Computer vision applications have heavily relied on the linear combination of Lambertian diffuse and microfacet specular reflection models for representing reflected radiance, which turns out to be physically incompatible and limited in applicability. In this paper, we derive a novel analytical reflectance model, which we refer to as Fresnel Microfacet BRDF model, that is physically accurate and generalizes to various real-world surfaces. Our key idea is to model the Fresnel reflection and transmission of the surface microgeometry with a collection of oriented mirror facets, both for body and surface reflections. We carefully derive the Fresnel reflection and transmission for each microfacet as well as the light transport between them in the subsurface. This physically-grounded modeling also allows us to express the polarimetric behavior of reflected light in addition to its radiometric behavior. That is, FMBRDF unifies not only body and surface reflections but also light reflection in radiometry and polarization and represents them in a single model. Experimental results demonstrate its effectiveness in accuracy, expressive power, image-based estimation, and geometry recovery.

We propose a method that trains a neural radiance field (NeRF) to encode not only the appearance of the scene but also semantic correlations between scene points, regions, or entities -- aiming to capture their mutual co-variation patterns. In contrast to the traditional first-order photometric reconstruction objective, our method explicitly regularizes the learning dynamics to align the Jacobians of highly-correlated entities, which proves to maximize the mutual information between them under random scene perturbations. By paying attention to this second-order information, we can shape a NeRF to express semantically meaningful synergies when the network weights are changed by a delta along the gradient of a single entity, region, or even a point. To demonstrate the merit of this mutual information modeling, we leverage the coordinated behavior of scene entities that emerges from our shaping to perform label propagation for semantic and instance segmentation. Our experiments show that a JacobiNeRF is more efficient in propagating annotations among 2D pixels and 3D points compared to NeRFs without mutual information shaping, especially in extremely sparse label regimes -- thus reducing annotation burden. The same machinery can further be used for entity selection or scene modifications. Our code is available at https://github.com/xxm19/jacobinerf.

Although many recent works have investigated generalizable NeRF-based novel view synthesis for unseen scenes, they seldom consider the synthetic-to-real generalization, which is desired in many practical applications. In this work, we first investigate the effects of synthetic data in synthetic-to-real novel view synthesis and surprisingly observe that models trained with synthetic data tend to produce sharper but less accurate volume densities. For pixels where the volume densities are correct, fine-grained details will be obtained. Otherwise, severe artifacts will be produced. To maintain the advantages of using synthetic data while avoiding its negative effects, we propose to introduce geometry-aware contrastive learning to learn multi-view consistent features with geometric constraints. Meanwhile, we adopt cross-view attention to further enhance the geometry perception of features by querying features across input views. Experiments demonstrate that under the synthetic-to-real setting, our method can render images with higher quality and better fine-grained details, outperforming existing generalizable novel view synthesis methods in terms of PSNR, SSIM, and LPIPS. When trained on real data, our method also achieves state-of-the-art results. https://haoy945.github.io/contranerf/

Neural radiance fields (NeRFs) have enabled high fidelity 3D reconstruction from multiple 2D input views. However, a well-known drawback of NeRFs is the less-than-ideal performance under a small number of views, due to insufficient constraints enforced by volumetric rendering. To address this issue, we introduce SCADE, a novel technique that improves NeRF reconstruction quality on sparse, unconstrained input views for in-the-wild indoor scenes. To constrain NeRF reconstruction, we leverage geometric priors in the form of per-view depth estimates produced with state-of-the-art monocular depth estimation models, which can generalize across scenes. A key challenge is that monocular depth estimation is an ill-posed problem, with inherent ambiguities. To handle this issue, we propose a new method that learns to predict, for each view, a continuous, multimodal distribution of depth estimates using conditional Implicit Maximum Likelihood Estimation (cIMLE). In order to disambiguate exploiting multiple views, we introduce an original space carving loss that guides the NeRF representation to fuse multiple hypothesized depth maps from each view and distill from them a common geometry that is consistent with all views. Experiments show that our approach enables higher fidelity novel view synthesis from sparse views. Our project page can be found at https://scade-spacecarving-nerfs.github.io.

Neural Radiance Fields (NeRFs) are emerging as a ubiquitous scene representation that allows for novel view synthesis. Increasingly, NeRFs will be shareable with other people. Before sharing a NeRF, though, it might be desirable to remove personal information or unsightly objects. Such removal is not easily achieved with the current NeRF editing frameworks. We propose a framework to remove objects from a NeRF representation created from an RGB-D sequence. Our NeRF inpainting method leverages recent work in 2D image inpainting and is guided by a user-provided mask. Our algorithm is underpinned by a confidence based view selection procedure. It chooses which of the individual 2D inpainted images to use in the creation of the NeRF, so that the resulting inpainted NeRF is 3D consistent. We show that our method for NeRF editing is effective for synthesizing plausible inpaintings in a multi-view coherent manner, outperforming competing methods. We validate our approach by proposing a new and still-challenging dataset for the task of NeRF inpainting.

We present an algorithm for reconstructing the radiance field of a large-scale scene from a single casually captured video. The task poses two core challenges. First, most existing radiance field reconstruction approaches rely on accurate pre-estimated camera poses from Structure-from-Motion algorithms, which frequently fail on in-the-wild videos. Second, using a single, global radiance field with finite representational capacity does not scale to longer trajectories in an unbounded scene. For handling unknown poses, we jointly estimate the camera poses with radiance field in a progressive manner. We show that progressive optimization significantly improves the robustness of the reconstruction. For handling large unbounded scenes, we dynamically allocate new local radiance fields trained with frames within a temporal window. This further improves robustness (e.g., performs well even under moderate pose drifts) and allows us to scale to large scenes. Our extensive evaluation on the Tanks and Temples dataset and our collected outdoor dataset, Static Hikes, show that our approach compares favorably with the state-of-the-art.

