generalist Anomaly Detection (GAD), aiming to train one single detection model that can generalize to detect anomalies in diverse datasets from different application domains without any further training on the target data. Some recent studies have shown that large pre-trained Visual-Language Models (VLMs) like CLIP have strong generalization capabilities on detecting industrial defects from various datasets, but their methods rely heavily on handcrafted text prompts about defects, making them difficult to generalize to anomalies in other applications, e.g., medical image anomalies or semantic anomalies in natural images. In this work, we propose to train a GAD model with few-shot normal images as sample prompts for AD on diverse datasets on the fly. To this end, we introduce a novel approach that learns an in-context residual learning model for GAD, termed InCTRL. It is trained on an auxiliary dataset to discriminate anomalies from normal samples based on a holistic evaluation of the residuals between query images and few-shot normal sample prompts. Regardless of the datasets, per definition of anomaly, larger residuals are expected for anomalies than normal samples, thereby enabling InCTRL to generalize across different domains without further training. Comprehensive experiments on nine AD datasets are performed to establish a GAD benchmark that encapsulate the detection of industrial defect anomalies, medical anomalies, and semantic anomalies in both one-vs-all and multi-class setting, on which InCTRL is the best performer and significantly outperforms state-of-the-art competing methods. Code is available at https://github.com/mala-lab/InCTRL.

In the context of autonomous navigation of terrestrial robots, the creation of realistic models for agent dynamics and sensing is a widespread habit in the robotics literature and in commercial applications, where they are used for model based control and/or for localization and mapping. The more recent Embodied AI literature, on the other hand, focuses on modular or end-to-end agents trained in simulators like Habitat or AI-Thor, where the emphasis is put on photo-realistic rendering and scene diversity, but high-fidelity robot motion is assigned a less privileged role. The resulting sim2real gap significantly impacts transfer of the trained models to real robotic platforms.In this work we explore end-to-end training of agents in simulation in settings which minimize the sim2real gap both, in sensing and in actuation. Our agent directly predicts (discretized) velocity commands, which are maintained through closed-loop control in the real robot. The behavior of the real robot (including the underlying low-level controller) is identified and simulated in a modified Habitat simulator. Noise models for odometry and localization further contribute in lowering the sim2real gap. We evaluate on real navigation scenarios, explore different localization and point goal calculation methods and report significant gains in performance and robustness compared to prior work.

Successfully addressing a wide variety of tasks is a core ability of autonomous agents, which requires flexibly adapting the underlying decision-making strategies and, as we argue in this work, also adapting the underlying perception modules. An analogical argument would be the human visual system, which uses top-down signals to focus attention determined by the current task. Similarly, in this work, we adapt pre-trained large vision models conditioned on specific downstream tasks in the context of multi-task policy learning. We introduce task-conditioned adapters that do not require finetuning any pre-trained weights, combined with a single policy trained with behavior cloning and capable of addressing multiple tasks. We condition the policy and visual adapters on task embeddings, which can be selected at inference if the task is known, or alternatively inferred from a set of example demonstrations. To this end, we propose a new optimization-based estimator. We evaluate the method on a wide variety of tasks of the CortexBench benchmark and show that, compared to existing work, it can be addressed with a single policy. In particular, we demonstrate that adapting visual features is a key design choice and that the method generalizes to unseen tasks given visual demonstrations.

3D correspondence, i.e., a pair of 3D points, is a fundamental concept in computer vision. A set of 3D correspondences, when equipped with compatibility edges, forms a correspondence graph. This graph is a critical component in several state-of-the-art 3D point cloud registration approaches, e.g., the one based on maximal cliques (MAC). However, its properties have not been well understood. So we present the first study that introduces graph signal processing into the domain of correspondence graph. We exploit the generalized degree signal on correspondence graph and pursue sampling strategies that preserve high-frequency components of this signal. To address time-consuming singular value decomposition in deterministic sampling, we resort to a stochastic approximate sampling strategy. As such, the core of our method is the stochastic spectral sampling of correspondence graph. As an application, we build a complete 3D registration algorithm termed as FastMAC, that reaches real-time speed while leading to little to none performance drop. Through extensive experiments, we validate that FastMAC works for both indoor and outdoor benchmarks. For example, FastMAC can accelerate MAC by 80 times while maintaining high registration success rate on KITTI. Codes will be publicly available.

We present FoundationPose, a unified foundation model for 6D object pose estimation and tracking, supporting both model-based and model-free setups. Our approach can be instantly applied at test-time to a novel object without fine-tuning, as long as its CAD model is given, or a small number of reference images are captured. We bridge the gap between these two setups with a neural implicit representation that allows for effective novel view synthesis, keeping the downstream pose estimation modules invariant under the same unified framework. Strong generalizability is achieved via large-scale synthetic training, aided by a large language model (LLM), a novel transformer-based architecture, and contrastive learning formulation. Extensive evaluation on multiple public datasets involving challenging scenarios and objects indicate our unified approach outperforms existing methods specialized for each task by a large margin. In addition, it even achieves comparable results to instance-level methods despite the reduced assumptions. Project page: https://foundationpose.github.io/foundationpose/

We address the challenge of generating 3D articulated objects in a controllable fashion. Currently, modeling articulated 3D objects is either achieved through laborious manual authoring, or using methods from prior work that are hard to scale and control directly. We leverage the interplay between part shape, connectivity, and motion using a denoising diffusion-based method with attention modules designed to extract correlations between part attributes. Our method takes an object category label and a part connectivity graph as input and generates an object's geometry and motion parameters. The generated objects conform to user-specified constraints on the object category, part shape, and part articulation. Our experiments show that our method outperforms the state-of-the-art in articulated object generation, producing more realistic objects while conforming better to user constraints.

There are five types of trajectory prediction tasks: deterministic, stochastic, domain adaptation, momentary observation, and few-shot. These associated tasks are defined by various factors, such as the length of input paths, data split and pre-processing methods. Interestingly, even though they commonly take sequential coordinates of observations as input and infer future paths in the same coordinates as output, designing specialized architectures for each task is still necessary. For the other task, generality issues can lead to sub-optimal performances. In this paper, we propose SingularTrajectory, a diffusion-based universal trajectory prediction framework to reduce the performance gap across the five tasks. The core of SingularTrajectory is to unify a variety of human dynamics representations on the associated tasks. To do this, we first build a Singular space to project all types of motion patterns from each task into one embedding space. We next propose an adaptive anchor working in the Singular space. Unlike traditional fixed anchor methods that sometimes yield unacceptable paths, our adaptive anchor enables correct anchors, which are put into a wrong location, based on a traversability map. Finally, we adopt a diffusion-based predictor to further enhance the prototype paths using a cascaded denoising process. Our unified framework ensures the generality across various benchmark settings such as input modality, and trajectory lengths. Extensive experiments on five public benchmarks demonstrate that SingularTrajectory substantially outperforms existing models, highlighting its effectiveness in estimating general dynamics of human movements. Code is publicly available at https://github.com/InhwanBae/SingularTrajectory.

Grasp detection is a persistent and intricate challenge with various industrial applications. Recently, many methods and datasets have been proposed to tackle the grasp detection problem. However, most of them do not consider using natural language as a condition to detect the grasp poses. In this paper, we introduce Grasp-Anything++, a new language-driven grasp detection dataset featuring 1M samples, over 3M objects, and upwards of 10M grasping instructions. We utilize foundation models to create a large-scale scene corpus with corresponding images and grasp prompts. We approach the language-driven grasp detection task as a conditional generation problem. Drawing on the success of diffusion models in generative tasks and given that language plays a vital role in this task, we propose a new language-driven grasp detection method based on diffusion models. Our key contribution is the contrastive training objective, which explicitly contributes to the denoising process to detect the grasp pose given the language instructions. We illustrate that our approach is theoretically supportive. The intensive experiments show that our method outperforms state-of-the-art approaches and allows real-world robotic grasping. Finally, we demonstrate our large-scale dataset enables zero-short grasp detection and is a challenging benchmark for future work.

Image-goal navigation is a challenging task that requires an agent to navigate to a goal indicated by an image in unfamiliar environments. Existing methods utilizing diverse scene memories suffer from inefficient exploration since they use all historical observations for decision-making without considering the goal-relevant fraction. To address this limitation, we present MemoNav, a novel memory model for image-goal navigation, which utilizes a working memory-inspired pipeline to improve navigation performance. Specifically, we employ three types of navigation memory. The node features on a map are stored in the short-term memory (STM), as these features are dynamically updated. A forgetting module then retains the informative STM fraction to increase efficiency. We also introduce long-term memory (LTM) to learn global scene representations by progressively aggregating STM features. Subsequently, a graph attention module encodes the retained STM and the LTM to generate working memory (WM) which contains the scene features essential for efficient navigation. The synergy among these three memory types boosts navigation performance by enabling the agent to learn and leverage goal-relevant scene features within a topological map. Our evaluation on multi-goal tasks demonstrates that MemoNav significantly outperforms previous methods across all difficulty levels in both Gibson and Matterport3D scenes. Qualitative results further illustrate that MemoNav plans more efficient routes.

The practicality of 3D object pose estimation remains limited for many applications due to the need for prior knowledge of a 3D model and a training period for new objects. To address this limitation, we propose an approach that takes a single image of a new object as input and predicts the relative pose of this object in new images without prior knowledge of the object's 3D model and without requiring training time for new objects and categories. We achieve this by training a model to directly predict discriminative embeddings for viewpoints surrounding the object. This prediction is done using a simple U-Net architecture with attention and conditioned on the desired pose, which yields extremely fast inference. We compare our approach to state-of-the-art methods and show it outperforms them both in terms of accuracy and robustness.

In this work, we propose a novel discriminative framework for dexterous grasp generation, named $\textbf{D}$exterous $\textbf{G}$rasp $\textbf{TR}$ansformer ($\textbf{DGTR}$), capable of predicting a diverse set of feasible grasp poses by processing the object point cloud with only $\textbf{one forward pass}$. We formulate dexterous grasp generation as a set prediction task and design a transformer-based grasping model for it. However, we identify that this set prediction paradigm encounters several optimization challenges in the field of dexterous grasping and results in restricted performance. To address these issues, we propose progressive strategies for both the training and testing phases. First, the dynamic-static matching training (DSMT) strategy is presented to enhance the optimization stability during the training phase. Second, we introduce the adversarial-balanced test-time adaptation (AB-TTA) with a pair of adversarial losses to improve grasping quality during the testing phase. Experimental results on the DexGraspNet dataset demonstrate the capability of DGTR to predict dexterous grasp poses with both high quality and diversity. Notably, while keeping high quality, the diversity of grasp poses predicted by DGTR significantly outperforms previous works in multiple metrics without any data pre-processing. Codes are available at https://github.com/iSEE-Laboratory/DGTR.

Route planning for navigation under partially observability plays a crucial role in modern robotics and autonomous driving. Existing route plannings can be categorized into two main classes: traditional autoregressive and diffusion-based methods. The former often fails due to its myopic nature, while the latter either assumes full observability or struggles to adapt to unfamiliar scenarios, due to strong couplings with behavior cloning from experts. To address these deficiencies, we propose a versatile diffusion-based approach for both 2D and 3D route planning under partial observability. Specifically, our value-guided diffusion policy first generates plans to predict actions across various timesteps, providing ample foresight to the planning. It then employs a differentiable planner with state estimations to derive a value function, directing the agent's exploration and goal-seeking behaviors without seeking for experts while explicitly addressing partial observability. During inference, our policy is further enhanced by a best-plan-selection strategy, substantially boosting the planning success rate. Moreover, we propose projecting point cloud, derived from RGB-D inputs, onto 2D grid-based bird-eye-view maps via semantic segmentation, generalizing to 3D environments. This simple yet effective adaption enables zero-shot transfer from 2D-trained policy to 3D, cutting across the laborious training for 3D policy, and thus certifying our versatility. Experimental results demonstrate our superior performance, particularly in navigating situations beyond expert demonstrations, surpassing state-of-the-art autoregressive and diffusion-based baselines for both 2D and 3D scenarios.

We introduce CyberDemo, a novel approach to robotic imitation learning that leverages simulated human demonstrations for real-world tasks. By incorporating extensive data augmentation in a simulated environment, CyberDemo outperforms traditional in-domain real-world demonstrations when transferred to the real world, handling diverse physical and visual conditions. Regardless of its affordability and convenience in data collection, CyberDemo outperforms baseline methods in terms of success rates across various tasks and exhibits generalizability with previously unseen objects. For example, it can rotate novel tetra-valve and penta-valve, despite human demonstrations only involving tri-valves. Our research demonstrates the significant potential of simulated human demonstrations for real-world dexterous manipulation tasks.

Accuracy and computational efficiency are the most important metrics to Visual Inertial Navigation System (VINS). The existing VINS algorithms with either high accuracy or low computational complexity, are difficult to provide the high precision localization in resource-constrained devices. To this end, we propose a novel filter-based VINS framework named SchurVINS, which could guarantee both high accuracy by building a complete residual model and low computational complexity with Schur complement. Technically, we first formulate the full residual model where Gradient, Hessian and observation covariance are explicitly modeled. Then Schur complement is employed to decompose the full model into ego-motion residual model and landmark residual model. Finally, Extended Kalman Filter (EKF) update is implemented in these two models with high efficiency. Experiments on EuRoC and TUM-VI datasets show that our method notably outperforms state-of-the-art (SOTA) methods in both accuracy and computational complexity. We will open source our experimental code to benefit the community.

This paper proposes Retrieval-Enhanced Asymmetric Diffusion (READ) for image-based robot motion planning. Given an image of the scene, READ retrieves an initial motion from a database of image-motion pairs, and uses a diffusion model to refine the motion for the given scene. Unlike prior retrieval-based diffusion models that require long forward-reverse diffusion paths, READ directly diffuses between the source (retrieved) and target motions, resulting in an efficient diffusion path. A second contribution of READ is its use of asymmetric diffusion, whereby it preserves the kinematic feasibility of the generated motion by forward diffusion in a low-dimensional latent space, while achieving high-resolution motion by reverse diffusion in the original task space using cold diffusion. Experimental results on various manipulation tasks demonstrate that READ outperforms state-of-the-art planning methods, while ablation studies elucidate the contributions of asymmetric diffusion. Code: https://anonymous.4open.science/r/READ-C42F

Embodied agents operating in complex and uncertain environments face considerable challenges. While some advanced agents handle complex manipulation tasks with proficiency, their success often hinges on extensive training data to develop their capabilities. In contrast, humans typically rely on recalling past experiences and analogous situations to solve new problems. Aiming to emulate this human approach in robotics, we introduce the Retrieval-Augmented Embodied Agent (RAEA). This innovative system equips robots with a form of shared memory, significantly enhancing their performance. Our approach integrates a policy retriever, allowing robots to access relevant strategies from an external policy memory bank based on multi-modal inputs. Additionally, a policy generator is employed to assimilate these strategies into the learning process, enabling robots to formulate effective responses to tasks. Extensive testing of RAEA in both simulated and real-world scenarios demonstrates its superior performance over traditional methods, representing a major leap forward in robotic technology.

Collaborative perception in automated vehicles leverages the exchange of information between agents, aiming to elevate perception results. Previous camera-based collaborative 3D perception methods typically employ 3D bounding boxes or bird's eye views as representations of the environment. However, these approaches fall short in offering a comprehensive 3D environmental prediction. To bridge this gap, we introduce the first method for collaborative 3D semantic occupancy prediction. Particularly, it improves local 3D semantic occupancy predictions by hybrid fusion of (i) semantic and occupancy task features, and (ii) compressed orthogonal attention features shared between vehicles. Additionally, due to the lack of a collaborative perception dataset designed for semantic occupancy prediction, we augment a current collaborative perception dataset to include 3D collaborative semantic occupancy labels for a more robust evaluation. The experimental findings highlight that: (i) our collaborative semantic occupancy predictions excel above the results from single vehicles by over 30%, and (ii) models anchored on semantic occupancy outpace state-of-the-art collaborative 3D detection techniques in subsequent perception applications, showcasing enhanced accuracy and enriched semantic-awareness in road environments.

Diffusion generative modeling has become a promising approach for learning robotic manipulation tasks from stochastic human demonstrations. In this paper, we present Diffusion-EDFs, a novel SE(3)-equivariant diffusion-based approach for visual robotic manipulation tasks. We show that our proposed method achieves remarkable data efficiency, requiring only 5 to 10 human demonstrations for effective end-to-end training in less than an hour. Furthermore, our benchmark experiments demonstrate that our approach has superior generalizability and robustness compared to state-of-the-art methods. Lastly, we validate our methods with real hardware experiments. The codes will be released upon acceptance.

Visual-inertial odometry (VIO) has demonstrated remarkable success due to its low-cost and complementary sensors. However, existing VIO methods lack the generalization ability to adjust to different environments and sensor attributes. In this paper, we propose Adaptive VIO, a new monocular visual-inertial odometry that combines online continual learning with traditional nonlinear optimization. Adaptive VIO comprises two networks to predict visual correspondence and IMU bias. Unlike end-to-end approaches that use networks to fuse the features from two modalities (camera and IMU) and predict poses directly, we combine neural networks with visual-inertial bundle adjustment in our VIO system. The optimized estimates will be fed back to the visual and IMU bias networks, refining the networks in a self-supervised manner. Such a learning-optimization-combined framework and feedback mechanism enable the system to perform online continual learning. Experiments demonstrate that our Adaptive VIO manifests adaptive capability on EuRoC and TUM-VI datasets. The overall performance exceeds the currently known learning-based VIO methods and is comparable to the state-of-the-art optimization-based methods.

In this paper we propose an efficient data-driven solution to self-localization within a floorplan.Floorplan data is readily available, long-term persistent and inherently robust to changes in the visual appearance.Our method does not require retraining per map and location or demand a large database of images of the area of interest.We propose a novel probabilistic model consisting of an observation and a novel temporal filtering module.Operating internally with an efficient ray-based representation, the observation module consists of a single and a multiview module to predict horizontal depth from images and fuses their results to benefit from advantages offered by either methodology.Our method operates on conventional consumer hardware and overcomes a common limitation of competing methods that often demand upright images.Our full system meets real-time requirements, while outperforming the state-of-the-art by a significant margin.