We present NeRFVS, a novel neural radiance fields (NeRF) based method to enable free navigation in a room. NeRF achieves impressive performance in rendering images for novel views similar to the input views while suffering for novel views that are significantly different from the training views. To address this issue, we utilize the holistic priors, including pseudo depth maps and view coverage information, from neural reconstruction to guide the learning of implicit neural representations of 3D indoor scenes. Concretely, an off-the-shelf neural reconstruction method is leveraged to generate a geometry scaffold. Then, two loss functions based on the holistic priors are proposed to improve the learning of NeRF: 1) A robust depth loss that can tolerate the error of the pseudo depth map to guide the geometry learning of NeRF; 2) A variance loss to regularize the variance of implicit neural representations to reduce the geometry and color ambiguity in the learning procedure. These two loss functions are modulated during NeRF optimization according to the view coverage information to reduce the negative influence brought by the view coverage imbalance. Extensive results demonstrate that our NeRFVS outperforms state-of-the-art view synthesis methods quantitatively and qualitatively on indoor scenes, achieving high-fidelity free navigation results.

Neural Radiance Field (NeRF) is a popular method in representing 3D scenes by optimising a continuous volumetric scene function. Its large success which lies in applying volumetric rendering (VR) is also its Achilles’ heel in producing view-dependent effects. As a consequence, glossy and transparent surfaces often appear murky. A remedy to reduce these artefacts is to constrain this VR equation by excluding volumes with back-facing normal. While this approach has some success in rendering glossy surfaces, translucent objects are still poorly represented. In this paper, we present an alternative to the physics-based VR approach by introducing a self-attention-based framework on volumes along a ray. In addition, inspired by modern game engines which utilise Light Probes to store local lighting passing through the scene, we incorporate Learnable Embeddings to capture view dependent effects within the scene. Our method, which we call ABLE-NeRF, significantly reduces ‘blurry’ glossy surfaces in rendering and produces realistic translucent surfaces which lack in prior art. In the Blender dataset, ABLE-NeRF achieves SOTA results and surpasses Ref-NeRF in all 3 image quality metrics PSNR, SSIM, LPIPS.

Neural Radiance Fields (NeRFs) have demonstrated amazing ability to synthesize images of 3D scenes from novel views. However, they rely upon specialized volumetric rendering algorithms based on ray marching that are mismatched to the capabilities of widely deployed graphics hardware. This paper introduces a new NeRF representation based on textured polygons that can synthesize novel images efficiently with standard rendering pipelines. The NeRF is represented as a set of polygons with textures representing binary opacities and feature vectors. Traditional rendering of the polygons with a z-buffer yields an image with features at every pixel, which are interpreted by a small, view-dependent MLP running in a fragment shader to produce a final pixel color. This approach enables NeRFs to be rendered with the traditional polygon rasterization pipeline, which provides massive pixel-level parallelism, achieving interactive frame rates on a wide range of compute platforms, including mobile phones.

Neural scene representations have achieved great success in parameterizing and reconstructing images, but current state of the art models are not optimized with the preservation of physical quantities in mind. While current architectures can reconstruct color images correctly, they create artifacts when trying to fit maps of polar quantities. We propose polarimetric coordinate networks (pCON), a new model architecture for neural scene representations aimed at preserving polarimetric information while accurately parameterizing the scene. Our model removes artifacts created by current coordinate network architectures when reconstructing three polarimetric quantities of interest.

We present EgoNeRF, a practical solution to reconstruct large-scale real-world environments for VR assets. Given a few seconds of casually captured 360 video, EgoNeRF can efficiently build neural radiance fields which enable high-quality rendering from novel viewpoints. Motivated by the recent acceleration of NeRF using feature grids, we adopt spherical coordinate instead of conventional Cartesian coordinate. Cartesian feature grid is inefficient to represent large-scale unbounded scenes because it has a spatially uniform resolution, regardless of distance from viewers. The spherical parameterization better aligns with the rays of egocentric images, and yet enables factorization for performance enhancement. However, the naïve spherical grid suffers from irregularities at two poles, and also cannot represent unbounded scenes. To avoid singularities near poles, we combine two balanced grids, which results in a quasi-uniform angular grid. We also partition the radial grid exponentially and place an environment map at infinity to represent unbounded scenes. Furthermore, with our resampling technique for grid-based methods, we can increase the number of valid samples to train NeRF volume. We extensively evaluate our method in our newly introduced synthetic and real-world egocentric 360 video datasets, and it consistently achieves state-of-the-art performance.

Relighting an outdoor scene is challenging due to the diverse illuminations and salient cast shadows. Intrinsic image decomposition on outdoor photo collections could partly solve this problem by weakly supervised labels with albedo and normal consistency from multi-view stereo. With neural radiance fields (NeRFs), editing the appearance code could produce more realistic results without explicitly interpreting the outdoor scene image formation. This paper proposes to complement the intrinsic estimation from volume rendering using NeRFs and from inversing the photometric image formation model using convolutional neural networks (CNNs). The former produces richer and more reliable pseudo labels (cast shadows and sky appearances in addition to albedo and normal) for training the latter to predict interpretable and editable lighting parameters via a single-image prediction pipeline. We demonstrate the advantages of our method for both intrinsic image decomposition and relighting for various real outdoor scenes.

Volumetric scene representations enable photorealistic view synthesis for static scenes and form the basis of several existing 6-DoF video techniques. However, the volume rendering procedures that drive these representations necessitate careful trade-offs in terms of quality, rendering speed, and memory efficiency. In particular, existing methods fail to simultaneously achieve real-time performance, small memory footprint, and high-quality rendering for challenging real-world scenes. To address these issues, we present HyperReel --- a novel 6-DoF video representation. The two core components of HyperReel are: (1) a ray-conditioned sample prediction network that enables high-fidelity, high frame rate rendering at high resolutions and (2) a compact and memory-efficient dynamic volume representation. Our 6-DoF video pipeline achieves the best performance compared to prior and contemporary approaches in terms of visual quality with small memory requirements, while also rendering at up to 18 frames-per-second at megapixel resolution without any custom CUDA code.

Neural volume rendering enables photo-realistic renderings of a human performer in free-view, a critical task in immersive VR/AR applications. But the practice is severely limited by high computational costs in the rendering process. To solve this problem, we propose the UV Volumes, a new approach that can render an editable free-view video of a human performer in real-time. It separates the high-frequency (i.e., non-smooth) human appearance from the 3D volume, and encodes them into 2D neural texture stacks (NTS). The smooth UV volumes allow much smaller and shallower neural networks to obtain densities and texture coordinates in 3D while capturing detailed appearance in 2D NTS. For editability, the mapping between the parameterized human model and the smooth texture coordinates allows us a better generalization on novel poses and shapes. Furthermore, the use of NTS enables interesting applications, e.g., retexturing. Extensive experiments on CMU Panoptic, ZJU Mocap, and H36M datasets show that our model can render 960 x 540 images in 30FPS on average with comparable photo-realism to state-of-the-art methods.

We present Tensor4D, an efficient yet effective approach to dynamic scene modeling. The key of our solution is an efficient 4D tensor decomposition method so that the dynamic scene can be directly represented as a 4D spatio-temporal tensor. To tackle the accompanying memory issue, we decompose the 4D tensor hierarchically by projecting it first into three time-aware volumes and then nine compact feature planes. In this way, spatial information over time can be simultaneously captured in a compact and memory-efficient manner. When applying Tensor4D for dynamic scene reconstruction and rendering, we further factorize the 4D fields to different scales in the sense that structural motions and dynamic detailed changes can be learned from coarse to fine. The effectiveness of our method is validated on both synthetic and real-world scenes. Extensive experiments show that our method is able to achieve high-quality dynamic reconstruction and rendering from sparse-view camera rigs or even a monocular camera.

Lighting effects such as shadows or reflections are key in making synthetic images realistic and visually appealing. To generate such effects, traditional computer graphics uses a physically-based renderer along with 3D geometry. To compensate for the lack of geometry in 2D Image compositing, recent deep learning-based approaches introduced a pixel height representation to generate soft shadows and reflections. However, the lack of geometry limits the quality of the generated soft shadows and constrains reflections to pure specular ones. We introduce PixHt-Lab, a system leveraging an explicit mapping from pixel height representation to 3D space. Using this mapping, PixHt-Lab reconstructs both the cutout and background geometry and renders realistic, diverse, lighting effects for image compositing. Given a surface with physically-based materials, we can render reflections with varying glossiness. To generate more realistic soft shadows, we further propose to use 3D-aware buffer channels to guide a neural renderer. Both quantitative and qualitative evaluations demonstrate that PixHt-Lab significantly improves soft shadow generation.

Flash is an essential tool as it often serves as the sole controllable light source in everyday photography. However, the use of flash is a binary decision at the time a photograph is captured with limited control over its characteristics such as strength or color. In this work, we study the computational control of the flash light in photographs taken with or without flash. We present a physically motivated intrinsic formulation for flash photograph formation and develop flash decomposition and generation methods for flash and no-flash photographs, respectively. We demonstrate that our intrinsic formulation outperforms alternatives in the literature and allows us to computationally control flash in in-the-wild images.

We present the first neural relighting approach for rendering high-fidelity personalized hands that can be animated in real-time under novel illumination. Our approach adopts a teacher-student framework, where the teacher learns appearance under a single point light from images captured in a light-stage, allowing us to synthesize hands in arbitrary illuminations but with heavy compute. Using images rendered by the teacher model as training data, an efficient student model directly predicts appearance under natural illuminations in real-time. To achieve generalization, we condition the student model with physics-inspired illumination features such as visibility, diffuse shading, and specular reflections computed on a coarse proxy geometry, maintaining a small computational overhead. Our key insight is that these features have strong correlation with subsequent global light transport effects, which proves sufficient as conditioning data for the neural relighting network. Moreover, in contrast to bottleneck illumination conditioning, these features are spatially aligned based on underlying geometry, leading to better generalization to unseen illuminations and poses. In our experiments, we demonstrate the efficacy of our illumination feature representations, outperforming baseline approaches. We also show that our approach can photorealistically relight two interacting hands at real-time speeds.

We present a new pipeline for acquiring a textured mesh in the wild with a single smartphone which offers access to images, depth maps, and valid poses. Our method first introduces an RGBD-aided structure from motion, which can yield filtered depth maps and refines camera poses guided by corresponding depth. Then, we adopt the neural implicit surface reconstruction method, which allows for high quality mesh and develops a new training process for applying a regularization provided by classical multi-view stereo methods. Moreover, we apply a differentiable rendering to fine-tune incomplete texture maps and generate textures which are perceptually closer to the original scene. Our pipeline can be applied to any common objects in the real world without the need for either in-the-lab environments or accurate mask images. We demonstrate results of captured objects with complex shapes and validate our method numerically against existing 3D reconstruction and texture mapping methods.

The success of the Neural Radiance Fields (NeRF) in novel view synthesis has inspired researchers to propose neural implicit scene reconstruction. However, most existing neural implicit reconstruction methods optimize per-scene parameters and therefore lack generalizability to new scenes. We introduce VolRecon, a novel generalizable implicit reconstruction method with Signed Ray Distance Function (SRDF). To reconstruct the scene with fine details and little noise, VolRecon combines projection features aggregated from multi-view features, and volume features interpolated from a coarse global feature volume. Using a ray transformer, we compute SRDF values of sampled points on a ray and then render color and depth. On DTU dataset, VolRecon outperforms SparseNeuS by about 30% in sparse view reconstruction and achieves comparable accuracy as MVSNet in full view reconstruction. Furthermore, our approach exhibits good generalization performance on the large-scale ETH3D benchmark.