We present the first application of 3D Gaussian Splatting in monocular SLAM, the most fundamental but the hardest setup for Visual SLAM. Our method, which runs live at 3fps, utilises Gaussians as the only 3D representation, unifying the required representation for accurate, efficient tracking, mapping, and high-quality rendering. Designed for challenging monocular settings, our approach is seamlessly extendable to RGB-D SLAM when an external depth sensor is available. Several innovations are required to continuously reconstruct 3D scenes with high fidelity from a live camera. First, to move beyond the original 3DGS algorithm, which requires accurate poses from an offline Structure from Motion (SfM) system, we formulate camera tracking for 3DGS using direct optimisation against the 3D Gaussians, and show that this enables fast and robust tracking with a wide basin of convergence. Second, by utilising the explicit nature of the Gaussians, we introduce geometric verification and regularisation to handle the ambiguities occurring in incremental 3D dense reconstruction. Finally, we introduce a full SLAM system which not only achieves state-of-the-art results in novel view synthesis and trajectory estimation but also reconstruction of tiny and even transparent objects.

Learning generalizable visual representations from Internet data has yielded promising results for robotics.Yet, prevailing approaches focus on pre-training 2D representations, being sub-optimal to deal with occlusions and accurately localize objects in complex 3D scenes. Meanwhile, 3D representation learning has been limited to single-object understanding.To address these limitations, we introduce a novel 3D pre-training framework for robotics named SUGAR that captures semantic, geometric and affordance properties of objects through 3D point clouds.We underscore the importance of cluttered scenes in 3D representation learning, and automatically construct a multi-object dataset benefiting from cost-free supervision in simulation.SUGAR employs a versatile transformer-based model to jointly address five pre-training tasks, namely cross-modal knowledge distillation for semantic learning, masked point modeling to understand geometry structures, grasping pose synthesis for object affordance, 3D instance segmentation and referring expression grounding to analyze cluttered scenes.We evaluate our learned representation on three robotic-related tasks, namely, zero-shot 3D object recognition, referring expression grounding, and language-driven robotic manipulation.Experimental results show that SUGAR's 3D representation outperforms state-of-the-art 2D and 3D representations.

Robot manipulation relies on accurately predicting contact points and end-effector directions to ensure successful operation. However, learning-based robot manipulation, trained on a limited category within a simulator, often struggles to achieve generalizability, especially when confronted with extensive categories.Therefore, we introduce an innovative approach for robot manipulation that leverages the robust reasoning capabilities of Multimodal Large Language Models (MLLMs) to enhance the stability and generalization of manipulation. By fine-tuning the injected adapters, we preserve the inherent common sense and reasoning ability of the MLLMs while equipping them with the ability for manipulation. The fundamental insight lies in the introduced fine-tuning paradigm, encompassing object category understanding, affordance prior reasoning, and object-centric pose prediction to stimulate the reasoning ability of MLLM in manipulation. During inference, our approach utilizes an RGB image and text prompt to predict the end effector's pose in chain of thoughts. After the initial contact is established, an active impedance adaptation policy is introduced to plan the upcoming waypoints in a closed-loop manner. Moreover, in real world, we design a test-time adaptation (TTA) strategy for manipulation to enable the model better adapt to the current real-world scene configuration. Experiments in simulator and real-world show the promising performance of ManipLLM.

We introduce the new setting of open-vocabulary object 6D pose estimation, in which a textual prompt is used to specify the object of interest.In contrast to existing approaches, in our setting(i) the object of interest is specified solely through the textual prompt,(ii) no object model (e.g. CAD or video sequence) is required at inference,(iii) the object is imaged from two different viewpoints of two different scenes, and(iv) the object was not observed during the training phase.To operate in this setting, we introduce a novel approach that leverages a Vision-Language Model to segment the object of interest from two distinct scenes and to estimate its relative 6D pose.The key of our approach is a carefully devised strategy to fuse object-level information provided by the prompt with local image features, resulting in a feature space that can generalize to novel concepts.We validate our approach on a new benchmark based on two popular datasets, REAL275 and Toyota-Light, which collectively encompass 39 object instances appearing in four thousand image pairs. The results demonstrate that our approach outperforms both a well-established hand-crafted method and a recent deep learning-based baseline in estimating the relative 6D pose of objects in different scenes.

This paper introduces Hierarchical Diffusion Policy (HDP), a hierarchical agent for multi-task robotic manipulation. HDP factorises a manipulation policy into a hierarchical structure: a high-level task-planning agent which predicts a distant next-best end-effector pose (NBP), and a low-level goal-conditioned diffusion policy which generates optimal motion trajectories. The factorised policy representation allows HDP to tackle both long-horizon task planning while generating fine-grained low-level actions. To generate context-aware motion trajectories while satisfying robot kinematics constraints, we present a novel kinematics-aware goal-conditioned control agent, Robot Kinematics Diffuser (RK-Diffuser). Specifically, RK-Diffuser learns to generate both the end-effector pose and joint position trajectories, and distill the accurate but kinematics-unaware end-effector pose diffuser to the kinematics-aware but less accurate joint position diffuser via differentiable kinematics. Empirically, we show that HDP achieves a significantly higher success rate than the state-of-the-art methods in both simulation and real-world.

Despite the great need for assistive technology among vulnerable groups (e.g., the elderly, children, and the disabled) in daily tasks, research into advanced AI-driven assistive solutions still remains sparse. Traditional multi-agent interaction tasks often require machines to simply help without nuanced consideration of human feelings (e.g., self-esteem, sense of control, opportunity for practice and learning, sense of self-improvement, etc.), while in real life there are always distinct types of human users with different capabilities, goals, etc. Addressing this mismatch, we define a pivotal and novel challenge ``Smart Help'', i.e., providing both proactive and adaptive support to human agents with diverse disabilities and dynamic goals in different tasks and environments. Leveraging AI2-THOR, an interactive 3D environment tailored for real-world home simulations, we introduce an innovative opponent modeling module, focusing on a nuanced understanding of the main agent's capabilities and goals, and optimize the assisting agent's helping policy. Rigorous experiments validate the efficacy of our model components and show the superiority of our holistic approach against established baselines. Our findings illustrate the potential of AI-imbued assistive robots for enhancing the well-being of vulnerable groups. Our environment, dataset, and codes will be released upon acceptance.

We focus on the generalization ability of the 6-DoF grasp detection method in this paper. While learning-based grasp detection methods can predict grasp poses for unseen objects using the grasp distribution learned from the training set, they often exhibit a significant performance drop when encountering objects with diverse shapes and structures. To enhance the grasp detection methods' generalization ability, we incorporate domain prior knowledge of robotic grasping, enabling better adaptation to objects with significant shape and structure differences. More specifically, we employ the physical constraint regularization during the training phase to guide the model towards predicting grasps that comply with the physical rule on grasping. For the unstable grasp poses predicted on novel objects, we design a contact-score joint optimization using the projection contact map to refine these poses in cluttered scenarios. Extensive experiments conducted on the GraspNet-1billion benchmark demonstrate a substantial performance gain on the novel object set and the real-world grasping experiments also demonstrate the effectiveness of our generalizing 6-DoF grasp detection method.

Event cameras exhibit remarkable attributes such as high dynamic range, asynchronicity, and low latency, making them highly suitable for vision tasks that involve high-speed motion in challenging lighting conditions. These cameras implicitly capture movement and depth information in events, making them appealing sensors for Camera Pose Relocalization (CPR) tasks. Nevertheless, existing CPR networks based on events neglect the pivotal fine-grained temporal information in events, resulting in unsatisfactory performance. Moreover, the energy-efficient features are further compromised by the use of excessively complex models, hindering efficient deployment on edge devices. In this paper, we introduce PEPNet, a simple and effective point-based network designed to regress six degrees of freedom (6-DOFs) event camera poses. We rethink the relationship between the event camera and CPR tasks, leveraging the raw point cloud directly as network input to harness the high-temporal resolution and inherent sparsity of events. PEPNet is adept at abstracting the spatial and implicit temporal features through hierarchical structure and explicit temporal features by Attentive Bi-directional Long Short-Term Memory (A-Bi-LSTM). By employing a carefully crafted lightweight design, PEPNet delivers state-of-the-art (SOTA) performance on public datasets with meager computational resources. Specifically, PEPNet attains a significant 38\% performance improvement on the random split DAVIS 240C CPR Dataset, utilizing merely 6\% of the parameters compared to traditional frame-based approaches. Moreover, the lightweight design version PEPNet$_{tiny}$ accomplishes results comparable to the SOTA while employing a mere 0.5\% of the parameters.

This paper presents Neural Visibility Field (NVF), a novel uncertainty quantification method for Neural Radiance Fields (NeRF) applied to active mapping. Our key insight is that regions not visible in the training views lead to inherently unreliable color predictions by NeRF at this region, resulting in increased uncertainty in the synthesized views. To address this, we propose to use Bayesian Networks to composite position-based field uncertainty into ray-based uncertainty in camera observations. Consequently, NVF naturally assigns higher uncertainty to unobserved regions, aiding robots to select the most informative next viewpoints. Extensive evaluations show that NVF excels not only in uncertainty quantification but also in scene reconstruction for active mapping, outperforming existing methods. More details can be found at https://sites.google.com/view/nvf-cvpr24/ .

While there has been remarkable progress recently in the fields of manipulation and locomotion, mobile manipulation remains a long-standing challenge. Compared to locomotion or static manipulation, a mobile system must make a diverse range of long-horizon tasks feasible in unstructured and dynamic environments. While the applications are broad and interesting, there are a plethora of challenges in developing these systems such as coordination between the base and arm, reliance on onboard perception for perceiving and interacting with the environment, and most importantly, simultaneously integrating all these parts together. Prior works approach the problem using disentangled modular skills for mobility and manipulation that are trivially tied together. This causes several limitations such as compounding errors, delays in decision-making, and no whole-body coordination. In this work, we present a reactive mobile manipulation framework that uses an active visual system to consciously perceive and react to its environment. Similar to how humans leverage whole-body and hand-eye coordination, we develop a mobile manipulator that exploits its ability to move and see, more specifically -- to move in order to see and to see in order to move. This allows it to not only move around and interact with its environment but also, choose "when" to perceive "what" using an active visual system. We observe that such an agent learns to navigate around complex cluttered scenarios while displaying agile whole-body coordination using only ego-vision without needing to create environment maps. Videos are available at https://spin-robot.github.io

Existing 3D scene understanding methods are heavily focused on 3D semantic and instance segmentation. However, identifying objects and their parts only constitutes an intermediate step towards a more fine-grained goal, which is effectively interacting with the functional interactive elements (e.g., handles, knobs, buttons) in the scene to accomplish diverse tasks. To this end, we introduce SceneFun3D, a large-scale dataset with more than 14.8k highly accurate interaction annotations for 710 high-resolution real-world 3D indoor scenes. We accompany the annotations with motion parameter information, describing how to interact with these elements, and a diverse set of natural language descriptions of tasks that involve manipulating them in the scene context. To showcase the value of our dataset, we introduce three novel tasks, namely functionality segmentation, task-driven affordance grounding and 3D motion estimation, and adapt existing state-of-the-art methods to tackle them. Our experiments show that solving these tasks in real 3D scenes remains challenging despite recent progress in closed-set and open-set 3D scene understanding methods.

Predictive learning models, which aim to predict future frames based on past observations, are crucial to constructing world models. These models need to maintain low-level consistency and capture high-level dynamics in unannotated spatiotemporal data. Transitioning from frame-wise to token-wise prediction presents a viable strategy for addressing these needs. How to improve token representation and optimize token decoding presents significant challenges. This paper introduces PredToken, a novel predictive framework that addresses these issues by decoupling space-time tokens into distinct components for iterative cascaded decoding. Concretely, we first design a "decomposition, quantization, and reconstruction" schema based on VQGAN to improve the token representation. This scheme disentangles low- and high-frequency representations and employs a dimension-aware quantization model, allowing more low-level details to be preserved. Building on this, we present a "coarse-to-fine iterative decoding" method. It leverages dynamic soft decoding to refine coarse tokens and static soft decoding for fine tokens, enabling more high-level dynamics to be captured. These designs make PredToken produce high-quality predictions. Extensive experiments demonstrate the superiority of our method on various real-world spatiotemporal predictive benchmarks. Furthermore, PredToken can also be extended to other visual generative tasks to yield realistic outcomes.

Diverse actions give rise to rich audio-visual signals in long videos. Recent works showcase that the two modalities of audio and video exhibit different temporal extents of events and distinct labels. We address the interplay between the two modalities in long videos by explicitly modelling the temporal extents of audio and visual events. We propose the Time Interval Machine (TIM) where a modality-specific time interval poses as a query to a transformer encoder that ingests a long video input. The encoder then attends to the specified interval, as well as the surrounding context in both modalities, in order to recognise the ongoing action. We test TIM on three long audio-visual video datasets: EPIC-KITCHENS, Perception Test, and AVE, reporting state-of-the-art (SOTA) for recognition. On EPIC-KITCHENS, we beat previous SOTA that utilises LLMs and significantly larger pre-training by 2.9% top-1 action recognition accuracy. Additionally, we show that TIM can be adapted for action detection, using dense multi-scale interval queries, outperforming SOTA on EPIC-KITCHENS-100 for most metrics, and showing strong performance on the Perception Test. Our ablations show the critical role of integrating the two modalities and modelling their time intervals in achieving this performance. Code and models at: https://github.com/JacobChalk/TIM.

Generating Audio Description (AD) for movies is a challenging task that requires fine-grained visual understanding and an awareness of the characters and their names. Currently, visual language models for AD generation are limited by a lack of suitable training data, and also their evaluation is hampered by using performance measures not specialized to the AD domain. In this paper, we make three contributions: (i) We propose two approaches for constructing AD datasets with aligned video data, and build training and evaluation datasets using these. These datasets will be publicly released; (ii) We develop a Q-former-based architecture which ingests raw video and generates AD, using frozen pre-trained visual encoders and large language models; and (iii) We provide new evaluation metrics to benchmark AD quality that are well matched to human performance. Taken together, we improve the state of the art on AD generation.

We study supervised action segmentation, whose goal is to predict framewise action labels of a video. To capture temporal dependencies over long horizons, prior works either improve framewise features with transformer or refine framewise predictions with learned action features. However, they ignore that frame and action features contain complimentary information, which can be leveraged to enhance both features and improve temporal modeling. Therefore, we propose an efficient Frame-Action Cross-attention Temporal modeling (FACT) framework that performs temporal modeling with frame and action features in parallel and leverage this parallelism to achieve iterative bidirectional information transfer between the features and refine them. FACT network contains (i) a frame branch to learn frame-level information with convolutions and frame features, (ii) an action branch to learn action-level dependencies with transformers and action tokens and (iii) cross-attentions to allow communication between the two branches. We also propose a new matching loss to ensure each action token uniquely encodes an action segment, thus better captures its semantics. Thanks to our architecture, we can also leverage textual transcripts of videos to help action segmentation. We evaluate FACT on four video datasets (two egocentric and two third-person) for action segmentation with and without transcripts, showing that FACT significantly improves the state-of-the-art accuracy while enjoys lower computational cost (3 times faster) than existing transformer-based methods.

We address the problem of online action segmentation for egocentric procedural task videos. While previous studies have mostly focused on offline action segmentation, where entire videos are available for both training and inference, the transition to online action segmentation is crucial for practical applications like AR/VR task assistants. Notably, applying an offline-trained model directly to online inference results in a significant performance drop due to the inconsistency between training and inference. We propose an online action segmentation framework by first modifying existing architectures to make them causal. Second, we develop a novel action progress prediction module to dynamically estimate the progress of ongoing actions and using them to refine the predictions of causal action segmentation. Third, we propose to learn task graphs from training videos and leverage them to obtain smooth and procedure-consistent segmentations. With the combination of progress and task graph with casual action segmentation, our framework effectively addresses prediction uncertainty and oversegmentation in online action segmentation and achieves significant improvement on three egocentric datasets.

Most video captioning models are designed to process short video clips of few seconds and output text describing low-level visual concepts (e.g., objects, scenes, atomic actions). However, most real-world videos last for minutes or hours and have a complex hierarchical structure spanning different temporal granularities. We propose Video ReCap, a recursive video captioning model that can process video inputs of dramatically different lengths (from 1 second to 2 hours) and output video captions at multiple hierarchy levels. The recursive video-language architecture exploits the synergy between different video hierarchies and can process hour-long videos efficiently. We utilize a curriculum learning training scheme to learn the hierarchical structure of videos, starting from clip-level captions describing atomic actions, then focusing on segment-level descriptions, and concluding with generating summaries for hour-long videos. Furthermore, we introduce Ego4D-HCap dataset by augmenting Ego4D with 8,267 manually collected long-range video summaries. Our recursive model can flexibly generate captions at different hierarchy levels while also being useful for other complex video understanding tasks, such as VideoQA on EgoSchema. Data, code, and models are publicly available at https://sites.google.com/view/vidrecap.

The core of video understanding tasks, such as recognition, captioning, and tracking, is to automatically detect objects or actions in a video and analyze their temporal evolution. Despite sharing a common goal, different tasks often rely on distinct model architectures and annotation formats. In contrast, natural language processing benefits from a unified output space, \textit{i.e.}, text sequences, which simplifies the training of powerful foundational language models, such as GPT-3, with extensive training corpora. Inspired by this, we seek to unify the output space of video understanding tasks by using languages as labels and additionally introducing \textit{time} and \textit{box} tokens. In this way, a variety of video tasks could be formulated as video-grounded token generation. This enables us to address various types of video tasks, including classification (such as action recognition), captioning (covering clip captioning, video question answering, and dense video captioning), and localization tasks (such as visual object tracking) within a fully shared encoder-decoder architecture, following a generative framework. Through comprehensive experiments, we demonstrate such a simple and straightforward idea is quite effective and can achieve state-of-the-art or competitive results on seven video benchmarks, providing a novel perspective for more universal video understanding.

Recently, integrating video foundation models and large language models to build a video understanding system can overcome the limitations of specific pre-defined vision tasks. Yet, existing systems can only handle videos with very few frames. For long videos, the computation complexity, memory cost, and long-term temporal connection impose additional challenges. Taking advantage of the Atkinson-Shiffrin memory model, with tokens in Transformers being employed as the carriers of memory in combination with our specially designed memory mechanism, we propose the MovieChat to overcome these challenges. MovieChat achieves state-of-the-art performance in long video understanding, along with the released MovieChat-1K benchmark with 1K long video and 14K manual annotations for validation of the effectiveness of our method.

This paper proposes Group Activity Feature (GAF) learning in which features of multi-person activity are learned as a compact latent vector. Unlike prior work in which the manual annotation of group activities is required for supervised learning, our method learns the GAF through person attribute prediction without group activity annotations. By learning the whole network in an end-to-end manner so that the GAF is required for predicting the person attributes of people in a group, the GAF is trained as the features of multi-person activity. As a person at- tribute, we propose to use a person’s action class and appearance features because the former is easy to annotate due to its simpleness, and the latter requires no manual annotation. In addition, we introduce a location-guided attribute prediction to disentangle the complex GAF for extracting the features of each target person properly. Various experimental results validate that our method outperforms SOTA methods quantitatively and qualitatively on two public datasets. Visualization of our GAF also demonstrates that our method learns the GAF representing fined-grained group activity classes. Code: https://github.com/chihina/GAFL-CVPR2024.

An ideal model for dense video captioning -- predicting captions localized temporally in a video -- should be able to handle long input videos, predict rich, detailed textual descriptions, and be able to produce outputs before processing the entire video. Current state-of-the-art models, however, process a fixed number of downsampled frames, and make a single full prediction after seeing the whole video. We propose a streaming dense video captioning model that consists of two novel components: First, we propose a new memory module, based on clustering incoming tokens, which can handle arbitrarily long videos as the memory is of a fixed size. Second, we develop a streaming decoding algorithm that enables our model to make predictions before the entire video has been processed. Our model achieves this streaming ability, and significantly improves the state-of-the-art on three dense video captioning benchmarks: ActivityNet, YouCook2 and ViTT. Our code is released at https://github.com/google-research/scenic.

Weakly-supervised action segmentation is a task of learning to partition a long video into several action segments, where training videos are only accompanied by transcripts (ordered list of actions). Most of existing methods need to infer pseudo segmentation for training by serial alignment between all frames and the transcript, which is time-consuming and hard to be parallelized while training. In this work, we aim to escape from this inefficient alignment with massive but redundant frames, and instead to directly localize a few action transitions for pseudo segmentation generation, where a transition refers to the change from an action segment to its next adjacent one in the transcript. As the true transitions are submerged in noisy boundaries due to intra-segment visual variation, we propose a novel Action-Transition-Aware Boundary Alignment (ATBA) framework to efficiently and effectively filter out noisy boundaries and detect transitions. In addition, to boost the semantic learning in the case that noise is inevitably present in the pseudo segmentation, we also introduce video-level losses to utilize the trusted video-level supervision. Extensive experiments show the effectiveness of our approach on both performance and training speed.

Temporal action detection (TAD) aims to locate action positions and recognize action categories in long-term untrimmed videos. Although many methods have achieved promising results, their robustness has not been thoroughly studied. In practice, we observe that temporal information in videos can be occasionally corrupted, such as missing or blurred frames. Interestingly, existing methods often incur a significant performance drop even if only one frame is affected. To formally evaluate the robustness, we establish two temporal corruption robustness benchmarks, namely THUMOS14-C and ActivityNet-v1.3-C. In this paper, we extensively analyze the robustness of seven leading TAD methods and obtain some interesting findings: 1) Existing methods are particularly vulnerable to temporal corruptions, and end-to-end methods are often more susceptible than those with a pre-trained feature extractor; 2) Vulnerability mainly comes from localization error rather than classification error; 3) When corruptions occur in the middle of an action instance, TAD models tend to yield the largest performance drop. Besides building a benchmark, we further develop a simple but effective robust training method to defend against temporal corruptions, through the FrameDrop augmentation and Temporal-Robust Consistency loss. Remarkably, our approach not only improves robustness but also yields promising improvements on clean data. We believe that this study will serve as a benchmark for future research in robust video analysis. Source code and models are available at https://github.com/Alvin-Zeng/temporal-robustness-benchmark.

Human comprehension of a video stream is naturally broad: in a few instants, we are able to understand what is happening, the relevance and relationship of objects, and forecast what will follow in the near future, everything all at once. We believe that - to effectively transfer such an holistic perception to intelligent machines - an important role is played by learning to correlate concepts and to abstract knowledge coming from different tasks, to synergistically exploit them when learning novel skills. To accomplish this, we seek for a unified approach to video understanding which combines shared temporal modelling of human actions with minimal overhead, to support multiple downstream tasks and enable cooperation when learning novel skills. We then propose EgoPack, a solution that creates a collection of task perspectives that can be carried across downstream tasks and used as a potential source of additional insights, as a backpack of skills that a robot can carry around and use when needed. We demonstrate the effectiveness and efficiency of our approach on four Ego4D benchmarks, outperforming current state-of-the-art methods. The code will be released upon acceptance.

We study object interaction anticipation in egocentric videos. This task requires an understanding of the spatio-temporal context formed by past actions on objects, coined "action context". We propose TransFusion, a multimodal transformer-based architecture for short-term object interaction anticipation. Our method exploits the representational power of language by summarizing the action context textually, after leveraging pre-trained vision-language foundation models to extract the action context from past video frames. The summarized action context and the last observed video frame are processed by the multimodal fusion module to forecast the next object interaction. Experiments on the Ego4D next active object interaction dataset show the effectiveness of our multimodal fusion model and highlight the benefits of using the power of foundation models and language-based context summaries in a task where vision may appear to suffice. Our novel approach outperforms all state-of-the-art methods on both versions of the Ego4D dataset.

Video anomaly detection (VAD) with weak supervision has achieved remarkable performance in utilizing video-level labels to discriminate whether a video frame is normal or abnormal. However, current approaches are inherently limited to a closed-set setting and may struggle in open-world applications where there can be anomaly categories in the test data unseen during training. A few recent studies attempt to tackle a more realistic setting, open-set VAD, which aims to detect unseen anomalies given seen anomalies and normal videos. However, such a setting focuses on predicting frame anomaly scores, having no ability to recognize the specific categories of anomalies, despite the fact that this ability is essential for building more informed video surveillance systems. This paper takes a step further and explores open-vocabulary video anomaly detection (OVVAD), in which we aim to leverage pre-trained large models to detect and categorize seen and unseen anomalies. To this end, we propose a model that decouples OVVAD into two mutually complementary tasks -- class-agnostic detection and class-specific classification -- and jointly optimizes both tasks. Particularly, we devise a semantic knowledge injection module to introduce semantic knowledge from large language models for the detection task, and design a novel anomaly synthesis module to generate pseudo unseen anomaly videos with the help of large vision generation models for the classification task. These semantic knowledge and synthesis anomalies substantially extend our model's capability in detecting and categorizing a variety of seen and unseen anomalies. Extensive experiments on three widely-used benchmarks demonstrate our model achieves state-of-the-art performance on OVVAD task.

Video moment retrieval and highlight detection are two highly valuable tasks in video understanding, but until recently they have been jointly studied. Although existing studies have made impressive advancement recently, they predominantly follow the data-driven bottom-up paradigm. Such paradigm overlooks task-specific and inter-task effects, resulting in poor model performance. In this paper, we propose a novel task-driven top-down framework TaskWeave for joint moment retrieval and highlight detection. The framework introduces a task-decoupled unit to capture task-specific and common representations. To investigate the interplay between the two tasks, we propose an inter-task feedback mechanism, which transforms the results of one task as guiding masks to assist the other task. Different from existing methods, we present a task-dependent joint loss function to optimize the model. Comprehensive experiments and in-depth ablation studies on QVHighlights, TVSum, and Charades-STA datasets corroborate the effectiveness and flexibility of the proposed framework. Codes are available at https://github.com/EdenGabriel/TaskWeave.

Weakly-supervised Video Anomaly Detection (wVAD) aims to detect frame-level anomalies using only video-level labels in training. Due to the limitation of weak-label, Multi-Instance Learning (MIL) is prevailing in wVAD which leverages binary labels. However, MIL suffers from insufficiency of binary supervision to model diverse abnormal patterns. Besides, the coupling between abnormality and its context hinders learning of clear abnormal event boundary. In this paper, we propose prompt-enhanced multi-instance learning, which is devised to detect various abnormal events while ensuring a clear event boundary. Concretely, we first design the abnormal-aware prompts by using abnormal class annotations together with trainable prompts, which can incorporate semantic priors into video features dynamically. The detector can utilize the semantic-rich features to capture diverse abnormal patterns. In addition, normal context prompt is introduced to amplify the distinction between abnormality and abnormal context, facilitating the generation of clear boundary. With the mutual enhancement of abnormal-aware and normal context prompts, the model can construct a discriminative representation to detect divergent anomalies without ambiguous event boundary. Extensive experiments demonstrate our method achieves SOTA performance on three public benchmarks.

Spatio-temporal video grounding (or STVG) task aims at locating a spatio-temporal tube for a specific instance given a text query. Despite advancements, current methods easily suffer the distractors or heavy object appearance variations in videos due to insufficient object information from the text, leading to degradation. Addressing this, we propose a novel framework, context-guided STVG (CG-STVG), which mines discriminative instance context for object in videos and applies it as a supplementary guidance for target localization. The key of CG-STVG lies in two specially designed modules, including instance context generation (ICG), which focuses on discovering visual context information (in both appearance and motion) of the instance, and instance context refinement (ICR), which aims to improve the instance context from ICG by eliminating irrelevant or even harmful information from the context. During grounding, ICG, together with ICR, are deployed at each decoding stage of a Transformer architecture for instance context learning. Particularly, instance context learned from one decoding stage is fed to the next stage, and leveraged as a guidance containing rich and discriminative object feature to enhance the target-awareness in decoding feature, which conversely benefits generating better new instance context for improving localization finally. Compared to existing methods, CG-STVG enjoys object information in text query and guidance from mined instance visual context for more accurate target localization. In our experiments on three benchmarks, including HCSTVG-v1/-v2 and VidSTG, CG-STVG sets new state-of-the-arts in mtIoU and mvIoU on all of them, showing efficacy. Code is released at https://github.com/HengLan/CGSTVG.

Video transformers have become the de facto standard for human action recognition, yet their exclusive reliance on the RGB modality still limits their adoption in certain domains. One such domain is Activities of Daily Living (ADL), where RGB alone is not sufficient to distinguish between visually similar actions, or actions observed from multiple viewpoints. To facilitate the adoption of video transformers for ADL, we hypothesize that the augmentation of RGB with human pose information, known for its sensitivity to fine-grained motion and multiple viewpoints, is essential. Consequently, we introduce the first Pose Induced Video Transformer: PI-ViT (or $\pi$-ViT), a novel approach that augments the RGB representations learned by video transformers with 2D and 3D pose information. The key elements of $\pi$-ViT are two plug-in modules, 2D Skeleton Induction Module and 3D Skeleton Induction Module, that are responsible for inducing 2D and 3D pose information into the RGB representations. These modules operate by performing pose-aware auxiliary tasks, a design choice that allows $\pi$-ViT to discard the modules during inference. Notably, $\pi$-ViT achieves the state-of-the-art performance on three prominent ADL datasets, encompassing both real-world and large-scale RGB-D datasets, without requiring poses or additional computational overhead at inference.

Action detection aims to localize the starting and ending points of action instances in untrimmed videos, and predict the classes of those instances. In this paper, we make the observation that the outputs of the action detection task can be formulated as images. Thus, from a novel perspective, we tackle action detection via a three-image generation process to generate starting point, ending point and action-class predictions as images via our proposed Action Detection Image Diffusion (ADI-Diff) framework. Furthermore, since our images differ from natural images and exhibit special properties, we further explore a Discrete Action-Detection Diffusion Process and a Row-Column Transformer design to better handle our images. Our ADI-Diff framework achieves state-of-the-art performance on two widely-used datasets.

Sign Language Translation (SLT) is a challenging task that aims to translate sign videos into spoken language. Inspired by the strong translation capabilities of large language models (LLMs) that are trained on extensive web-scale multilingual text corpora, we aim to harness off-the-shelf LLMs to handle SLT. In this paper, we regularize the sign videos to embody linguistic characteristics of spoken language, and propose a novel SignLLM framework to transform sign videos into a language-like representation for improved readability by off-the-shelf LLMs. SignLLM comprises two key modules: (1) The Vector-Quantized Visual Sign module converts sign videos into a sequence of discrete character-level sign tokens, and (2) the Codebook Reconstruction and Alignment module converts these character-level tokens into word-level sign representations using an optimal transport formulation. A sign-text alignment loss further bridges the gap between sign and text tokens, enhancing semantic compatibility. We achieve state-of-the-art results on two widely-used SLT benchmarks.

The most performant spatio-temporal action localisation models use external person proposals and complex external memory banks. We propose a fully end-to-end, transformer based model that directly ingests an input video, and outputs tubelets -- a sequence of bounding boxes and the action classes at each frame. Our flexible model can be trained with either sparse bounding-box supervision on individual frames, or full tubelet annotations. And in both cases, it predicts coherent tubelets as the output. Moreover, our end-to-end model requires no additional pre-processing in the form of proposals, or post-processing in terms of non-maximal suppression. We perform extensive ablation experiments, and significantly advance the state-of-the-art on five different spatio-temporal action localisation benchmarks with both sparse keyframes and full tubelet annotations.

Visual interactivity understanding within visual scenes presents a significant challenge in computer vision. Existing methods focus on complex interactivities while leveraging a simple relationship model. These methods, however, struggle with a diversity of appearance, situation, position, interaction, and relation in videos. This limitation hinders the ability to fully comprehend the interplay within the complex visual dynamics of subjects. In this paper, we delve into interactivities understanding within visual content by deriving scene graph representations from dense interactivities among humans and objects. To achieve this goal, we first present a new dataset containing Appearance-Situation-Position-Interaction-Relation predicates, named ASPIRe, offering an extensive collection of videos marked by a wide range of interactivities. Then, we propose a new approach named Hierarchical Interlacement Graph (HIG), which leverages a unified layer and graph within a hierarchical structure to provide deep insights into scene changes across five distinct tasks. Our approach demonstrates superior performance to other methods through extensive experiments conducted in various scenarios.

Skeleton-based action recognition has attracted lots of research attention. Recently, to build an accurate skeleton-based action recognizer, a variety of works have been proposed. Among them, some works use large model architectures as backbones of their recognizers to boost the skeleton data representation capability, while some other works pre-train their recognizers on external data to enrich the knowledge. In this work, we observe that large language models which have been extensively used in various natural language processing tasks generally hold both large model architectures and rich implicit knowledge. Motivated by this, we propose a novel LLM-AR framework, in which we investigate treating the Large Language Model as an Action Recognizer. In our framework, we propose a linguistic projection process to project each input action signal (i.e., each skeleton sequence) into its ''sentence format'' (i.e., an ''action sentence''). Moreover, we also incorporate our framework with several designs to further facilitate this linguistic projection process. Extensive experiments demonstrate the efficacy of our proposed framework.

Recent progress in large multimodal models (LMMs) has demonstrated exceptional potentials in visual comprehension to be a general-purpose assistant. However, existing LMM cannot easily be adapted to provide a frame-aligned, concise and timely answer for an online continuous incoming video stream. In this paper, we present LIVE, a novel Learning-In-Video-strEam framework, that can enable LMMs to address this challenge by carefully considering the training sequence format, dataset creation and inference optimization. First, we propose a novel streaming video dialogue format that encourages the model to produce frame-aligned responses for any incoming query. Second, we propose an improved autoregressive training objective that learns to predict the concise answer respect to key event frame and remains silent for redundant frames between key frames. To further speed up the inference, we propose a key-value strategy with only key-frame context. Compared to LMMs trained in conventional framework, we demonstrate that LIVE can provide more frame-aligned concise answer at high accuracy and can support real-time synchronized decoding for an online video stream in inference. We demonstrate LIVE can tackle general long-video understanding tasks with capabilities in captioning and forecasting. LIVE shows superior performance in a variety tasks such as temporal alignment, anticipation, and summarization on COIN datasets and Ego4D benchmarks.

Spatio-temporal grounding describes the task of localizing events in space and time, e.g., in video data, based on verbal descriptions only. Models for this task are usually trained with human-annotated sentences and bounding box supervision. This work addresses this task from a multimodal supervision perspective, proposing a framework for spatio-temporal action grounding trained on loose video and subtitle supervision only, without human annotation. To this end, we combine local representation learning, which focuses on leveraging fine-grained spatial information, with a global representation encoding that captures higher-level representations and incorporates both in a joint approach. To evaluate this challenging task in a real-life setting, a new benchmark dataset is proposed providing dense spatio-temporal grounding annotations in long, untrimmed, multi-action instructional videos for over 5K events. We evaluate the proposed approach and other methods on the proposed and standard downstream tasks showing that our method improves over current baselines in various settings, including spatial, temporal, and untrimmed multi-action spatio-temporal grounding.

In this paper, we investigate a new problem called narrative action evaluation (NAE). NAE aims to generate professional commentary that evaluates the execution of an action. Unlike traditional tasks such as score-based action quality assessment and video captioning involving superficial sentences, NAE focuses on creating detailed narratives in natural language. These narratives provide intricate descriptions of actions along with objective evaluations. NAE is a more challenging task because it requires both narrative flexibility and evaluation rigor. One existing possible solution is to use multi-task learning, where narrative language and evaluative information are predicted separately. However, this approach results in reduced performance for individual tasks because of variations between tasks and differences in modality between language information and evaluation information. To address this, we propose a prompt-guided multimodal interaction framework. This framework utilizes a pair of transformers to facilitate the interaction between different modalities of information. It also uses prompts to transform the score regression task into a video-text matching task, thus enabling task interactivity. To support further research in this field, we re-annotate the MTL-AQA and FineGym datasets with high-quality and comprehensive action narration. Additionally, we establish benchmarks for NAE. Extensive experiment results prove that our method outperforms separate learning methods and naive multi-task learning methods. Data and code are released at here.