In this paper, we investigate a new optimization framework for multi-view 3D shape reconstructions. Recent differentiable rendering approaches have provided breakthrough performances with implicit shape representations though they can still lack precision in the estimated geometries. On the other hand multi-view stereo methods can yield pixel wise geometric accuracy with local depth predictions along viewing rays. Our approach bridges the gap between the two strategies with a novel volumetric shape representation that is implicit but parameterized with pixel depths to better materialize the shape surface with consistent signed distances along viewing rays. The approach retains pixel-accuracy while benefiting from volumetric integration in the optimization. To this aim, depths are optimized by evaluating, at each 3D location within the volumetric discretization, the agreement between the depth prediction consistency and the photometric consistency for the corresponding pixels. The optimization is agnostic to the associated photo-consistency term which can vary from a median-based baseline to more elaborate criteria, learned functions. Our experiments demonstrate the benefit of the volumetric integration with depth predictions. They also show that our approach outperforms existing approaches over standard 3D benchmarks with better geometry estimations.

The Multiplane Image (MPI), containing a set of fronto-parallel RGBA layers, is an effective and efficient representation for view synthesis from sparse inputs. Yet, its fixed structure limits the performance, especially for surfaces imaged at oblique angles. We introduce the Structural MPI (S-MPI), where the plane structure approximates 3D scenes concisely. Conveying RGBA contexts with geometrically-faithful structures, the S-MPI directly bridges view synthesis and 3D reconstruction. It can not only overcome the critical limitations of MPI, i.e., discretization artifacts from sloped surfaces and abuse of redundant layers, and can also acquire planar 3D reconstruction. Despite the intuition and demand of applying S-MPI, great challenges are introduced, e.g., high-fidelity approximation for both RGBA layers and plane poses, multi-view consistency, non-planar regions modeling, and efficient rendering with intersected planes. Accordingly, we propose a transformer-based network based on a segmentation model. It predicts compact and expressive S-MPI layers with their corresponding masks, poses, and RGBA contexts. Non-planar regions are inclusively handled as a special case in our unified framework. Multi-view consistency is ensured by sharing global proxy embeddings, which encode plane-level features covering the complete 3D scenes with aligned coordinates. Intensive experiments show that our method outperforms both previous state-of-the-art MPI-based view synthesis methods and planar reconstruction methods.

We address the problem of surface reconstruction from unoriented point clouds. Implicit neural representations (INRs) have become popular for this task, but when information relating to the inside versus outside of a shape is not available (such as shape occupancy, signed distances or surface normal orientation) optimization relies on heuristics and regularizers to recover the surface. These methods can be slow to converge and easily get stuck in local minima. We propose a two-step approach, OG-INR, where we (1) construct a discrete octree and label what is inside and outside (2) optimize for a continuous and high-fidelity shape using an INR that is initially guided by the octree’s labelling. To solve for our labelling, we propose an energy function over the discrete structure and provide an efficient move-making algorithm that explores many possible labellings. Furthermore we show that we can easily inject knowledge into the discrete octree, providing a simple way to influence the result from the continuous INR. We evaluate the effectiveness of our approach on two unoriented surface reconstruction datasets and show competitive performance compared to other unoriented, and some oriented, methods. Our results show that the exploration by the move-making algorithm avoids many of the bad local minima reached by purely gradient descent optimized methods (see Figure 1).

Deep neural networks (DNNs) are widely applied for nowadays 3D surface reconstruction tasks and such methods can be further divided into two categories, which respectively warp templates explicitly by moving vertices or represent 3D surfaces implicitly as signed or unsigned distance functions. Taking advantage of both advanced explicit learning process and powerful representation ability of implicit functions, we propose a novel 3D representation method, Neural Vector Fields (NVF). It not only adopts the explicit learning process to manipulate meshes directly, but also leverages the implicit representation of unsigned distance functions (UDFs) to break the barriers in resolution and topology. Specifically, our method first predicts the displacements from queries towards the surface and models the shapes as Vector Fields. Rather than relying on network differentiation to obtain direction fields as most existing UDF-based methods, the produced vector fields encode the distance and direction fields both and mitigate the ambiguity at “ridge” points, such that the calculation of direction fields is straightforward and differentiation-free. The differentiation-free characteristic enables us to further learn a shape codebook via Vector Quantization, which encodes the cross-object priors, accelerates the training procedure, and boosts model generalization on cross-category reconstruction. The extensive experiments on surface reconstruction benchmarks indicate that our method outperforms those state-of-the-art methods in different evaluation scenarios including watertight vs non-watertight shapes, category-specific vs category-agnostic reconstruction, category-unseen reconstruction, and cross-domain reconstruction. Our code is released at https://github.com/Wi-sc/NVF.

We present a neural technique for learning to select a local sub-region around a point which can be used for mesh parameterization. The motivation for our framework is driven by interactive workflows used for decaling, texturing, or painting on surfaces. Our key idea to to learn a local parameterization in a data-driven manner, using a novel differentiable parameterization layer within a neural network framework. We train a segmentation network to select 3D regions that are parameterized into 2D and penalized by the resulting distortion, giving rise to segmentations which are distortion-aware. Following training, a user can use our system to interactively select a point on the mesh and obtain a large, meaningful region around the selection which induces a low-distortion parameterization. Our code and project page are publicly available.

We introduce SceneDiffuser, a conditional generative model for 3D scene understanding. SceneDiffuser provides a unified model for solving scene-conditioned generation, optimization, and planning. In contrast to prior works, SceneDiffuser is intrinsically scene-aware, physics-based, and goal-oriented. With an iterative sampling strategy, SceneDiffuser jointly formulates the scene-aware generation, physics-based optimization, and goal-oriented planning via a diffusion-based denoising process in a fully differentiable fashion. Such a design alleviates the discrepancies among different modules and the posterior collapse of previous scene-conditioned generative models. We evaluate SceneDiffuser with various 3D scene understanding tasks, including human pose and motion generation, dexterous grasp generation, path planning for 3D navigation, and motion planning for robot arms. The results show significant improvements compared with previous models, demonstrating the tremendous potential of SceneDiffuser for the broad community of 3D scene understanding.