Point-level weakly-supervised temporal action localization (P-TAL) aims to localize action instances in untrimmed videos through the use of single-point annotations in each instance. Existing methods predict the class activation sequences without any boundary information, and the unreliable sequences result in a significant misalignment between the quality of proposals and their corresponding confidence. In this paper, we surprisingly observe the most salient frame tend to appear in the central region of the each instance and is easily annotated by humans. Guided by the temporal saliency information, we present a novel proposal-level plug-in framework to relearn the aligned confidence of proposals generated by the base locators. The proposed approach consists of Center Score Learning (CSL) and Alignment-based Boundary Adaptation (ABA). In CSL, we design a novel center label generated by the point annotations for predicting aligned center scores. During inference, we first fuse the center scores with the predicted action probabilities to obtain the aligned confidence. ABA utilizes the both aligned confidence and IoU information to enhance localization completeness. Extensive experiments demonstrate the generalization and effectiveness of the proposed framework, showcasing state-of-the-art or competitive performances across three benchmarks. Our code is available at https://github.com/zyxia1009/CVPR2024-TSPNet.

In this paper, we study multi-label atomic activity recognition. Despite the notable progress in action recognition algorithms, they struggle to recognize atomic activities due to a deficiency in a holistic understanding of both multiple road users’ motions and their contextual information. We introduce Action-slot, a slot attention-based approach that learns visual action-centric representations, capturing both motion and contextual information. Our key idea is to design action slots that are capable of paying attention to regions where atomic activities occur, without the need for explicit perception guidance. To further enhance slot attention, we introduce a background slot that competes with action slots, aiding the training process in avoiding unnecessary focus on background regions devoid of activities. Yet, the imbalanced class distribution in the existing OATS dataset hampers the assessment of rare activities. To address the limitation, we collect a synthetic dataset called TACO, which is four times larger than OATS and features a balanced distribution of atomic activities. To validate the effectiveness of our method, we conduct comprehensive experiments and ablation studies against various action recognition baselines on OATS and TACO. We also show that the performance of multi-label atomic activity recognition on real-world datasets can be improved by pre-training representations on TACO. We will release our code and datasets upon acceptance.

Active Speaker Detection (ASD) aims to identify who is speaking in each frame of a video. Solving ASD involves using audio and visual information in two complementary contexts: long-term intra-speaker context models the temporal dependencies of the same speaker, while short-term inter-speaker context models the interactions of speakers in the same scene. Motivated by these observations, we propose LoCoNet, a simple but effective Long-Short Context Network that leverages Long-term Intra-speaker Modeling (LIM) and Short-term Inter-speaker Modeling (SIM) in an interleaved manner. LIM employs self-attention for long-range temporal dependencies modeling and cross-attention for audio-visual interactions modeling. SIM incorporates convolutional blocks that capture local patterns for short-term inter-speaker context. Experiments show that LoCoNet achieves state-of-the-art performance on multiple datasets, with 95.2%(+0.3%) mAP on AVA-ActiveSpeaker, 97.2%(+2.7%) mAP on Talkies and 68.4%(+7.7%) mAP on Ego4D. Moreover, in challenging cases where multiple speakers are present, LoCoNet outperforms previous state-of-the-art methods by 3.0% mAP on AVA-ActiveSpeaker. The code will be released.

Video scene detection aims to temporally link shots for obtaining semantically compact scenes.It is essential for this task to capture scene-distinguishable affinity among shots by similarity assessment.However, most methods relies on ordinary shot-to-shot similarities, which may inveigle similar shots into being linked even though they are from different scenes, and meanwhile hinder dissimilar shots from being blended into a complete scene.In this paper, we propose NeighborNet to inject shot contexts into shot-to-shot similarities through carefully exploring the relations between semantic/temporal neighbors of shots over a local time period.In this way, shot-to-shot similarities are remeasured as semantic/temporal neighbor-aware similarities so that NeighborNet can learn context embedding into shot features using graph convolutional network.As a result, not only do the learned shot features suppress the affinity among similar shots from different scenes, but they also promote the affinity among dissimilar shots in the same scene.Experimental results on public benchmark datasets show that our proposed NeighborNet yields substantial improvements in video scene detection, especially outperforms released state-of-the-arts by at least 6\% in Average Precision (AP).The code is available at https://github.com/ExMorgan-Alter/NeighborNet.

Promptly identifying procedural errors from egocentric videos in an online setting is highly challenging and valuable for detecting mistakes as soon as they happen. This capability has a wide range of applications across various fields, such as manufacturing and healthcare. The nature of procedural mistakes is open-set since novel types of failures might occur, which calls for one-class classifiers trained on correctly executed procedures. However, no technique can currently detect open-set procedural mistakes online.We propose PREGO, the first online one-class classification model for mistake detection in procedural egocentric videos. PREGO is based on an online action recognition component to model the current action, and a symbolic reasoning module to predict the next actions Mistake prediction is performed by comparing the recognized current action with the expected future one. We evaluate PREGO on two procedural egocentric video datasets, ASSEMBLY101 and Epic-tent, which we rearrange for online benchmarking of procedural mistake detection to establish suitable benchmarks, thus defining the ASSEMBLY101-O and Epic-tent-O datasets, respectively.

Object State Changes (OSCs) are pivotal for video understanding. While humans can effortlessly generalize OSC understanding from familiar to unknown objects, current approaches are confined to a closed vocabulary. Addressing this gap, we introduce a novel open-world formulation for the video OSC problem. The goal is to temporally localize the three stages of an OSC---the object's initial state, its transitioning state, and its end state---whether or not the object has been observed during training. Towards this end, we develop VidOSC, a holistic learning approach that: (1) leverages text and vision-language models for supervisory signals to obviate manually labeling OSC training data, and (2) abstracts fine-grained shared state representations from objects to enhance generalization. Furthermore, we present HowToChange, the first open-world benchmark for video OSC localization, which offers an order of magnitude increase in the label space and annotation volume compared to the best existing benchmark. Experimental results demonstrate the efficacy of our approach, in both traditional closed-world and open-world scenarios.

Extending large image-text pre-trained models (e.g., CLIP) for video understanding has made significant advancements. To enable the capability of CLIP to perceive dynamic information in videos, existing works are dedicated to equipping the visual encoder with various temporal modules. However, these methods exhibit "asymmetry" between the visual and textual sides, with neither temporal descriptions in input texts nor temporal module in text encoder. This limitation hinders the potential of language supervision emphasized in CLIP, and restricts the learning of temporal features, as the text encoder has demonstrated limited proficiency in motion understanding. To address this issue, we propose leveraging "MoTion-Enhanced Descriptions" (MoTED) to facilitate the extraction of distinctive temporal features in videos. Specifically, we first generate discriminative motion-related descriptions via querying GPT-4 to compare easy-confusing action categories. Then, we incorporate both the visual and textual encoders with additional perception modules to process the video frames and generated descriptions, respectively. Finally, we adopt a contrastive loss to align the visual and textual motion features. Extensive experiments on five benchmarks show that MoTED surpasses state-of-the-art methods with convincing gaps, laying a solid foundation for empowering CLIP with strong temporal modeling.

Self-supervised foundation models have shown great potential in computer vision thanks to the pre-training paradigm of masked autoencoding. Scale is a primary factor influencing the performance of these foundation models. However, these large foundation models often result in high computational cost that might limit their deployment. This paper focuses on pre-training relatively small vision transformer models that could be efficiently adapted to downstream tasks. Specifically, taking inspiration from knowledge distillation in model compression, we propose a new asymmetric masked distillation(AMD) framework for pre-training relatively small models with autoencoding. The core of AMD is to devise an asymmetric masking strategy, where the teacher model is enabled to see more context information with a lower masking ratio, while the student model still with high masking ratio to the original masked pre-training. We design customized multi-layer feature alignment between the teacher encoder and student encoder to regularize the pre-training of student MAE. To demonstrate the effectiveness and versatility of AMD, we apply it to both ImageMAE and VideoMAE for pre-training relatively small ViT models. AMD achieved 84.6% classification accuracy on IN1K using the ViT-B model. And AMD achieves 73.3% classification accuracy using the ViT-B model on the Something-in-Something V2 dataset, a 3.7% improvement over the original ViT-B model from VideoMAE. We also transfer AMD pre-trained models to downstream tasks and obtain consistent performance improvement over the standard pre-training.

Video anomaly detection (VAD) aims to temporally locate abnormal events in a video. Existing works mostly rely on training deep models to learn the distribution of normality with either video-level supervision, one-class supervision, or in an unsupervised setting. Training-based methods are prone to be domain-specific, thus being costly for practical deployment as any domain change will involve data collection and model training. In this paper, we radically depart from previous efforts and propose LAnguage-based VAD (LAVAD), a method tackling VAD in a novel, training-free paradigm, exploiting the capabilities of pre-trained large language models (LLMs) and existing vision-language models (VLMs). We leverage VLM-based captioning models to generate textual descriptions for each frame of any test video. With the textual scene description, we then devise a prompting mechanism to unlock the capability of LLMs in terms of temporal aggregation and anomaly score estimation, turning LLMs into an effective video anomaly detector. We further leverage modality-aligned VLMs and propose effective techniques based on cross-modal similarity for cleaning noisy captions and refining the LLM-based anomaly scores. We evaluate LAVAD on two large datasets featuring real-world surveillance scenarios (UCF-Crime and XD-Violence), showing that it outperforms both unsupervised and one-class methods without requiring any training or data collection.

Video visual relation detection tasks, such as video scene graph generation, play an important role in fine-grained video understanding. However, current video visual relation detection datasets have two main limitations that hinder the progress of research in this area. Firstly, they do not explore complex human-human interactions in multi-person scenarios. Secondly, the relation types they define have a relatively low semantic level and can often be recognized by appearance or prior information, without the need for detailed spatio-temporal context reasoning. Nevertheless, comprehending high-level interactions between humans is crucial for understanding complex multi-person videos, such as sports and surveillance videos. To address this gap, we propose a new video visual relation detection task: video human-human interaction detection, and introduce a novel dataset named SportsHHI. SportsHHI contains 34 high-level interaction classes from basketball and volleyball sports. 118075 human bounding boxes and 50649 interaction instances are annotated on 11398 keyframes. To benchmark this, we propose a two-stage baseline method and conduct extensive experiments to reveal the key factors for a successful human-human interaction detector. We hope that SportsHHI can stimulate research on human interaction understanding in videos and promote the development of spatio-temporal context modeling techniques in video visual relation detection.

Vision-Language models (VLMs) have excelled in the image-domain--- especially in zero-shot settings--- thanks to the availability of vast pretraining data (i.e., paired image-text samples). However for videos, such paired data is not as abundant. Therefore, video-VLMs are usually designed by adapting pretrained image-VLMs to the video-domain, instead of training from scratch. All such recipes rely on augmenting visual embeddings with temporal information (i.e., image $\rightarrow$ video), often keeping text embeddings unchanged or even being discarded. In this paper, we argue the contrary, that better video-VLMs can be designed by focusing more on augmenting text, rather than visual information. More specifically, we introduce Video-conditioned Text Representations (VicTR): a form of text embeddings optimized w.r.t. visual embeddings, creating a more-flexible contrastive latent space. Our model can further make use of freely-available semantic information, in the form of visually-grounded auxiliary text (e.g. object or scene information). We evaluate our model on few-shot, zero-shot (HMDB-51, UCF-101), short-form (Kinetics-400) and long-form (Charades) activity recognition benchmarks, showing strong performance among video-VLMs.

Temporal Action Detection (TAD) aims to identify the action boundaries and the corresponding category within untrimmed videos. Inspired by the success of DETR in object detection, previous methods adapted the query-based framework to TAD tasks. However, these approaches primarily followed DETR to predict actions at the instance level (i.e., identify each action by its center point), leading to sub-optimal boundary localization. To address this issue, we propose a novel Dual-level query-based TAD framework, namely DualDETR, to detect actions from both instance-level and boundary-level. Decoding at different levels requires semantics of different granularity, therefore we introduce a two-branch decoding structure. This structure creates distinctive decoding processes for different levels, facilitating explicit capture of temporal cues and semantics at each level. On top of the two-branch design, we present a joint query initialization strategy to align queries from both levels. Specifically, we leverage encoder proposals to match queries from each level in a one-to-one manner. Then, the matched queries are initialized using position and content prior from the matched action proposal. The aligned dual-level queries can refine the matched proposal with complementary efforts during subsequent decoding. We evaluate DualDETR on three challenging multi-label TAD benchmarks. The experimental results demonstrate the superior performance of DualDETR compared to the existing state-of-the-art methods, achieving a substantial improvement under det-mAP and delivering impressive results under seg-mAP.

Vision transformer (ViT) has shown high potential in video recognition, owing to its straightforward design, adaptable self-attention mechanisms, and the efficacy of masked pre-training. Yet, it still remains unclear how to adapt these pre-trained short-term ViTs for temporal action detection (TAD) in untrimmed videos. The existing works treat them as off-the-shelf feature extractors for each extremely short trimmed snippet without capturing the fine-grained relation among different snippets in a broader temporal context.To mitigate this issue, this paper focuses on designing a new mechanism for adapting these pre-trained ViT models as a unified long-form video transformer, to fully unleash its modeling power in capturing inter-snippet relation while still keeping low computation overhead and memory consumption for efficient TAD. To this end, we designed effective cross-snippet propagation modules to gradually exchange short-term video information among different snippets from two levels. For inner-backbone information propagation, we introduced a cross-snippet propagation strategy to enable multi-snippet temporal feature interaction inside the backbone.For post-backbone information propagation, we proposed temporal transformer layers for further clip-level modeling.With the plain ViT-B pre-trained with VideoMAE, our end-to-end temporal action detector (ViT-TAD) yields a very competitive performance to previous temporal action detectors, riching up to 69.0 average mAP on THUMOS14, 37.12 average mAP on ActivityNet-1.3 and 17.20 average mAP on FineAction.

We introduce PlausiVL, a large video-language model for anticipating action sequences that are plausible in the real-world. While significant efforts have been made towards anticipating future actions, prior approaches do not take into account the aspect of plausibility in an action sequence. To address this limitation, we explore the generative capability of a large video-language model in our work and further, develop the understanding of plausibility in an action sequence by introducing two objective functions, a counterfactual-based plausible action sequence learning loss and a long-horizon action repetition loss. We utilize temporal logical constraints as well as verb-noun action pair logical constraints to create implausible/counterfactual action sequences and use them to train the model with plausible action sequence learning loss. This loss helps the model to differentiate between plausible and not plausible action sequences and also helps the model to learn implicit temporal cues crucial for the task of action anticipation. The long-horizon action repetition loss puts a higher penalty on the actions that are more prone to repetition over a longer temporal window. With this penalization, the model is able to generate diverse, plausible action sequences. We evaluate our approach on two large-scale datasets, Ego4D and EPIC-Kitchens-100 and show improvements on the task of action anticipation.

Recently, temporal action detection (TAD) has seen significant performance improvement with end-to-end training. However, due to the memory bottleneck, only models with limited scales and limited data volumes can afford end-to-end training, which inevitably restricts TAD performance. In this paper, we reduce the memory consumption for end-to-end training, and manage to scale up the TAD backbone to 1 billion parameters and the input video to 1,536 frames, leading to significant detection performance.The key to our approach lies in our proposed temporal-informative adapter (TIA), which is a novel lightweight module that reduces training memory. Using TIA, we free the humongous backbone from learning to adapt to the TAD task by only updating the parameters in TIA. TIA also leads to better TAD representation by temporally aggregating context from adjacent frames throughout the backbone.We evaluate our model across four representative datasets. Owing to our efficient design, we are able to train end-to-end on VideoMAEv2-giant and achieve 75.4\% mAP on THUMOS14, being the first end-to-end model to outperform the best feature-based methods.

With recent video object segmentation (VOS) benchmarks evolving to challenging scenarios, we revisit a simple yet previously underestimated strategy: restricting the size of memory banks. This diverges from the prevalent practice of expanding memory banks to accommodate extensive historical information. Our specially designed "memory deciphering" study offers a pivotal insight underpinning such a strategy: expanding memory banks, while seemingly beneficial, actually increases the difficulty for VOS modules to decode relevant features due to the confusion from redundant information. By restricting memory banks to a limited number of essential frames, we achieve a notable improvement in VOS accuracy. This process balances the importance and freshness of frames to maintain an informative memory bank within a bounded capacity. Additionally, restricted memory banks reduce the training-inference discrepancy in memory lengths compared with continuous expansion. This fosters new opportunities in temporal reasoning and enables us to introduce the previously overlooked "temporal positional embedding." Finally, our insights are embodied in "RMem" ("R" for restricted), a simple yet effective VOS modification that excels at challenging VOS scenarios and establishes new state of the art for object state changes (VOST dataset) and long videos (the Long Videos dataset). Our codes are available at https://github.com/Restricted-Memory/RMem and our demo can be watched on https://youtu.be/V3tCFQsJrrM.

Researchers in natural science need reliable methods for quantifying animal behavior. Recently, numerous computer vision methods emerged to automate the process. However, observing wild species at remote locations remains a challenging task due to difficult lighting conditions and constraints on power supply and data storage. Event cameras offer unique advantages for battery-dependent remote monitoring due to their low power consumption and high dynamic range capabilities. We use this novel sensor to quantify a behavior in Chinstrap penguins called ecstatic display. We formulate the problem as a temporal action detection task, determining the start and end times of the behavior. For this purpose, we recorded a colony of breeding penguins in Antarctica during several weeks and labeled event data on 16 nests. The developed method consists of a generator of candidate time intervals (proposals) and a classifier of the actions within them. The experiments show that the event cameras’ natural response to motion is effective for continuous behavior monitoring and detection, reaching a mean average precision (mAP) of 58% (which increases to 63% in good weather conditions). The results also demonstrate the robustness against various lighting conditions contained in the challenging dataset. The low-power capabilities of the event camera allows to record three times longer than with a conventional camera. We make our code and “Event Penguins” dataset available. This work pioneers the use of event cameras for remote wildlife observation, opening new interdisciplinary opportunities.

We present Egocentric Action Scene Graphs (EASGs), a new representation for long-form understanding of egocentric videos. EASGs extend standard manually-annotated representations of egocentric videos, such as verb-noun action labels, by providing a temporally evolving graph-based description of the actions performed by the camera wearer, including interacted objects, their relationships, and how actions unfold in time. Through a novel annotation procedure, we extend the Ego4D dataset by adding manually labeled Egocentric Action Scene Graphs offering a rich set of annotations designed for long-from egocentric video understanding. We hence define the EASG generation task and provide a baseline approach, establishing preliminary benchmarks. Experiments on two downstream tasks, egocentric action anticipation and egocentric activity summarization, highlight the effectiveness of EASGs for long-form egocentric video understanding. We will release the dataset and the code to replicate experiments and annotations.

Event cameras have recently been shown beneficial for practical vision tasks, such as action recognition, thanks to their high temporal resolution, power efficiency, and reduced privacy concerns. However, current research is hindered by 1) the difficulty in processing events because of their prolonged duration and dynamic actions with complex and ambiguous semantics; and 2) the redundant action depiction of the event frame representation with fixed stacks. We find language naturally conveys abundant semantic information, rendering it stunningly superior in reducing semantic uncertainty. In light of this, we propose ExACT, a novel approach that, for the first time, tackles event-based action recognition from a cross-modal conceptualizing perspective. Our ExACT brings two technical contributions. Firstly, we propose an adaptive fine-grained event (AFE) representation to adaptively filter out the repeated events for the stationary objects while preserving dynamic ones. This subtly enhances the performance of ExACT without extra computational cost. Then, we propose a conceptual reasoning-based uncertainty estimation module, which simulates the recognition process to enrich the semantic representation. In particular, conceptual reasoning builds the temporal relation based on the action semantics, and uncertainty estimation tackles the semantic uncertainty of actions based on the distributional representation. Experiments show that our ExACT achieves superior recognition accuracy of 94.83%(+2.23%), 90.10%(+37.47%) and 67.24% on PAF, HARDVS and our SeAct datasets respectively.

Human action anticipation aims at predicting what people will do in the future based on past observations. In this paper, we introduce Uncertainty-aware Action Decoupling Transformer (UADT) for action anticipation. Unlike existing methods that directly predict action in a verb-noun pair format, we decouple the action anticipation task into verb and noun anticipations separately. The objective is to make the two decoupled tasks assist each other and eventually improve the action anticipation task. Specifically, we propose a two-stream Transformer-based architecture which is composed of a verb-to-noun model and a noun-to-verb model. The verb-to-noun model leverages the verb information to improve the noun prediction and the other way around. We extend the model in a probabilistic manner and quantify the predictive uncertainty of each decoupled task to select features. In this way, the noun prediction leverages the most informative and redundancy-free verb features and verb prediction works similarly. Finally, the two streams are combined dynamically based on their uncertainties to make the joint action anticipation. We demonstrate the efficacy of our method by achieving state-of-the-art performance on action anticipation benchmarks including EPIC-KITCHENS, EGTEA Gaze+, and 50 Salads.

We present a new egocentric procedural error dataset containing videos with various types of errors as well as normal videos and propose a new framework for procedural error detection using error-free training videos only. Our framework consists of an action segmentation model and a contrastive step prototype learning module to segment actions and learn useful features for error detection. Based on the observation that interactions between hands and objects often inform action and error understanding, we propose to combine holistic frame features with relations features, which we learn by building a graph using active object detection followed by a Graph Convolutional Network. To handle errors, unseen during training, we use our contrastive step prototype learning to learn multiple prototypes for each step, capturing variations of error-free step executions. At inference time, we use feature-prototype similarities for error detection. By experiments on three datasets, we show that our proposed framework outperforms state-of-the-art video anomaly detection methods for error detection and provides smooth action and error predictions.

In this paper, we tackle the problem of self-supervised video alignment and activity progress prediction using in-the-wild videos. Our proposed self-supervised representation learning method carefully addresses different action orderings, redundant actions, and background frames to generate improved video representations compared to previous methods. Our model generalizes temporal cycle-consistency learning to allow for more flexibility in determining cycle-consistent neighbors. More specifically, to handle repeated actions, we propose a multi-neighbor cycle consistency and a multi-cycle-back regression loss by finding multiple soft nearest neighbors using a Gaussian Mixture Model. To handle background and redundant frames, we introduce a context-dependent drop function in our framework, discouraging the alignment of droppable frames. On the other hand, to learn from videos of multiple activities jointly, we propose a multi-head crosstask network, allowing us to embed a video and estimate progress without knowing its activity label. Experiments on multiple datasets show that our method outperforms the state-of-the-art for video alignment and progress prediction.

Current transformer-based skeletal action recognition models tend to focus on a limited set of joints and low-level motion patterns to predict action classes. This results in significant performance degradation under small skeleton perturbations or changing the pose estimator between training and testing. In this work, we introduce MaskCLR, a new Masked Contrastive Learning approach for Robust skeletal action recognition. We propose an Attention-Guided Probabilistic Masking (AGPM) strategy to occlude the most important joints and encourage the model to explore a larger set of discriminative joints. Furthermore, we propose a Multi-Level Contrastive Learning (MLCL) paradigm to enforce the representations of standard and occluded skeletons to be class-discriminative, i.e, more compact within each class and more dispersed across different classes. Our approach helps the model capture the high-level action semantics instead of low-level joint variations, and can be conveniently incorporated into transformer-based models. Without loss of generality, we combine MaskCLR with three transformer backbones: the vanilla transformer, DSTFormer, and STTFormer. Extensive experiments on NTU60, NTU120, and Kinetics400 show that MaskCLR consistently outperforms previous state-of-the-art methods on standard and perturbed skeletons from different pose estimators, showing improved accuracy, generalization, and robustness to skeleton perturbations. Codes will be available upon publication..

Large-scale visual-language pre-trained models have achieved significant success in various video tasks. However, most existing methods follow an "adapt then align" paradigm, which adapts pre-trained image encoders to model video-level representations and utilizes one-hot or text embedding of the action labels for supervision. This paradigm overlooks the challenge of mapping from static images to complicated activity concepts. In this paper, we propose a novel "Align before Adapt" (ALT) paradigm. Prior to adapting to video representation learning, we exploit the entity-to-region alignments for each frame. The alignments are fulfilled by matching the region-aware image embeddings to an offline-constructed text corpus. With the aligned entities, we feed their text embeddings to a transformer-based video adapter as the queries, which can help extract the semantics of the most important entities from a video to a vector. This paradigm reuses the visual-language alignment of VLP during adaptation and tries to explain an action by the underlying entities. This helps understand actions by bridging the gap with complex activity semantics, particularly when facing unfamiliar or unseen categories. ALT demonstrates competitive performance while maintaining remarkably low computational costs. In fully supervised experiments, it achieves 88.1% top-1 accuracy on Kinetics-400 with only 4947 GFLOPs. Moreover, ALT outperforms the previous state-of-the-art methods in both zero-shot and few-shot experiments, emphasizing its superior generalizability across various learning scenarios.

We present Dive Into the BoundrieS (DIBS), a novel pretraining framework for dense video captioning (DVC), that elaborates on improving the quality of the generated event captions and their associated pseudo event boundaries from unlabeled videos. By leveraging the capabilities of diverse large language models (LLMs), we generate rich DVC-oriented caption candidates and optimize the corresponding pseudo boundaries jointly under several meticulously designed objectives, considering diversity, event-centricity, temporal ordering, and coherence. Moreover, we further introduce a novel online boundary refinement strategy that iteratively improves the quality of pseudo boundaries during training. Comprehensive experiments have been conducted to examine the effectiveness of the proposed technique components. By leveraging a substantial amount of unlabeled video data, such as HowTo100M, we achieve a remarkable advancement on standard DVC datasets like YouCook2 and ActivityNet. We outperform the previous state-of-the-art Vid2Seq across a majority of metrics, achieving this with just 0.4\% of the unlabeled video data used for pre-training by Vid2Seq.

Video Moment Retrieval (MR) and Highlight Detection (HD) have attracted significant attention due to the growing demand for video analysis. Recent approaches treat MR and HD as similar video grounding problems and address them together with transformer-based architecture. However, we observe that the emphasis of MR and HD differs, with one necessitating the perception of local relationships and the other prioritizing the understanding of global contexts. Consequently, the lack of task-specific design will inevitably lead to limitations in associating the intrinsic specialty of two tasks. To tackle the issue, we propose a $\textbf{U}$nified $\textbf{V}$ideo $\textbf{COM}$prehension framework (UVCOM) to bridge the gap and jointly solve MR and HD effectively. By performing progressive integration on intra and inter-modality across multi-granularity, UVCOM achieves the comprehensive understanding in processing a video. Moreover, we present multi-aspect contrastive learning to consolidate the local relation modeling and global knowledge accumulation via well aligned multi-modal space. Extensive experiments on QVHighlights, Charades-STA, TACoS , YouTube Highlights and TVSum datasets demonstrate the effectiveness and rationality of UVCOM which outperforms the state-of-the-art methods by a remarkable margin.

Zero-Shot Temporal Action Localization (ZS-TAL) seeks to identify and locate actions in untrimmed videos unseen during training. Existing ZS-TAL methods involve fine-tuning a model on a large amount of annotated training data. While effective, training-based ZS-TAL approaches assume the availability of labeled data for supervised learning, which can be impractical in some applications. Furthermore, the training process naturally induces a domain bias into the learned model, which may adversely affect the model's generalization ability to arbitrary videos. These considerations prompt us to approach the ZS-TAL problem from a radically novel perspective, relaxing the requirement for training data. To this aim, we introduce a novel method that performs $\textbf{T}$est-$\textbf{T}$ime adaptation for $\textbf{T}$emporal $\textbf{A}$ction $\textbf{L}$ocalization ($\textbf{T3AL}$). In a nutshell, $T3AL$ adapts a pre-trained Vision and Language Model (VLM) at inference time on a sample basis. $T3AL$ operates in three steps. First, a video-level pseudo-label of the action category is computed by aggregating information from the entire video. Then, action localization is performed adopting a novel procedure inspired by self-supervised learning. Finally, frame-level textual descriptions extracted with a state-of-the-art captioning model are employed for refining the action region proposals. We validate the effectiveness of $T3AL$ by conducting experiments on the THUMOS14 and the ActivityNet-v1.3 datasets. Our results demonstrate that $T3AL$ significantly outperforms zero-shot baselines based on state-of-the-art VLMs, confirming the benefit of a test-time adaptation approach.

Concerns for the privacy of individuals captured in public imagery have led to privacy-preserving action recognition. Existing approaches often suffer from issues arising through obfuscation being applied globally and a lack of interpretability. Global obfuscation hides privacy sensitive regions, but also contextual regions important for action recognition. Lack of interpretability erodes trust in these new technologies. We highlight the limitations of current paradigms and propose a solution: Human selected privacy templates that yield interpretability by design, an obfuscation scheme that selectively hides attributes and also induces temporal consistency, which is important in action recognition. Our approach is architecture agnostic and directly modifies input imagery, while existing approaches generally require architecture training. Our approach offers more flexibility, as no retraining is required, and outperforms alternatives on three widely used datasets.

Comparing a user video to a reference how-to video is a key requirement for AR/VR technology delivering personalized assistance tailored to the user's progress. However, current approaches for language-based assistance can only answer questions about a single video. We propose an approach that first automatically generates large amounts of visual instruction tuning data involving pairs of videos from HowTo100M by leveraging existing step annotations and accompanying narrations, and then trains a video-conditioned language model to jointly reason across multiple raw videos. Our model achieves state-of-the-art performance at identifying differences between video pairs and ranking videos based on the severity of these differences, and shows promising ability to perform general reasoning over multiple videos.

In this paper, we suggest a new novel method to understand complex semantic structures through long video inputs.Conventional methods for understanding videos have been focused on short-term clips, and trained to get visual representations for the short clips using convolutional neural networks or transformer architectures.However, most real-world videos are composed of long videos ranging from minutes to hours, therefore, it essentially brings limitations to understanding the overall semantic structures of the long videos by dividing them into small clips and learning the representations of them.To tackle this problem, we suggested a new algorithm to learn representations of videos at the object level, defining spatiotemporal high-order relationships among the representations as semantic units.The proposed method includes a new transformer architecture capable of learning spatiotemporal graphs, and a compositional learning method to learn disentangled features for each semantic unit.Using the suggested method, we resolve the challenging video task, which is compositional generalization understanding of unseen videos.In experiments using CATER and newly suggested compositional generalization datasets, we demonstrate new state-of-the-art performance.

We propose a novel approach to the action segmentation task for long, untrimmed videos, based on solving an optimal transport problem. By encoding a temporal consistency prior into a Gromov-Wasserstein problem, we are able to decode a temporally consistent segmentation from a noisy affinity/matching cost matrix between video frames and action classes. Unlike previous approaches, our method does not require knowing the action order for a video to attain temporal consistency. Furthermore, our resulting (fused) Gromov-Wasserstein problem can be efficiently solved on GPUs using a few iterations of projected mirror descent. We demonstrate the effectiveness of our method in an unsupervised learning setting, where our method is used to generate pseudo-labels for self-training. We evaluate our segmentation approach and unsupervised learning pipeline on the Breakfast, 50-Salads, YouTube Instructions and Desktop Assembly datasets, yielding state-of-the-art results for the unsupervised video action segmentation task.

Existing action quality assessment (AQA) methods mainly learn deep representations at the video level for scoring diverse actions. Due to the lack of a fine-grained understanding of actions in videos, they harshly suffer from low credibility and interpretability, thus insufficient for stringent applications, such as Olympic diving events. We argue that a fine-grained understanding of actions requires the model to perceive and parse actions in both time and space, which is also the key to the credibility and interpretability of the AQA technique. Based on this insight, we propose a new fine-grained spatial-temporal action parser named FineParser. It learns human-centric foreground action representations by focusing on target action regions within each frame and exploiting their fine-grained alignments in time and space to minimize the impact of invalid backgrounds during the assessment. In addition, we construct fine-grained annotations of human-centric foreground action masks for the FineDiving dataset, called FineDiving-HM. With refined annotations on diverse target action procedures, FineDiving-HM can promote the development of real-world AQA systems. Through extensive experiments, we demonstrate the effectiveness of FineParser, which outperforms state-of-the-art methods while supporting more tasks of fine-grained action understanding. Data and code are available at https://github.com/PKU-ICST-MIPL/FineParser_CVPR2024.

Skeleton-based action recognition is gaining popularity as a viable alternative to RGB-based classification due to its lower computational complexity and compact data structure. While remarkable progress has been made on supervised skeleton-based action recognition, the challenge of zero-shot action recognition remains relatively unexplored. In this paper, we argue that relying solely on aligning label-level semantics and global features from all skeleton joints is insufficient to effectively transfer locally consistent visual knowledge from seen to unseen classes. To address this limitation, we introduce Part-aware Unified Representation between Language and Skeleton (PURLS) to explore visual-semantic alignment at both local and global scales. PURLS incorporates a prompting module and a partitioning module to generate textual and visual representations for visual-text alignment. The prompting module leverages GPT-3 to generate refined global/local descriptions from original action labels and extracts their language embeddings using CLIP. The partitioning module employs an adaptive sampling strategy to group visual features of semantic-relevant body joints for a given description. During training, PURLS is trained to project aligned visual-textual encoding manifolds of global, body-part-based local, and temporal-interval-based local representations in a balanced manner. Our approach is evaluated on three large-scale datasets, i.e., NTU-RGB+D 60, NTU-RGB+D 120, and a newly curated dataset Kinetics-skeleton 200. The results showcase the superior performance of PURLS, surpassing prior skeleton-based solutions and standard baselines from other domains.

Video Transformers have become the prevalent solution for various video downstream tasks with superior expressive power and flexibility. However, these video transformers suffer from heavy computational costs induced by the massive number of tokens across the entire video frames, which has been the major barrier to training the model. Further, the patches irrelevant to the main contents, e.g., backgrounds, degrade the generalization performance of models. To tackle these issues, we propose training free token merging for lightweight video Transformer (vid-TLDR) that aims to enhance the efficiency of video Transformers by merging the background tokens without additional training. For vidTLDR, we introduce a novel approach to capture the salient regions in videos only with the attention map. Further, we introduce the saliency-aware token merging strategy by dropping the background tokens and sharpening the object scores. Our experiments show that vid-TLDR significantly mitigates the computational complexity of video Transformers while achieving competitive performance compared to the base model without vid-TLDR.

Fine-grained medical action analysis plays a vital role in improving medical skill training efficiency, but it faces the problems of data and algorithm shortage. Cardiopulmonary Resuscitation (CPR) is an essential skill in emergency treatment. Currently, the assessment of CPR skills mainly depends on dummies and trainers, leading to high training costs and low efficiency. For the first time, this paper constructs a vision-based system to complete error action recognition and skill assessment in CPR. Specifically, we define 13 types of single-error actions and 74 types of composite error actions during external cardiac compression and then develop a video dataset named CPR-Coach. By taking the CPR-Coach as a benchmark, this paper investigates and compares the performance of existing action recognition models based on different data modalities. To solve the unavoidable Single-class Training & Multi-class Testing problem, we propose a human-cognition-inspired framework named ImagineNet to improve the model's multi-error recognition performance under restricted supervision. Extensive comparison and actual deployment experiments verify the effectiveness of the framework. We hope this work could bring new inspiration to the computer vision and medical skills training communities simultaneously. The dataset and the code are publicly available on https://github.com/Shunli-Wang/CPR-Coach.

Video anomaly understanding (VAU) aims to automatically comprehend unusual occurrences in videos, thereby enabling various applications such as traffic surveillance and industrial manufacturing. While existing VAU benchmarks primarily concentrate on anomaly detection and localization, our focus is on more practicality, prompting us to raise the following crucial questions: “what anomaly occurred?”, “why did it happen?”, and “how severe is this abnormal event?”. In pursuit of these answers, we present a comprehensive benchmark for Causation Understanding of Video Anomaly (CUVA). Specifically, each instance of the proposed benchmark involves three sets of human annotations to indicate the “what”, “why” and “how” of an anomaly, including 1) anomaly type, start and end times, and event descriptions, 2) natural language explanations for the cause of an anomaly, and 3) free text reflecting the effect of the abnormality. In addition, we also introduce MMEval, a novel evaluation metric designed to better align with human preferences for CUVA, facilitating the measurement of existing LLMs in comprehending the underlying cause and corresponding effect of video anomalies. Finally, we propose a novel prompt-based method that can serve as a baseline approach for the challenging CUVA. We conduct extensive experiments to show the superiority of our evaluation metric and the prompt-based approach.