We target a 3D generative model for general natural scenes that are typically unique and intricate. Lacking the necessary volumes of training data, along with the difficulties of having ad hoc designs in presence of varying scene characteristics, renders existing setups intractable. Inspired by classical patch-based image models, we advocate for synthesizing 3D scenes at the patch level, given a single example. At the core of this work lies important algorithmic designs w.r.t the scene representation and generative patch nearest-neighbor module, that address unique challenges arising from lifting classical 2D patch-based framework to 3D generation. These design choices, on a collective level, contribute to a robust, effective, and efficient model that can generate high-quality general natural scenes with both realistic geometric structure and visual appearance, in large quantities and varieties, as demonstrated upon a variety of exemplar scenes. Data and code can be found at http://wyysf-98.github.io/Sin3DGen.

Novel view synthesis from a single image has been a cornerstone problem for many Virtual Reality applications that provide immersive experiences. However, most existing techniques can only synthesize novel views within a limited range of camera motion or fail to generate consistent and high-quality novel views under significant camera movement. In this work, we propose a pose-guided diffusion model to generate a consistent long-term video of novel views from a single image. We design an attention layer that uses epipolar lines as constraints to facilitate the association between different viewpoints. Experimental results on synthetic and real-world datasets demonstrate the effectiveness of the proposed diffusion model against state-of-the-art transformer-based and GAN-based approaches. More qualitative results are available at https://poseguided-diffusion.github.io/.

We develop a generalized 3D shape generation prior model, tailored for multiple 3D tasks including unconditional shape generation, point cloud completion, and cross-modality shape generation, etc. On one hand, to precisely capture local fine detailed shape information, a vector quantized variational autoencoder (VQ-VAE) is utilized to index local geometry from a compactly learned codebook based on a broad set of task training data. On the other hand, a discrete diffusion generator is introduced to model the inherent structural dependencies among different tokens. In the meantime, a multi-frequency fusion module (MFM) is developed to suppress high-frequency shape feature fluctuations, guided by multi-frequency contextual information. The above designs jointly equip our proposed 3D shape prior model with high-fidelity, diverse features as well as the capability of cross-modality alignment, and extensive experiments have demonstrated superior performances on various 3D shape generation tasks.

Despite the impressive performance obtained by recent single-image hand modeling techniques, they lack the capability to capture sufficient details of the 3D hand mesh. This deficiency greatly limits their applications when high fidelity hand modeling is required, e.g., personalized hand modeling. To address this problem, we design a frequency split network to generate 3D hand mesh using different frequency bands in a coarse-to-fine manner. To capture high-frequency personalized details, we transform the 3D mesh into the frequency domain, and propose a novel frequency decomposition loss to supervise each frequency component. By leveraging such a coarse-to-fine scheme, hand details that correspond to the higher frequency domain can be preserved. In addition, the proposed network is scalable, and can stop the inference at any resolution level to accommodate different hardwares with varying computational powers. To quantitatively evaluate the performance of our method in terms of recovering personalized shape details, we introduce a new evaluation metric named Mean Signal-to-Noise Ratio (MSNR) to measure the signal-to-noise ratio of each mesh frequency component. Extensive experiments demonstrate that our approach generates fine-grained details for high fidelity 3D hand reconstruction, and our evaluation metric is more effective for measuring mesh details compared with traditional metrics.

In this paper, we investigate an open research task of generating controllable 3D textured shapes from the given textual descriptions. Previous works either require ground truth caption labeling or extensive optimization time. To resolve these issues, we present a novel framework, TAPS3D, to train a text-guided 3D shape generator with pseudo captions. Specifically, based on rendered 2D images, we retrieve relevant words from the CLIP vocabulary and construct pseudo captions using templates. Our constructed captions provide high-level semantic supervision for generated 3D shapes. Further, in order to produce fine-grained textures and increase geometry diversity, we propose to adopt low-level image regularization to enable fake-rendered images to align with the real ones. During the inference phase, our proposed model can generate 3D textured shapes from the given text without any additional optimization. We conduct extensive experiments to analyze each of our proposed components and show the efficacy of our framework in generating high-fidelity 3D textured and text-relevant shapes.

Reverse engineering CAD models from raw geometry is a classic but strenuous research problem. Previous learning-based methods rely heavily on labels due to the supervised design patterns or reconstruct CAD shapes that are not easily editable. In this work, we introduce SECAD-Net, an end-to-end neural network aimed at reconstructing compact and easy-to-edit CAD models in a self-supervised manner. Drawing inspiration from the modeling language that is most commonly used in modern CAD software, we propose to learn 2D sketches and 3D extrusion parameters from raw shapes, from which a set of extrusion cylinders can be generated by extruding each sketch from a 2D plane into a 3D body. By incorporating the Boolean operation (i.e., union), these cylinders can be combined to closely approximate the target geometry. We advocate the use of implicit fields for sketch representation, which allows for creating CAD variations by interpolating latent codes in the sketch latent space. Extensive experiments on both ABC and Fusion 360 datasets demonstrate the effectiveness of our method, and show superiority over state-of-the-art alternatives including the closely related method for supervised CAD reconstruction. We further apply our approach to CAD editing and single-view CAD reconstruction. The code is released at https://github.com/BunnySoCrazy/SECAD-Net.

Cartoonization is a task that renders natural photos into cartoon styles. Previous deep methods only have focused on end-to-end translation, disabling artists from manipulating results. To tackle this, in this work, we propose a novel solution with editing features of texture and color based on the cartoon creation process. To do that, we design a model architecture to have separate decoders, texture and color, to decouple these attributes. In the texture decoder, we propose a texture controller, which enables a user to control stroke style and abstraction to generate diverse cartoon textures. We also introduce an HSV color augmentation to induce the networks to generate consistent color translation. To the best of our knowledge, our work is the first method to control the cartoonization during the inferences step, generating high-quality results compared to baselines.

We propose a novel method for high-quality facial texture reconstruction from RGB images using a novel capturing routine based on a single smartphone which we equip with an inexpensive polarization foil. Specifically, we turn the flashlight into a polarized light source and add a polarization filter on top of the camera. Leveraging this setup, we capture the face of a subject with cross-polarized and parallel-polarized light. For each subject, we record two short sequences in a dark environment under flash illumination with different light polarization using the modified smartphone. Based on these observations, we reconstruct an explicit surface mesh of the face using structure from motion. We then exploit the camera and light co-location within a differentiable renderer to optimize the facial textures using an analysis-by-synthesis approach. Our method optimizes for high-resolution normal textures, diffuse albedo, and specular albedo using a coarse-to-fine optimization scheme. We show that the optimized textures can be used in a standard rendering pipeline to synthesize high-quality photo-realistic 3D digital humans in novel environments.