We introduce the video detours problem for navigating instructional videos. Given a source video and a natural language query asking to alter the how-to video's current path of execution in a certain way, the goal is to find a related "detour video" that satisfies the requested alteration. To address this challenge, we propose VidDetours, a novel video-language approach that learns to retrieve the targeted temporal segments from a large repository of how-to's using video-and-text conditioned queries. Furthermore, we devise a language-based pipeline that exploits how-to video narration text to create weakly supervised training data. We demonstrate our idea applied to the domain of how-to cooking videos, where a user can detour from their current recipe to find steps with alternate ingredients, tools, and techniques. Validating on a ground truth annotated dataset of 16K samples, we show our model's significant improvements over best available methods for video retrieval and question answering, with recall rates exceeding the state of the art by 35%.

In this paper, we explore the capability of an agent to construct a logical sequence of action steps, thereby assembling a strategic procedural plan. This plan is crucial for navigating from an initial visual observation to a target visual outcome, as depicted in real-life instructional videos. Existing works have attained partial success by extensively leveraging various sources of information available in the datasets, such as heavy intermediate visual observations, procedural names, or natural language step-by-step instructions, for features or supervision signals. However, the task remains formidable due to the implicit causal constraints in the sequencing of steps and the variability inherent in multiple feasible plans. To tackle these intricacies that previous efforts have overlooked, we propose to enhance the agent’s capabilities by infusing it with procedural knowledge. This knowledge, sourced from training procedure plans and structured as a directed weighted graph, equips the agent to better navigate the complexities of step sequencing and its potential variations. We coin our approach KEPP, a novel Knowledge-Enhanced Procedure Planning system, which harnesses a probabilistic procedural knowledge graph extracted from training data, effectively acting as a comprehensive textbook for the training domain. Experimental evaluations across three widely-used datasets under settings of varying complexity reveal that KEPP attains superior, state-of-the-art results while requiring only minimal supervision. Code and trained model are available at https://github.com/Ravindu-Yasas-Nagasinghe/KEPP

Action Localization is a challenging problem that combines detection and recognition tasks, which are often addressed separately. State-of-the-art methods rely on off-the-shelf bounding box detections pre-computed at high resolution, and propose transformer models that focus on the classification task alone. Such two-stage solutions are prohibitive for real-time deployment. On the other hand, single-stage methods target both tasks by devoting part of the network (generally the backbone) to sharing the majority of the workload, compromising performance for speed. These methods build on adding a DETR head with learnable queries that after cross- and self-attention can be sent to corresponding MLPs for detecting a person’s bounding box and action. However, DETR-like architectures are challenging to train and can incur in big complexity.In this paper, we observe that a straight bipartite matching loss can be applied to the output tokens of a vision transformer. This results in a backbone + MLP architecture that can do both tasks without the need of an extra encoder-decoder head and learnable queries. We show that a single MViTv2-S architecture trained with bipartite matching to perform both tasks surpasses the same MViTv2-S when trained with RoI align on pre-computed bounding boxes. With a careful design of token pooling and the proposed training pipeline, our Bipartite-Matching Vision Transformer model, BMViT, achieves +3 mAP on AVA2.2. w.r.t. the two-stage MViTv2-S counterpart. Code is available athttps://github.com/IoannaNti/BMViT

In this paper, we investigate that the normalized coordinate expression is a key factor as reliance on hand-crafted components in query-based detectors for temporal action detection (TAD).Despite significant advancements towards an end-to-end framework in object detection, query-based detectors have been limited in achieving full end-to-end modeling in TAD.To address this issue, we propose TE-TAD, a full end-to-end action detection transformer that integrates time-aligned coordinate expression.We reformulate coordinate expression utilizing actual timeline values, ensuring length-invariant representations from the extremely diverse video duration environment.Furthermore, our proposed adaptive query selection dynamically adjusts the number of queries based on video length, providing a suitable solution for varying video durations compared to a fixed query set.Our approach not only simplifies the TAD process by eliminating the need for hand-crafted components but also significantly improves the performance of query-based detectors.Our TE-TAD outperforms the previous query-based detectors and achieves competitive performance compared to state-of-the-art methods on popular benchmark datasets.

Video summarization aims to generate a concise representation of a video, capturing its essential content and key moments while reducing its overall length. Although several methods employ attention mechanisms to handle long-term dependencies, they often fail to capture the visual significance inherent in frames. To address this limitation, we propose a CNN-based SpatioTemporal Attention (CSTA) method that stacks each feature of frames from a single video to form image-like frame representations and applies 2D CNN to these frame features. Our methodology relies on CNN to comprehend the inter and intra-frame relations and to find crucial attributes in videos by exploiting its ability to learn absolute positions within images. In contrast to previous work compromising efficiency by designing additional modules to focus on spatial importance, CSTA requires minimal computational overhead as it uses CNN as a sliding window. Extensive experiments on two benchmark datasets (SumMe and TVSum) demonstrate that our proposed approach achieves state-of-the-art performance with fewer MACs compared to previous methods. Codes are available at https://github.com/thswodnjs3/CSTA.

Recent progress in Vision-Language Foundation (VLF) models has revealed the great advantages of cross-modality learning. However, due to a large gap between vision and text, they might not be able to sufficiently utilize the benefits of cross-modality information. In the field of human action recognition, the additional pose modality may bridge the gap between vision and text to improve the effectiveness of cross-modality learning. In this paper, we propose a novel framework, called Pose-enhanced Vision-Language (PeVL) model, to adapt the VL model with pose modality to learn effective knowledge of fine-grained human actions. Our PeVL model includes two novel components: an Unsymmetrical Cross-Modality Refinement (UCMR) block and a Semantic-Guided Multi-level Contrastive (SGMC) module. The UCMR block includes Pose-guided Visual Refinement (P2V-R) and Visual-enriched Pose Refinement (V2P-R) for effective cross-modality learning. The SGMC module includes Multi-level Contrastive Associations of vision-text and pose-text at both action and sub-action levels, and a Semantic-Guided Loss, enabling effective contrastive learning among three modalities (i.e., video, pose, and text). Built upon a pre-trained VL foundation model, our model integrates trainable adapters and can be trained end-to-end. Our novel PeVL design over VL foundation model yields remarkable performance gains on four fine-grained human action recognition datasets, achieving new SOTA with a significantly small number of tunable parameters for low-cost re-training.

We propose a novel approach to video anomaly detection: we treat feature vectors extracted from videos as realizations of a random variable with a fixed distribution and model this distribution with a neural network. This lets us estimate the likelihood of test videos and detect video anomalies by thresholding the likelihood estimates. We train our video anomaly detector using a modification of denoising score matching, a method that injects training data with noise to facilitate modeling its distribution. To eliminate hyperparameter selection, we model the distribution of noisy video features across a range of noise levels and introduce a regularizer that tends to align the models for different levels of noise. At test time, we combine anomaly indications at multiple noise scales with a Gaussian mixture model. Running our video anomaly detector induces minimal delays as inference requires merely extracting the features and forward-propagating them through a shallow neural network and a Gaussian mixture model. Our experiments on five popular video anomaly detection benchmarks demonstrate state-of-the-art performance both in the object-centric and in the frame-centric setup.

Although neural networks excel in video action recognition tasks, their ”black-box” nature makes it challenging to understand the rationale behind their decisions. Recent approaches used inherently interpretable models to analyze video actions in a manner akin to human reasoning. However, it has been observed that these interpretable models tend to underperform when compared to their black-box counterparts. In this work, we present a new framework, called Language-guided Interpretable Action Recognition framework (LaIAR). This framework leverages knowledge from language models to enhance both the recognition capabilities and the interpretability of video models. In essence, we reframe the challenge of understanding video model decisions as a task of aligning video and language models. Using the logical reasoning captured by the language model, we steer the training of the video model. This integrated approach not only improves the video model’s adaptability to different domains but also boosts its overall performance. Extensive experiments on Charades and CAD-120 datasets demonstrate the superior performance and interpretability of our proposed method. The code of LaIAR is available at https://anonymous.

Due to the resource-intensive nature of training vision-language models on expansive video data, a majority of studies have centered on adapting pre-trained image-language models to the video domain. Dominant pipelines propose to tackle the visual discrepancies with additional temporal learners while overlooking the substantial discrepancy for web-scaled descriptive narratives and concise action category names, leading to less distinct semantic space and potential performance limitations. In this work, we prioritize the refinement of text knowledge to facilitate generalizable video recognition. To address the limitations of the less distinct semantic space of category names, we prompt a large language model (LLM) to augment action class names into Spatio-Temporal Descriptors thus bridging the textual discrepancy and serving as a knowledge base for general recognition. Moreover, to assign the best descriptors with different video instances, we propose Optimal Descriptor Solver, forming the video recognition problem as solving the optimal matching flow across frame-level representations and descriptors. Comprehensive evaluations in zero-shot, few-shot, and fully supervised video recognition highlight the effectiveness of our approach. Our best model achieves a state-of-the-art zero-shot accuracy of 75.1% on Kinetics-600.

Weakly supervised video anomaly detection (WSVAD) is a challenging task. Generating fine-grained pseudo-labels based on weak-label and then self-training a classifier is currently a promising solution. However, since the existing methods use only RGB visual modality and the utilization of category text information is neglected, thus limiting the generation of more accurate pseudo-labels and affecting the performance of self-training. Inspired by the manual labeling process based on the event description, in this paper, we propose a novel pseudo-label generation and self-training framework based on Text Prompt with Normality Guidance (TPWNG) for WSVAD. Our idea is to transfer the rich language-visual knowledge of the contrastive language-image pre-training (CLIP) model for aligning the video event description text and corresponding video frames to generate pseudo-labels. Specifically, We first fine-tune the CLIP for domain adaptation by designing two ranking losses and a distributional inconsistency loss. Further, we propose a learnable text prompt mechanism with the assist of a normality visual prompt to further improve the matching accuracy of video event description text and video frames. Then, we design a pseudo-label generation module based on the normality guidance to infer reliable frame-level pseudo-labels. Finally, we introduce a temporal context self-adaptive learning module to learn the temporal dependencies of different video events more flexibly and accurately. Extensive experiments show that our method achieves state-of-the-art performance on two benchmark datasets, UCF-Crime and XD-Violence, demonstrating the effectiveness of our proposed method.

Video grounding aims to localize a spatio-temporal section in a video corresponding to an input text query. This paper addresses a critical limitation in current video grounding methodologies by introducing an Open-Vocabulary Spatio-Temporal Video Grounding task. Unlike prevalent closed-set approaches that struggle with open-vocabulary scenarios due to limited training data and predefined vocabularies, our model leverages pre-trained representations from foundational spatial grounding models. This empowers it to effectively bridge the semantic gap between natural language and diverse visual content, achieving strong performance in closed-set and open-vocabulary settings. Our contributions include a novel spatio-temporal video grounding model, surpassing state-of-the-art results in closed-set evaluations on multiple datasets and demonstrating superior performance in open-vocabulary scenarios. Notably, the proposed model outperforms state-of-the-art methods in closed-set settings on VidSTG (Declarative and Interrogative) and HC-STVG (V1 and V2) datasets. Furthermore, in open-vocabulary evaluations on HC-STVG V1 and YouCook-Interactions, our model surpasses the recent best-performing models by 4.88 m_vIoU and 1.83\% accuracy, demonstrating its efficacy in handling diverse linguistic and visual concepts for improved video understanding. Our codes will be publicly released.

In this work, we tackle the problem of unsupervised domain adaptation (UDA) for video action recognition. Our approach, which we call UNITE, uses an image teacher model to adapt a video student model to the target domain. UNITE first employs self-supervised pre-training to promote discriminative feature learning on target domain videos using a teacher-guided masked distillation objective. We then perform self-training on masked target data, using the video student model and image teacher model together to generate improved pseudolabels for unlabeled target videos. Our self-training process successfully leverages the strengths of both models to achieve strong transfer performance across domains. We evaluate our approach on multiple video domain adaptation benchmarks and observe significant improvements upon previously reported results.

Temporal grounding of text descriptions in video is an important task amid vision-language learning, and remains a challenging problem in video understanding. Existing methods focus on grounding a few text queries within minute-long videos, yet fail to scale up to hour-long videos with hundreds of queries. In this paper, we present a systematic study for the design of scalable video grounding models. We compare design choices for cross-modal fusion, analyze their computational cost, and point out key insight and a new training scheme that enables scalable video grounding. We further present a simple model following our key findings. Our model attains superior accuracy and efficiency on recent benchmarks for long-form video grounding, while remaining highly competitive on previous benchmarks comprising short videos.

We introduce a new attention mechanism, dubbed structural self-attention (StructSA), that leverages rich correlation patterns naturally emerging in key-query interactions of attention. StructSA generates attention maps by recognizing space-time structures of key-query correlations via convolution and uses them to dynamically aggregate local contexts of value features. This effectively leverages rich structural patterns in images and videos such as scene layouts, object motion, and inter-object relations.Using StructSA as a main building block, we develop the structural vision transformer (StructViT) and evaluate its effectiveness on both image and video classification tasks, achieving state-of-the-art results on ImageNet-1K, Kinetics-400, Something-Something V1 & V2, Diving-48, and FineGym.

In this paper, we address the weakly-supervised Audio-Visual Video Parsing (AVVP) problem, which aims at labeling events in a video as audible, visible, or both, and temporally localizing and classifying them into known categories. This is challenging since we only have access to video-level (weak) event labels when training but need to predict event labels at the segment (frame) level at test time. Recent methods employ multiple-instance learning (MIL) techniques that tend to focus solely on the most discriminative segments, resulting in frequent misclassifications. Our idea is to first construct several "prototype'' features for each event class by clustering key segments identified for the event in the training data. We then assign pseudo labels to all training segments based on their feature similarities with these prototypes and re-train the model under weak and strong supervision. We facilitate this by structuring the feature space with contrastive learning using pseudo labels. Experiments show that we outperform existing methods for weakly-supervised AVVP. We also show that learning with weak and iteratively re-estimated pseudo labels can be interpreted as an expectation-maximization (EM) algorithm, providing further insight for our training procedure.

The robust association of the same objects across video frames in complex scenes is crucial for many applications, especially multiple object tracking (MOT). Current methods predominantly rely on labeled domain-specific video datasets, which limits the cross-domain generalization of learned similarity embeddings.We propose MASA, a novel method for robust instance association learning, capable of matching any objects within videos across diverse domains without tracking labels. Leveraging the rich object segmentation from the Segment Anything Model (SAM), MASA learns instance-level correspondence through exhaustive data transformations. We treat the SAM outputs as dense object region proposals and learn to match those regions from a vast image collection.We further design a universal MASA adapter which can work in tandem with foundational segmentation or detection models and enable them to track any detected objects. Those combinations present strong zero-shot tracking ability in complex domains.Extensive tests on multiple challenging MOT and MOTS benchmarks indicate that the proposed method, using only unlabelled static images, achieves even better performance than state-of-the-art methods trained with fully annotated in-domain video sequences, in zero-shot association. Our code is available at https://github.com/siyuanliii/masa.

This paper presents the first 3D feature tracking method with the corresponding dataset. Our proposed method takes event streams from stereo event cameras as input to predict 3D trajectories of the target features with high-speed motion. To achieve this, our method leverages a joint framework to predict the 2D feature motion offsets and the 3D feature spatial position simultaneously. A motion compensation module is leveraged to overcome the feature deformation. A patch matching module based on bi-polarity hypergraph modeling is proposed to robustly estimate the feature spatial position. Meanwhile, we collect the first 3D feature tracking dataset with high-speed moving objects and ground truth 3D feature trajectories at 250 FPS, named E-3DTrack, which can be used as the first high-speed 3D feature tracking benchmark. Our code and dataset could be found at: https://github.com/lisiqi19971013/E-3DTrack.

Video Motion Magnification (VMM) aims to reveal subtle and imperceptible motion information of objects in the macroscopic world. Prior methods directly model the motion field from the Eulerian perspective by Representation Learning that separates shape and texture or Multi-domain Learning from phase fluctuations. Inspired by the frequency spectrum, we observe that the low-frequency components with stable energy always possess spatial structure and less noise, making them suitable for modeling the subtle motion field. To this end, we present FD4MM, a new paradigm of Frequency Decoupling for Motion Magnification with a Multi-level Isomorphic Architecture to capture multi-level high-frequency details and a stable low-frequency structure (motion field) in video space. Since high-frequency details and subtle motions are susceptible to information degradation due to their inherent subtlety and unavoidable external interference from noise, we carefully design Sparse High/Low-pass Filters to enhance the integrity of details and motion structures, and a Sparse Frequency Mixer to promote seamless recoupling. Besides, we innovatively design a contrastive regularization for this task to strengthen the model's ability to discriminate irrelevant features, reducing undesired motion magnification. Extensive experiments on both Real-world and Synthetic Datasets show that our FD4MM outperforms SOTA methods. Meanwhile, FD4MM reduces FLOPs by 1.63$\times$ and boosts inference speed by 1.68$\times$ than the latest method. Our code is available at https://github.com/Jiafei127/FD4MM.

Multi-Object Tracking (MOT) encompasses various tracking scenarios, each characterized by unique traits. Effective trackers should demonstrate a high degree of generalizability across diverse scenarios. However, existing trackers struggle to accommodate all aspects or necessitate hypothesis and experimentation to customize the association information (motion and/or appearance) for a given scenario, leading to narrowly tailored solutions with limited generalizability. In this paper, we investigate the factors that influence trackers' generalization to different scenarios and concretize them into a set of tracking scenario attributes to guide the design of more generalizable trackers. Furthermore, we propose some designs to address the challenges of various scenarios and integrate them within a unified tracker, i.e., GeneralTrack, which can generalize across diverse scenarios while eliminating the need to balance motion and appearance. Thanks to its superior generalizability, our proposed GeneralTrack achieves state-of-the-art performance on five public datasets, i.e., BDD100K, SportsMOT, DanceTrack, MOT17 and MOT20. Moreover, we also demonstrate the strong potential of our tracker for domain generalization. We will make the code public.

Analyzing and forecasting trajectories of agents like pedestrians and cars in complex scenes has become more and more significant in many intelligent systems and applications. The diversity and uncertainty in socially interactive behaviors among a rich variety of agents make this task more challenging than other deterministic computer vision tasks. Researchers have made a lot of efforts to quantify the effects of these interactions on future trajectories through different mathematical models and network structures, but this problem has not been well solved. Inspired by marine animals that localize the positions of their companions underwater through echoes, we build a new anglebased trainable social interaction representation, named SocialCircle, for continuously reflecting the context of social interactions at different angular orientations relative to the target agent. We validate the effect of the proposed SocialCircle by training it along with several newly released trajectory prediction models, and experiments show that the SocialCircle not only quantitatively improves the prediction performance, but also qualitatively helps better simulate social interactions when forecasting pedestrian trajectories in a way that is consistent with human intuitions.

In this paper, we propose a novel concept of path consistency to learn robust object association without using the manual object identity supervision. Our key idea is that, to track a object through frames, we can obtain multiple different association results from a model by varying the frames it can observe, i.e., skipping frames in observation. As the difference in observations does not alter the identities of objects, the obtained association results should be consistent. Based on this rationale, we generate multiple paths by skipping observations in intermediate frames and formulate the Path Consistency Loss that enforces the association results are consistent with those different observation paths. We train an object matching model with the proposed loss, and with extensive experiments on three tracking datasets (MOT17, PersonPath22, KITTI), we demonstrate that our method outperforms existing unsupervised methods with consistent margins on various evaluation metrics, and even achieves performance close to supervised methods.

Traditional unsupervised optical flow methods are vulnerable to occlusions and motion boundaries due to lack of object-level information. Therefore, we propose UnSAMFlow, an unsupervised flow network that also leverages object information from the latest foundation model Segment Anything Model (SAM). We first include a self-supervised semantic augmentation module tailored to SAM masks. We also analyze the poor gradient landscapes of traditional smoothness losses and propose a new smoothness definition based on homography instead. A simple yet effective mask feature module has also been added to further aggregate features on the object level. With all these adaptations, our method produces clear optical flow estimation with sharp boundaries around objects, which outperforms state-of-the-art methods on both KITTI and Sintel datasets. Our method also generalizes well across domains and runs very efficiently.

Existing tracking methods mainly focus on learning better target representation or developing more robust prediction models to improve tracking performance. While tracking performance has significantly improved, the target loss issue occurs frequently due to tracking failures, complete occlusion, or out-of-view situations. However, considerably less attention is paid to the self-recovery issue of tracking methods, which is crucial for practical applications. To this end, we propose a recoverable tracking framework, RTracker, that uses a tree-structured memory to dynamically associate a tracker and a detector to enable self-recover ability. Specifically, we propose a Positive-Negative Tree-structured memory to chronologically store and maintain positive and negative target samples. Upon the PN tree memory, we develop corresponding walking rules for determining the state of the target and define a set of control flows to unite the tracker and the detector in different tracking scenarios. Our core idea is to use the support samples of positive and negative target categories to establish a relative distance-based criterion for a reliable assessment of target loss. The favorable performance in comparison against the state-of-the-art methods on numerous challenging benchmarks demonstrates the effectiveness of the proposed algorithm. All the source code and trained models will be released.

We present ARTrackV2, which integrates two pivotal aspects of tracking: determining where to look (localization) and how to describe (appearance analysis) the target object across video frames. Building on the foundation of its predecessor, ARTrackV2 extends the concept by introducing a unified generative framework to "read out" object's trajectory and "retell" its appearance in an autoregressive manner. This approach fosters a time-continuous methodology that models the joint evolution of motion and visual features, guided by previous estimates. Furthermore, ARTrackV2 stands out for its efficiency and simplicity, obviating the less efficient intra-frame autoregression and hand-tuned parameters for appearance updates. Despite its simplicity, ARTrackV2 achieves state-of-the-art performance on prevailing benchmark datasets while demonstrating remarkable efficiency improvement. In particular, ARTrackV2 achieves AO score of 79.5\% on GOT-10k, and AUC of 86.1\% on TrackingNet while being $3.6 \times$ faster than ARTrack. The code will be released.

The Segment Anything Model (SAM), a prompt-driven foundational model, has demonstrated remarkable performance in natural image segmentation. However, its application in video camouflaged object detection (VCOD) encounters challenges, chiefly stemming from the overlooked temporal-spatial associations and the unreliability of user-provided prompts for camouflaged objects that are difficult to discern with the naked eye. To tackle the above issues, we endow SAM with keen eyes and propose the Temporal-spatial Prompt SAM (TSP-SAM), a novel approach tailored for VCOD via an ingenious prompted learning scheme.Firstly, motion-driven self-prompt learning is employed to capture the camouflaged object, thereby bypassing the need for user-provided prompts. With the detected subtle motion cues across consecutive video frames, the overall movement of the camouflaged object is captured for more precise spatial localization. Subsequently, to eliminate the prompt bias resulting from inter-frame discontinuities, the long-range consistency within the video sequences is taken into account to promote the robustness of the self-prompts. It is also injected into the encoder of SAM to enhance the representational capabilities. Extensive experimental results on two benchmarks demonstrate that the proposed TSP-SAM achieves a significant improvement over the state-of-the-art methods. With the mIoU metric increasing by 7.8% and 9.6%, TSP-SAM emerges as a groundbreaking step forward in the field of VCOD.

Optical flow is a classical task that is important to the vision community. Classical optical flow estimation uses two frames as input, whilst some recent methods consider multiple frames to explicitly model long-range information. The former ones limit their ability to fully leverage temporal coherence along the video sequence; and the latter ones incur heavy computational overhead, typically not possible for real-time flow estimation. Some multi-frame-based approaches even necessitate unseen future frames for current estimation, compromising real-time applicability in safety-critical scenarios. To this end, we present MemFlow, a real-time method for optical flow estimation and prediction with memory. Our method enables memory read-out and update modules for aggregating historical motion information in real-time. Furthermore, we integrate resolution-adaptive re-scaling to accommodate diverse video resolutions. Besides, our approach seamlessly extends to the future prediction of optical flow based on past observations. Leveraging effective historical motion aggregation, our method outperforms VideoFlow with fewer parameters and faster inference speed on Sintel and KITTI-15 datasets in terms of generalization performance. At the time of submission, MemFlow also leads in performance on the 1080p Spring dataset. Codes and models will be available at: https://dqiaole.github.io/MemFlow/.

Visual object tracking aims to localize the target object of each frame based on its initial appearance in the first frame. Depending on the input modility, tracking tasks can be divided into RGB tracking and RGB+X (e.g. RGB+N, and RGB+D) tracking. Despite the different input modalities, the core aspect of tracking is the temporal matching. Based on this common ground, we present a general framework to unify various tracking tasks, termed as OneTracker. OneTracker first performs a large-scale pre-training on a RGB tracker called Foundation Tracker. This pretraining phase equips the Foundation Tracker with a stable ability to estimate the location of the target object. Then we regard other modality information as prompt and build Prompt Tracker upon Foundation Tracker. Through freezing the Foundation Tracker and only adjusting some additional trainable parameters, Prompt Tracker inhibits the strong localization ability from Foundation Tracker and achieves parameter-efficient finetuning on downstream RGB+X tracking tasks. To evaluate the effectiveness of our general framework OneTracker, which is consisted of Foundation Tracker and Prompt Tracker, we conduct extensive experiments on 6 popular tracking tasks across 11 benchmarks and our OneTracker outperforms other models and achieves state-of-the-art performance.

Clustering multiple motions from observed point trajectories is a fundamental task in understanding dynamic scenes. Most motion models require multiple tracks to estimate their parameters, hence identifying clusters when multiple motions are observed is a very challenging task. This is even aggravated for high-dimensional motion models. The starting point of our work is that this high-dimensionality of motion model can actually be leveraged to our advantage as sufficiently long trajectories identify the underlying motion uniquely in practice. Consequently, we propose to learn a mapping from trajectories to embedding vectors that represent the generating motion. The obtained trajectory embeddings are useful for clustering multiple observed motions, but are also trained to contain sufficient information to recover the parameters of the underlying motion by utilizing a geometric loss. We therefore are able to use only weak supervision from given motion segmentation to train this mapping. The entire algorithm consisting of trajectory embedding, clustering and motion parameter estimation is highly efficient. We conduct experiments on the Hopkins155, Hopkins12 and KT3DMoSeg datasets and show state-of-the-art performance of our proposed method for trajectory-based motion segmentation on full sequences and its competitiveness on the occluded sequences.

The primary focus of Neural Representation for Videos (NeRV) is the effective modeling of spatiotemporal consistency. However, present NeRV systems often face a significant issue of spatial inconsistency, which leads to a decrease in perceptual quality. To combat this issue, we introduce the Pyramidal Neural Representation for Videos (PNeRV), which is built on a multi-scale information connection and comprises a lightweight rescaling operator, Kronecker Fully-connected layer (KFc), and a Benign Selective Memory (BSM) mechanism. The KFc, inspired by the tensor decomposition of the vanilla Fully-connected layer, facilitates low-cost rescaling and global correlation modeling. BSM merges high-level features with granular ones adaptively. Furthermore, we provide the analysis based on Universal Approximation Theory towards NeRV system and validate the effectiveness of the proposed PNeRV. We have conducted comprehensive experiments to illustrate that PNeRV surpasses the performance of contemporary NeRV models, achieving the best results in video regression on UVG and DAVIS under various metrics (PSNR, SSIM, LPIPS, and FVD). When compared to the vanilla NeRV, PNeRV achieves +4.49db gain in PSNR and a 231% increase in FVD on UVG, along with +3.28db PSNR and 634% FVD increase on DAVIS.

Existing Siamese or transformer trackers commonly pose visual object tracking as a one-shot detection problem, i.e., locating the target object in a \textbf{single forward evaluation} scheme. Despite the demonstrated success, these trackers may easily drift towards distractors with similar appearance due to the single forward evaluation scheme lacking self-correction. To address this issue, we cast visual tracking as a point-set based denoising diffusion process and propose a novel generative learning based tracker, dubbed DiffusionTrack. Our DiffusionTrack possesses two appealing properties: 1) It takes a novel noise-to-target tracking paradigm that leverages multiple denoising diffusion steps to localize the target in a dynamic searching manner per frame. 2) It learns diffusion models on point-set, which can better handle appearance variations for more precise localization. One side benefit is that DiffusionTrack greatly simplifies the post-processing of locating the target in bounding boxes as the widely used window penalty scheme is no longer needed for point-set. Without bells and whistles, our proposed DiffusionTrack achieves the leading performance over the state-of-the-art trackers and runs in real-time.

The challenge posed by large motion plays a crucial role in the task of Video Frame Interpolation (VFI) for handling the potentially significant temporal gap between input inference frames. Existing algorithms are often constrained by limited receptive fields or rely on local refinement, resulting in suboptimal performance when dealing with scenarios with large motion. In this paper, we introduce a sparse global matching algorithm to specifically address Video Frame Interpolation (VFI) challenges associated with large motion. Our two-step modeling approach efficiently captures both local details and global motion correlations to overcome the limitations of existing methods in handling large motion. First, we estimate a pair of initial intermediate flows using a hybrid structure of CNN and Transformer for local details. Then, we incorporate a sparse global matching block to identify unmatched flaws in flow estimation and generate sparse flow fixes within a global receptive field. Our method not only achieves the state-of-the-art performance in the most challenging subset of the commonly used large motion benchmarks, X-Test, Xiph and SNU-FILM hard and extreme, but also maintains very competitive performance on SNU-FILM easy and medium.

Referring multi-object tracking (RMOT) aims to track multiple objects based on input textual descriptions. Previous works realize it by simply integrating an extra textual module into the multi-object tracker. However, they typically need to retrain the entire framework and have difficulties in optimization. In this work, we propose an insertable Knowledge Unification Network, termed iKUN, to enable communication with off-the-shelf trackers in a plug-and-play manner. Concretely, a knowledge unification module (KUM) is designed to adaptively extract visual features based on textual guidance. Meanwhile, to improve the localization accuracy, we present a neural version of Kalman filter (NKF) to dynamically adjust process noise and observation noise based on the current motion status. Moreover, to address the problem of open-set long-tail distribution of textual descriptions, a test-time similarity calibration method is proposed to refine the confidence score with pseudo frequency. Extensive experiments on Refer-KITTI dataset verify the effectiveness of our framework. Finally, to speed up the development of RMOT, we also contribute a more challenging dataset, Refer-Dance, by extending public DanceTrack dataset with motion and dressing descriptions. The codes and dataset are available at https://github.com/dyhBUPT/iKUN.

The complex dynamicity of open-world objects presents non-negligible challenges for multi-object tracking (MOT), often manifested as severe deformations, fast motion, and occlusions. Most methods that solely depend on coarse-grained object cues, such as boxes and the overall appearance of the object, are susceptible to degradation due to distorted internal relationships of dynamic objects.To address this problem, this work proposes NetTrack, an efficient, generic, and affordable tracking framework to introduce fine-grained learning that is robust to dynamicity. Specifically, NetTrack constructs a dynamicity-aware association with a fine-grained Net, leveraging point-level visual cues. Correspondingly, a fine-grained sampler and matching method have been incorporated. Furthermore, NetTrack learns object-text correspondence for fine-grained localization. To evaluate MOT in extremely dynamic open-world scenarios, a bird flock tracking (BFT) dataset is constructed, which exhibits high dynamicity with diverse species and open-world scenarios. Comprehensive evaluation on BFT validates the effectiveness of fine-grained learning on object dynamicity, and thorough transfer experiments on challenging open-world benchmarks, i.e., TAO, TAO-OW, AnimalTrack, and GMOT-40, validate the strong generalization ability of NetTrack even without finetuning. Code, dataset, and evaluation results are anonymously available in the supplementary materials.

In the realm of video object tracking, auxiliary modalities such as depth, thermal, or event data have emerged as valuable assets to complement the RGB trackers. In practice, most existing RGB trackers learn a single set of parameters to use them across datasets and applications. However, a similar single-model unification for multi-modality tracking presents several challenges. These challenges stem from the inherent heterogeneity of inputs -- each with modality-specific representations, the scarcity of multi-modal datasets, and the absence of all the modalities at all times. In this work, we introduce Un-Track, a \underline{Un}ified Tracker of a single set of parameters for any modality. To handle any modality, our method learns their common latent space through low-rank factorization and reconstruction techniques. More importantly, we use only the RGB-X pairs to learn the common latent space. This unique shared representation seamlessly binds all modalities together, enabling effective unification and accommodating any missing modality, all within a single transformer-based architecture and without the need for modality-specific fine-tuning. Our Un-Track achieves +8.1 absolute F-score gain, on the DepthTrack dataset, by introducing only +2.14 (over 21.50) GFLOPs with +6.6M (over 93M) parameters, through a simple yet efficient prompting strategy. Extensive comparisons on five benchmark datasets with different modalities show that Un-Track surpasses both SOTA unified trackers and modality-specific finetuned counterparts, validating our effectiveness and practicality. The source code will be made publicly available.

Optical flow estimation, a process of predicting pixel-wise displacement between consecutive frames, has commonly been approached as a regression task in the age of deep learning. Despite notable advancements, this de-facto paradigm unfortunately falls short in generalization performance when trained on synthetic or constrained data. Pioneering a paradigm shift, we recast optical flow estimation as a conditional flow generation challenge, unveiling FlowDiffuser --- a new family of optical flow models that could have stronger learning and generalization capabilities. FlowDiffuser estimates optical flow through a `noise-to-flow' strategy, progressively eliminating noise from randomly generated flows conditioned on the provided pairs. To optimize accuracy and efficiency, our FlowDiffuser incorporates a novel Conditional Recurrent Denoising Decoder (Conditional-RDD), streamlining the flow estimation process. It incorporates a unique Hidden State Denoising (HSD) paradigm, effectively leveraging the information from previous time steps. Moreover, FlowDiffuser can be easily integrated into existing flow networks, leading to significant improvements in performance metrics compared to conventional implementations. Experiments on challenging benchmarks, including Sintel and KITTI, demonstrate the effectiveness of our FlowDiffuser with superior performance to existing state-of-the-art models. Our code will be made publicly available upon acceptance.

Video harmonization is an important and challenging task that aims to obtain visually realistic composite videos by automatically adjusting the foreground's appearance to harmonize with the background. Inspired by the short-term and long-term gradual adjustment process of manual harmonization, we present a Video Triplet Transformer framework to model three spatio-temporal variation patterns within videos, \ie, short-term spatial as well as long-term global and dynamic, for video-to-video tasks like video harmonization. Specifically, for short-term harmonization, we adjust foreground appearance to consist with background in spatial dimension based on the neighbor frames; for long-term harmonization, we not only explore global appearance variations to enhance temporal consistency but also alleviate motion offset constraints to align similar contextual appearances dynamically. Extensive experiments and ablation studies demonstrate the effectiveness of our method, achieving state-of-the-art performance in video harmonization, video enhancement, and video demoireing tasks. We also propose a temporal consistency metric to better evaluate the harmonized videos. Code is available at https://github.com/zhenglab/VideoTripletTransformer.

Recent approaches to point tracking are able to recover the trajectory of any scene point through a large portion of a video despite the presence of occlusions.They are, however, too slow in practice to track every point observed in a single frame in a reasonable amount of time.This paper introduces DOT, a novel, simple and efficient method for solving this problem.It first extracts a small set of tracks from key regions at motion boundaries using an off-the-shelf point tracking algorithm.Given source and target frames, DOT then computes rough initial estimates of a dense flow field and visibility mask through nearest-neighbor interpolation, before refining them using a learnable optical flow estimator that explicitly handles occlusions and can be trained on synthetic data with ground-truth correspondences.We show that DOT is significantly more accurate than current optical flow techniques, outperforms sophisticated ``universal'' trackers like OmniMotion, and is on par with, or better than, the best point tracking algorithms like CoTracker while being at least two orders of magnitude faster.Quantitative and qualitative experiments with synthetic and real videos validate the promise of the proposed approach.