We present GlassesGAN, a novel image editing framework for custom design of glasses, that sets a new standard in terms of output-image quality, edit realism, and continuous multi-style edit capability. To facilitate the editing process with GlassesGAN, we propose a Targeted Subspace Modelling (TSM) procedure that, based on a novel mechanism for (synthetic) appearance discovery in the latent space of a pre-trained GAN generator, constructs an eyeglasses-specific (latent) subspace that the editing framework can utilize. Additionally, we also introduce an appearance-constrained subspace initialization (SI) technique that centers the latent representation of the given input image in the well-defined part of the constructed subspace to improve the reliability of the learned edits. We test GlassesGAN on two (diverse) high-resolution datasets (CelebA-HQ and SiblingsDB-HQf) and compare it to three state-of-the-art baselines, i.e., InterfaceGAN, GANSpace, and MaskGAN. The reported results show that GlassesGAN convincingly outperforms all competing techniques, while offering functionality (e.g., fine-grained multi-style editing) not available with any of the competitors. The source code for GlassesGAN is made publicly available.

Neural networks for facial landmark detection are notoriously limited to a fixed set of landmarks in a dedicated layout, which must be specified at training time. Dedicated datasets must also be hand-annotated with the corresponding landmark configuration for training. We propose the first facial landmark detection network that can predict continuous, unlimited landmarks, allowing to specify the number and location of the desired landmarks at inference time. Our method combines a simple image feature extractor with a queried landmark predictor, and the user can specify any continuous query points relative to a 3D template face mesh as input. As it is not tied to a fixed set of landmarks, our method is able to leverage all pre-existing 2D landmark datasets for training, even if they have inconsistent landmark configurations. As a result, we present a very powerful facial landmark detector that can be trained once, and can be used readily for numerous applications like 3D face reconstruction, arbitrary face segmentation, and is even compatible with helmeted mounted cameras, and therefore could vastly simplify face tracking workflows for media and entertainment applications.

Realistic face rendering from multi-view images is beneficial to various computer vision and graphics applications. Due to the complex spatially-varying reflectance properties and geometry characteristics of faces, however, it remains challenging to recover 3D facial representations both faithfully and efficiently in the current studies. This paper presents a novel 3D face rendering model, namely NeuFace, to learn accurate and physically-meaningful underlying 3D representations by neural rendering techniques. It naturally incorporates the neural BRDFs into physically based rendering, capturing sophisticated facial geometry and appearance clues in a collaborative manner. Specifically, we introduce an approximated BRDF integration and a simple yet new low-rank prior, which effectively lower the ambiguities and boost the performance of the facial BRDFs. Extensive experiments demonstrate the superiority of NeuFace in human face rendering, along with a decent generalization ability to common objects. Code is released at https://github.com/aejion/NeuFace.

In this work, we present a multimodal solution to the problem of 4D face reconstruction from monocular videos. 3D face reconstruction from 2D images is an under-constrained problem due to the ambiguity of depth. State-of-the-art methods try to solve this problem by leveraging visual information from a single image or video, whereas 3D mesh animation approaches rely more on audio. However, in most cases (e.g. AR/VR applications), videos include both visual and speech information. We propose AVFace that incorporates both modalities and accurately reconstructs the 4D facial and lip motion of any speaker, without requiring any 3D ground truth for training. A coarse stage estimates the per-frame parameters of a 3D morphable model, followed by a lip refinement, and then a fine stage recovers facial geometric details. Due to the temporal audio and video information captured by transformer-based modules, our method is robust in cases when either modality is insufficient (e.g. face occlusions). Extensive qualitative and quantitative evaluation demonstrates the superiority of our method over the current state-of-the-art.

We propose a method to learn a high-quality implicit 3D head avatar from a monocular RGB video captured in the wild. The learnt avatar is driven by a parametric face model to achieve user-controlled facial expressions and head poses. Our hybrid pipeline combines the geometry prior and dynamic tracking of a 3DMM with a neural radiance field to achieve fine-grained control and photorealism. To reduce over-smoothing and improve out-of-model expressions synthesis, we propose to predict local features anchored on the 3DMM geometry. These learnt features are driven by 3DMM deformation and interpolated in 3D space to yield the volumetric radiance at a designated query point. We further show that using a Convolutional Neural Network in the UV space is critical in incorporating spatial context and producing representative local features. Extensive experiments show that we are able to reconstruct high-quality avatars, with more accurate expression-dependent details, good generalization to out-of-training expressions, and quantitatively superior renderings compared to other state-of-the-art approaches.

Controllability, generalizability and efficiency are the major objectives of constructing face avatars represented by neural implicit field. However, existing methods have not managed to accommodate the three requirements simultaneously. They either focus on static portraits, restricting the representation ability to a specific subject, or suffer from substantial computational cost, limiting their flexibility. In this paper, we propose One-shot Talking face Avatar (OTAvatar), which constructs face avatars by a generalized controllable tri-plane rendering solution so that each personalized avatar can be constructed from only one portrait as the reference. Specifically, OTAvatar first inverts a portrait image to a motion-free identity code. Second, the identity code and a motion code are utilized to modulate an efficient CNN to generate a tri-plane formulated volume, which encodes the subject in the desired motion. Finally, volume rendering is employed to generate an image in any view. The core of our solution is a novel decoupling-by-inverting strategy that disentangles identity and motion in the latent code via optimization-based inversion. Benefiting from the efficient tri-plane representation, we achieve controllable rendering of generalized face avatar at 35 FPS on A100. Experiments show promising performance of cross-identity reenactment on subjects out of the training set and better 3D consistency. The code is available at https://github.com/theEricMa/OTAvatar.