In this paper, we explore the problem of event-based meshflow estimation, a novel task that involves predicting a spatially smooth sparse motion field from event cameras. To start, we generate a large-scale High-Resolution Event Meshflow (HREM) dataset, which showcases its superiority by encompassing the merits of high resolution at 1280x720, handling dynamic objects and complex motion patterns, and offering both optical flow and meshflow labels. These aspects have not been fully explored in previous works. Besides, we propose Efficient Event-based MeshFlow (EEMFlow) network, a lightweight model featuring a specially crafted encoder-decoder architecture to facilitate swift and accurate meshflow estimation. Furthermore, we upgrade EEMFlow network to support dense event optical flow, in which a Confidence-induced Detail Completion (CDC) module is proposed to preserve sharp motion boundaries. We conduct comprehensive experiments to show the exceptional performance and runtime efficiency (39x faster) of our EEMFlow model compared to recent state-of-the-art flow methods. Code and dataset will be released.

Tracking by natural language specification (TNL) aims to consistently localize a target in a video sequence given a linguistic description in the initial frame. Existing methodologies perform language-based and template-based matching for target reasoning separately and merge the matching results from two sources, which suffer from tracking drift when language and visual templates miss-align with the dynamic target state and ambiguity in the later merging stage. To tackle the issues, we propose a joint multi-modal tracking framework with 1) a prompt modulation module to leverage the complementarity between temporal visual templates and language expressions, enabling precise and context-aware appearance and linguistic cues, and 2) a unified target decoding module to integrate the multi-modal reference cues and executes the integrated queries on the search image to predict the target location in an end-to-end manner directly.This design ensures spatio-temporal consistency by leveraging historical visual information and introduces an integrated solution, generating predictions in a single step. Extensive experiments conducted on TNL2K, OTB-Lang, LaSOT, and RefCOCOg validate the efficacy of our proposed approach. The results demonstrate competitive performance against state-of-the-art methods for both tracking and grounding.

Zero-shot Video Object Segmentation (ZSVOS) aims at segmenting the primary moving object without any human annotations. Mainstream solutions mainly focus on learning a single model on large-scale video datasets, which struggle to generalize to unseen videos. In this work, we introduce a test-time training (TTT) strategy to address the problem. Our key insight is to enforce the model to predict consistent depth during the TTT process. In detail, we first train a single network to perform both segmentation and depth prediction tasks. This can be effectively learned with our specifically designed depth modulation layer. Then, for the TTT process, the model is updated by predicting consistent depth maps for the same frame under different data augmentations. In addition, we explore different TTT weight update strategies. Our empirical results suggest that the momentum-based weight initialization and looping-based training scheme lead to more stable improvements. Experiments show that the proposed method achieves clear improvements on ZSVOS. Our proposed video TTT strategy provides significant superiority over state-of-the-art TTT methods. Our code is available at: https://nifangbaage.github.io/DATTT/.

Video Individual Counting (VIC) aims to predict the number of unique individuals in a single video.Existing methods learn representations based on trajectory labels for individuals, which are annotation-expensive. To provide a more realistic reflection of the underlying practical challenge, we introduce a weakly supervised VIC task, wherein trajectory labels are not provided. Instead, two types of labels are provided to indicate traffic entering the field of view (inflow) and leaving the field view (outflow).We also propose the first solution as a baseline that formulates the task as a weakly supervised contrastive learning problem under group-level matching. In doing so, we devise an end-to-end trainable soft contrastive loss to drive the network to distinguish inflow, outflow, and the remaining.To facilitate future study in this direction, we generate annotations from the existing VIC datasets SenseCrowd and CroHD and also build a new dataset, UAVVIC.Extensive results show that our baseline weakly supervised method outperforms supervised methods, and thus, little information is lost in the transition to the more practically relevant weakly supervised task.

Unsupervised video object segmentation (VOS) aims to detect and segment the most salient object in videos. The primary techniques used in unsupervised VOS are 1) the collaboration of appearance and motion information; and 2) temporal fusion between different frames. This paper proposes two novel prototype-based attention mechanisms, inter-modality attention (IMA) and inter-frame attention (IFA), to incorporate these techniques via dense propagation across different modalities and frames. IMA densely integrates context information from different modalities based on a mutual refinement. IFA injects global context of a video to the query frame, enabling a full utilization of useful properties from multiple frames. Experimental results on public benchmark datasets demonstrate that our proposed approach outperforms all existing methods by a substantial margin. The proposed two components are also thoroughly validated via ablative study.

Tracking with bio-inspired event cameras has garnered increasing interest in recent years. Existing works either utilize aligned RGB and event data for accurate tracking or directly learn an event-based tracker. The former incurs higher inference costs while the latter may be susceptible to the impact of noisy events or sparse spatial resolution. In this paper, we propose a novel hierarchical knowledge distillation framework that can fully utilize multi-modal / multi-view information during training to facilitate knowledge transfer, enabling us to achieve high-speed and low-latency visual tracking during testing by using only event signals. Specifically, a teacher Transformer-based multi-modal tracking framework is first trained by feeding the RGB frame and event stream simultaneously. Then, we design a new hierarchical knowledge distillation strategy which includes pairwise similarity, feature representation, and response maps-based knowledge distillation to guide the learning of the student Transformer network. In particular, since existing event-based tracking datasets are all low-resolution ($346 \times 260$), we propose the first large-scale high-resolution ($1280 \times 720$) dataset named EventVOT. It contains 1141 videos and covers a wide range of categories such as pedestrians, vehicles, UAVs, ping pong, etc. Extensive experiments on both low-resolution (FE240hz, VisEvent, COESOT), and our newly proposed high-resolution EventVOT dataset fully validated the effectiveness of our proposed method. The dataset, evaluation toolkit, and source code will be released.

Trackers that follow Siamese paradigm utilize similarity matching between template and search region features for tracking. Many methods have been explored to enhance tracking performance by incorporating tracking history to better handle scenarios involving target appearance variations such as deformation and occlusion. However, the utilization of historical information in existing methods is insufficient and incomprehensive, which typically requires repetitive training and introduces a large amount of computation. In this paper, we show that by providing a tracker that follows Siamese paradigm with precise and updated historical information, a significant performance improvement can be achieved with completely unchanged parameters. Based on this, we propose a historical prompt network that uses refined historical foreground masks and historical visual features of the target to provide comprehensive and precise prompts for the tracker. We build a novel tracker called HIPTrack based on the historical prompt network, which achieves considerable performance improvements without the need to retrain the entire model. We conduct experiments on seven datasets and experimental results demonstrate that our method surpasses the current state-of-the-art trackers on LaSOT, LaSOText, GOT-10k and NfS. Furthermore, the historical prompt network can seamlessly integrate as a plug-and-play module into existing trackers, providing performance enhancements. The source code is available at https://github.com/WenRuiCai/HIPTrack.

In the domain of video tracking, existing methods often grapple with a trade-off between spatial density and temporal range. Current approaches in dense optical flow estimators excel in providing spatially dense tracking but are limited to short temporal spans. Conversely, recent advancements in long-range trackers offer extended temporal coverage but at the cost of spatial sparsity. This paper introduces FlowTrack, a novel framework designed to bridge this gap. FlowTrack combines the strengths of both paradigms by 1) chaining confident flow predictions to maximize efficiency and 2) automatically switching to an error compensation module in instances of flow prediction inaccuracies. This dual strategy not only offers efficient dense tracking over extended temporal spans but also ensures robustness against error accumulations and occlusions, common pitfalls of naive flow chaining. Furthermore, we demonstrate that chained flow itself can serve as an effective guide for an error compensation module, even for occluded points. Our framework achieves state-of-the-art accuracy for long-range tracking on the DAVIS dataset, and renders 50\% speed-up when performing dense tracking.

Recent advancements in video modeling extensively rely on optical flow to represent the relationships across frames, but this approach often lacks efficiency and fails to model the probability of the intrinsic motion of objects. In addition, conventional encoder-decoder frameworks in video processing focus on modeling the correlation in the encoder, leading to limited generative capabilities and redundant intermediate representations. To address these challenges, this paper proposes a novel Implicit Motion Function (IMF) method. Our approach utilizes a low-dimensional latent token as the implicit representation, along with the use of cross-attention, to implicitly model the correlation between frames. This enables the implicit modeling of temporal correlations and understanding of object motions. Our method not only improves sparsity and efficiency in representation but also explores the generative capabilities of the decoder by integrating correlation modeling within it. The IMF framework facilitates video editing and other generative tasks by allowing the direct manipulation of latent tokens. We validate the effectiveness of IMF through extensive experiments on multiple video tasks, demonstrating superior performance in terms of reconstructed video quality, compression efficiency and generation ability.

Accurate data association is crucial in reducing confusion, such as ID switches and assignment errors, in multi-object tracking (MOT). However, existing advanced methods often overlook the diversity among trajectories and the ambiguity and conflicts present in motion and appearance cues, leading to confusion among detections, trajectories, and associations when performing simple global data association. To address this issue, we propose a simple, versatile, and highly interpretable data association approach called Decomposed Data Association (DDA). DDA decomposes the traditional association problem into multiple sub-problems using a series of non-learning-based modules and selectively addresses the confusion in each sub-problem by incorporating targeted exploitation of new cues. Additionally, we introduce Occlusion-aware Non-Maximum Suppression (ONMS) to retain more occluded detections, thereby increasing opportunities for association with trajectories and indirectly reducing the confusion caused by missed detections. Finally, based on DDA and ONMS, we design a powerful multi-object tracker named DeconfuseTrack, specifically focused on resolving confusion in MOT. Extensive experiments conducted on the MOT17 and MOT20 datasets demonstrate that our proposed DDA and ONMS significantly enhance the performance of several popular trackers. Moreover, DeconfuseTrack achieves state-of-the-art performance on the MOT17 and MOT20 test sets, significantly outperforms the baseline tracker ByteTrack in metrics such as HOTA, IDF1, AssA. This validates that our tracking design effectively reduces confusion caused by simple global association.

The rich spatio-temporal information is crucial to capture the complicated target appearance variations in visual tracking. However, most top-performing tracking algorithms rely on many hand-crafted components for spatio-temporal information aggregation. Consequently, the spatio-temporal information is far away from being fully explored. To alleviate this issue, we propose an adaptive tracker with spatio-temporal transformers (named AQATrack), which adopts simple autoregressive queries to effectively learn spatio-temporal information without many hand-designed components. Firstly, we introduce a set of learnable and autoregressive queries to capture the instantaneous target appearance changes in a sliding window fashion. Then, we design a novel attention mechanism for the interaction of existing queries to generate a new query in current frame. Finally, based on the initial target template and learnt autoregressive queries, a spatio-temporal information fusion module (STM) is designed for spatiotemporal formation aggregation to locate a target object. Benefiting from the STM, we can effectively combine the static appearance and instantaneous changes to guide robust tracking. Extensive experiments show that our method significantly improves the tracker’s performance on six popular tracking benchmarks: LaSOT, LaSOText, TrackingNet, GOT-10k, TNL2K, and UAV123. Code and models will be https://github.com/orgs/GXNU-ZhongLab.

Video prediction is a challenging task due to its nature of uncertainty, especially for forecasting a long period. To model the temporal dynamics, advanced methods benefit from the recent success of diffusion models, and repeatedly refine the predicted future frames with 3D spatiotemporal U-Net. However, there exists a gap between the present and future and the repeated usage of U-Net brings a heavy computation burden. To address this, we propose a diffusion-based video prediction method that predicts future frames by extrapolating the present distribution of features, namely ExtDM. Specifically, our method consists of three components: (i) a motion autoencoder conducts a bijection transformation between video frames and motion cues; (ii) a layered distribution adaptor module extrapolates the present features in the guidance of Gaussian distribution; (iii) a 3D U-Net architecture specialized for jointly fusing guidance and features among the temporal dimension by spatiotemporal-window attention. Extensive experiments on five popular benchmarks covering short- and long-term video prediction verify the effectiveness of ExtDM.

In Multiple Object Tracking, objects often exhibit non-linear motion of acceleration and deceleration, with irregular direction changes. Tacking-by-detection (TBD) trackers with Kalman Filter motion prediction work well in pedestrian-dominant scenarios but fall short in complex situations when multiple objects perform non-linear and diverse motion simultaneously. To tackle the complex non-linear motion, we propose a real-time diffusion-based MOT approach named DiffMOT. Specifically, for the motion predictor component, we propose a novel Decoupled Diffusion-based Motion Predictor (D$^2$MP). It models the entire distribution of various motion presented by the data as a whole. It also predicts an individual object's motion conditioning on an individual's historical motion information. Furthermore, it optimizes the diffusion process with much fewer sampling steps. As a MOT tracker, the DiffMOT is real-time at 22.7FPS, and also outperforms the state-of-the-art on DanceTrack and SportsMOT datasets with 62.3% and 76.2% in HOTA metrics, respectively. To the best of our knowledge, DiffMOT is the first to introduce a diffusion probabilistic model into the MOT to tackle non-linear motion prediction.

Pedestrian trajectory prediction is a well-established area in human behavior research and various industries, with significant recent advancements. While attention has been shifted to predict long-term trajectories, existing datasets are unable to fulfill the demand for studying minute-level trajectory prediction, mainly due to the lack of high-resolution trajectory observation in wide field of view (FoV). To bridge this gap, we introduce a novel dataset named GigaTraj, specifically tailored for long-term trajectory predictions. GigaTraj provides videos with an expansive FoV spanning a vast area of $4 \times 10^4 ~m^2$ and high-resolution imagery at the gigapixel level. Furthermore, it includes comprehensive annotations such as bounding boxes, identity associations, world coordinates, face orientations, complex group/interaction relationships, and scene semantics. Leveraging these multimodal annotations, we evaluate and validate the state-of-the-art approaches for minute-level long-term trajectory prediction in large-scale scenes. Extensive experiments and analyses have revealed that long-term prediction for pedestrian trajectories present numerous challenges and unresolved issues, indicating a vital new direction for trajectory research. The dataset and our code will be made publicly available upon paper acceptance.

Multiple Object Tracking (MOT) is a critical area within computer vision, with a broad spectrum of practical implementations. Current research has primarily focused on the development of tracking algorithms and enhancement of post-processing techniques. Yet, there has been a lack of thorough examination concerning the nature of tracking data it self. In this study, we pioneer an exploration into the distribution patterns of tracking data and identify a pronounced long-tail distribution issue within existing MOT datasets. We note a significant imbalance in the distribution of trajectory lengths across different pedestrians, a phenomenon we refer to as ``pedestrians trajectory long-tail distribution''. Addressing this challenge, we introduce a bespoke strategy designed to mitigate the effects of this skewed distribution. Specifically, we propose two data augmentation strategies, including Stationary Camera View Data Augmentation (SVA) and Dynamic Camera View Data Augmentation (DVA) , designed for viewpoint states and the Group Softmax (GS) module for Re-ID. SVA is to backtrack and predict the pedestrian trajectory of tail classes, and DVA is to use diffusion model to change the background of the scene. GS divides the pedestrians into unrelated groups and performs softmax operation on each group individually. Our proposed strategies can be integrated into numerous existing tracking systems, and extensive experimentation validates the efficacy of our method in reducing the influence of long-tail distribution on multi-object tracking performance. The code is available at https://github.com/chen-si-jia/Trajectory-Long-tail-Distribution-for-MOT.

The scarcity of ground-truth labels poses one major challenge in developing optical flow estimation models that are both generalizable and robust. While current methods rely on data augmentation, they have yet to fully exploit the rich information available in labeled video sequences. We propose OCAI, a method that supports robust frame interpolation by generating intermediate video frames alongside optical flows in between. Utilizing a forward warping approach, OCAI employs occlusion awareness to resolve ambiguities in pixel values and fills in missing values by leveraging the forward-backward consistency of optical flows. Additionally, we introduce a teacher-student style semi-supervised learning method on top of the interpolated frames. Using a pair of unlabeled frames and the teacher model's predicted optical flow, we generate interpolated frames and flows to train a student model. The teacher's weights are maintained using Exponential Moving Averaging of the student. Our evaluations demonstrate perceptually superior interpolation quality and enhanced optical flow accuracy on established benchmarks such as Sintel and KITTI.

Understanding social interactions involving both verbal and non-verbal cues is essential for effectively interpreting social situations. However, most prior works on multimodal social cues focus predominantly on single-person behaviors or rely on holistic visual representations that are not aligned to utterances in multi-party environments. Consequently, they are limited in modeling the intricate dynamics of multi-party interactions. In this paper, we introduce three new challenging tasks to model the fine-grained dynamics between multiple people: speaking target identification, pronoun coreference resolution, and mentioned player prediction. We contribute extensive data annotations to curate these new challenges in social deduction game settings. Furthermore, we propose a novel multimodal baseline that leverages densely aligned language-visual representations by synchronizing visual features with their corresponding utterances. This facilitates concurrently capturing verbal and non-verbal cues pertinent to social reasoning. Experiments demonstrate the effectiveness of the proposed approach with densely aligned multimodal representations in modeling fine-grained social interactions. Project website: https://sangmin-git.github.io/projects/MMSI.

Egocentric videos provide a first-person perspective of the wearer's activities, involving simultaneous interactions with multiple objects. In this work, we propose the task of weakly-supervised Narration-based Video Object Segmentation (NVOS). Given an egocentric video clip and a narration of the wearer's activities, our aim is to segment object instances mentioned in the narration, without using any spatial annotations during training. Existing weakly-supervised video object grounding methods typically yield bounding boxes for referred objects. In contrast, we propose ROSA, a weakly-supervised pixel-level grounding framework learning alignments between referred objects and segmentation mask proposals. Our model harnesses vision-language models pre-trained on image-text pairs to embed region masks and object phrases. During training, we combine (a) a video-narration contrastive loss that implicitly supervises the alignment between regions and phrases, and (b) a region-phrase contrastive loss based on inferred latent alignments. To address the lack of annotated NVOS datasets in egocentric videos, we create a new evaluation benchmark, VISOR-NVOS, leveraging existing annotations of segmentation masks from VISOR alongside 12k newly-collected, object-based video clip narrations. Our approach achieves state-of-the-art zero-shot pixel-level grounding performance compared to strong baselines under similar supervision. Additionally, we demonstrate generalization capabilities for zero-shot video object grounding on YouCook2, a third-person instructional video dataset.