We present X-Avatar, a novel avatar model that captures the full expressiveness of digital humans to bring about life-like experiences in telepresence, AR/VR and beyond. Our method models bodies, hands, facial expressions and appearance in a holistic fashion and can be learned from either full 3D scans or RGB-D data. To achieve this, we propose a part-aware learned forward skinning module that can be driven by the parameter space of SMPL-X, allowing for expressive animation of X-Avatars. To efficiently learn the neural shape and deformation fields, we propose novel part-aware sampling and initialization strategies. This leads to higher fidelity results, especially for smaller body parts while maintaining efficient training despite increased number of articulated bones. To capture the appearance of the avatar with high-frequency details, we extend the geometry and deformation fields with a texture network that is conditioned on pose, facial expression, geometry and the normals of the deformed surface. We show experimentally that our method outperforms strong baselines both quantitatively and qualitatively on the animation task. To facilitate future research on expressive avatars we contribute a new dataset, called X-Humans, containing 233 sequences of high-quality textured scans from 20 participants, totalling 35,500 data frames.

In this paper, we take one step further towards real-world applicability of monocular neural avatar reconstruction by contributing InstantAvatar, a system that can reconstruct human avatars from a monocular video within seconds, and these avatars can be animated and rendered at an interactive rate. To achieve this efficiency we propose a carefully designed and engineered system, that leverages emerging acceleration structures for neural fields, in combination with an efficient empty-space skipping strategy for dynamic scenes. We also contribute an efficient implementation that we will make available for research purposes. Compared to existing methods, InstantAvatar converges 130x faster and can be trained in minutes instead of hours. It achieves comparable or even better reconstruction quality and novel pose synthesis results. When given the same time budget, our method significantly outperforms SoTA methods. InstantAvatar can yield acceptable visual quality in as little as 10 seconds training time. For code and more demo results, please refer to https://ait.ethz.ch/InstantAvatar.

This paper presents JAWS, an optimzation-driven approach that achieves the robust transfer of visual cinematic features from a reference in-the-wild video clip to a newly generated clip. To this end, we rely on an implicit-neural-representation (INR) in a way to compute a clip that shares the same cinematic features as the reference clip. We propose a general formulation of a camera optimization problem in an INR that computes extrinsic and intrinsic camera parameters as well as timing. By leveraging the differentiability of neural representations, we can back-propagate our designed cinematic losses measured on proxy estimators through a NeRF network to the proposed cinematic parameters directly. We also introduce specific enhancements such as guidance maps to improve the overall quality and efficiency. Results display the capacity of our system to replicate well known camera sequences from movies, adapting the framing, camera parameters and timing of the generated video clip to maximize the similarity with the reference clip.

Animating virtual avatars with free-view control is crucial for various applications like virtual reality and digital entertainment. Previous studies have attempted to utilize the representation power of the neural radiance field (NeRF) to reconstruct the human body from monocular videos. Recent works propose to graft a deformation network into the NeRF to further model the dynamics of the human neural field for animating vivid human motions. However, such pipelines either rely on pose-dependent representations or fall short of motion coherency due to frame-independent optimization, making it difficult to generalize to unseen pose sequences realistically. In this paper, we propose a novel framework MonoHuman, which robustly renders view-consistent and high-fidelity avatars under arbitrary novel poses. Our key insight is to model the deformation field with bi-directional constraints and explicitly leverage the off-the-peg keyframe information to reason the feature correlations for coherent results. Specifically, we first propose a Shared Bidirectional Deformation module, which creates a pose-independent generalizable deformation field by disentangling backward and forward deformation correspondences into shared skeletal motion weight and separate non-rigid motions. Then, we devise a Forward Correspondence Search module, which queries the correspondence feature of keyframes to guide the rendering network. The rendered results are thus multi-view consistent with high fidelity, even under challenging novel pose settings. Extensive experiments demonstrate the superiority of our proposed MonoHuman over state-of-the-art methods.

We introduce Structured 3D Features, a model based on a novel implicit 3D representation that pools pixel-aligned image features onto dense 3D points sampled from a parametric, statistical human mesh surface. The 3D points have associated semantics and can move freely in 3D space. This allows for optimal coverage of the person of interest, beyond just the body shape, which in turn, additionally helps modeling accessories, hair, and loose clothing. Owing to this, we present a complete 3D transformer-based attention framework which, given a single image of a person in an unconstrained pose, generates an animatable 3D reconstruction with albedo and illumination decomposition, as a result of a single end-to-end model, trained semi-supervised, and with no additional postprocessing. We show that our S3F model surpasses the previous state-of-the-art on various tasks, including monocular 3D reconstruction, as well as albedo & shading estimation. Moreover, we show that the proposed methodology allows novel view synthesis, relighting, and re-posing the reconstruction, and can naturally be extended to handle multiple input images (e.g. different views of a person, or the same view, in different poses, in video). Finally, we demonstrate the editing capabilities of our model for 3D virtual try-on applications.

We propose a method that leverages graph neural networks, multi-level message passing, and unsupervised training to enable real-time prediction of realistic clothing dynamics. Whereas existing methods based on linear blend skinning must be trained for specific garments, our method is agnostic to body shape and applies to tight-fitting garments as well as loose, free-flowing clothing. Our method furthermore handles changes in topology (e.g., garments with buttons or zippers) and material properties at inference time. As one key contribution, we propose a hierarchical message-passing scheme that efficiently propagates stiff stretching modes while preserving local detail. We empirically show that our method outperforms strong baselines quantitatively and that its results are perceived as more realistic than state-of-the-art methods.

Recent works have proposed to craft adversarial clothes for evading person detectors, while they are either only effective at limited viewing angles or very conspicuous to humans. We aim to craft adversarial texture for clothes based on 3D modeling, an idea that has been used to craft rigid adversarial objects such as a 3D-printed turtle. Unlike rigid objects, humans and clothes are non-rigid, leading to difficulties in physical realization. In order to craft natural-looking adversarial clothes that can evade person detectors at multiple viewing angles, we propose adversarial camouflage textures (AdvCaT) that resemble one kind of the typical textures of daily clothes, camouflage textures. We leverage the Voronoi diagram and Gumbel-softmax trick to parameterize the camouflage textures and optimize the parameters via 3D modeling. Moreover, we propose an efficient augmentation pipeline on 3D meshes combining topologically plausible projection (TopoProj) and Thin Plate Spline (TPS) to narrow the gap between digital and real-world objects. We printed the developed 3D texture pieces on fabric materials and tailored them into T-shirts and trousers. Experiments show high attack success rates of these clothes against multiple detectors.

3D human body representation learning has received increasing attention in recent years. However, existing works cannot flexibly, controllably and accurately represent human bodies, limited by coarse semantics and unsatisfactory representation capability, particularly in the absence of supervised data. In this paper, we propose a human body representation with fine-grained semantics and high reconstruction-accuracy in an unsupervised setting. Specifically, we establish a correspondence between latent vectors and geometric measures of body parts by designing a part-aware skeleton-separated decoupling strategy, which facilitates controllable editing of human bodies by modifying the corresponding latent codes. With the help of a bone-guided auto-encoder and an orientation-adaptive weighting strategy, our representation can be trained in an unsupervised manner. With the geometrically meaningful latent space, it can be applied to a wide range of applications, from human body editing to latent code interpolation and shape style transfer. Experimental results on public datasets demonstrate the accurate reconstruction and flexible editing abilities of the proposed method. The code will be available at http://cic.tju.edu.cn/faculty/likun/projects/SemanticHuman.

Building animatable 3D models is challenging due to the need for 3D scans, laborious registration, and manual rigging. Recently, differentiable rendering provides a pathway to obtain high-quality 3D models from monocular videos, but these are limited to rigid categories or single instances. We present RAC, a method to build category-level 3D models from monocular videos, disentangling variations over instances and motion over time. Three key ideas are introduced to solve this problem: (1) specializing a category-level skeleton to instances, (2) a method for latent space regularization that encourages shared structure across a category while maintaining instance details, and (3) using 3D background models to disentangle objects from the background. We build 3D models for humans, cats, and dogs given monocular videos. Project page: gengshan-y.github.io/rac-www/

We present Deformable mesh transFormer (DeFormer), a novel vertex-based approach to monocular 3D human mesh recovery. DeFormer iteratively fits a body mesh model to an input image via a mesh alignment feedback loop formed within a transformer decoder that is equipped with efficient body mesh driven attention modules: 1) body sparse self-attention and 2) deformable mesh cross attention. As a result, DeFormer can effectively exploit high-resolution image feature maps and a dense mesh model which were computationally expensive to deal with in previous approaches using the standard transformer attention. Experimental results show that DeFormer achieves state-of-the-art performances on the Human3.6M and 3DPW benchmarks. Ablation study is also conducted to show the effectiveness of the DeFormer model designs for leveraging multi-scale feature maps. Code is available at https://github.com/yusukey03012/DeFormer.

We propose Hi4D, a method and dataset for the auto analysis of physically close human-human interaction under prolonged contact. Robustly disentangling several in-contact subjects is a challenging task due to occlusions and complex shapes. Hence, existing multi-view systems typically fuse 3D surfaces of close subjects into a single, connected mesh. To address this issue we leverage i) individually fitted neural implicit avatars; ii) an alternating optimization scheme that refines pose and surface through periods of close proximity; and iii) thus segment the fused raw scans into individual instances. From these instances we compile Hi4D dataset of 4D textured scans of 20 subject pairs, 100 sequences, and a total of more than 11K frames. Hi4D contains rich interaction-centric annotations in 2D and 3D alongside accurately registered parametric body models. We define varied human pose and shape estimation tasks on this dataset and provide results from state-of-the-art methods on these benchmarks.

Despite recent achievements, existing 3D interacting hands recovery methods have shown results mainly on motion capture (MoCap) environments, not on in-the-wild (ITW) ones. This is because collecting 3D interacting hands data in the wild is extremely challenging, even for the 2D data. We present InterWild, which brings MoCap and ITW samples to shared domains for robust 3D interacting hands recovery in the wild with a limited amount of ITW 2D/3D interacting hands data. 3D interacting hands recovery consists of two sub-problems: 1) 3D recovery of each hand and 2) 3D relative translation recovery between two hands. For the first sub-problem, we bring MoCap and ITW samples to a shared 2D scale space. Although ITW datasets provide a limited amount of 2D/3D interacting hands, they contain large-scale 2D single hand data. Motivated by this, we use a single hand image as an input for the first sub-problem regardless of whether two hands are interacting. Hence, interacting hands of MoCap datasets are brought to the 2D scale space of single hands of ITW datasets. For the second sub-problem, we bring MoCap and ITW samples to a shared appearance-invariant space. Unlike the first sub-problem, 2D labels of ITW datasets are not helpful for the second sub-problem due to the 3D translation’s ambiguity. Hence, instead of relying on ITW samples, we amplify the generalizability of MoCap samples by taking only a geometric feature without an image as an input for the second sub-problem. As the geometric feature is invariant to appearances, MoCap and ITW samples do not suffer from a huge appearance gap between the two datasets. The code is available in https://github.com/facebookresearch/InterWild.

We present a novel method for recovering the absolute pose and shape of a human in a pre-scanned scene given a single image. Unlike previous methods that perform sceneaware mesh optimization, we propose to first estimate absolute position and dense scene contacts with a sparse 3D CNN, and later enhance a pretrained human mesh recovery network by cross-attention with the derived 3D scene cues. Joint learning on images and scene geometry enables our method to reduce the ambiguity caused by depth and occlusion, resulting in more reasonable global postures and contacts. Encoding scene-aware cues in the network also allows the proposed method to be optimization-free, and opens up the opportunity for real-time applications. The experiments show that the proposed network is capable of recovering accurate and physically-plausible meshes by a single forward pass and outperforms state-of-the-art methods in terms of both accuracy and speed. Code is available on our project page: https://zju3dv.github.io/sahmr/.

Real-time 3D hand mesh reconstruction is challenging, especially when the hand is holding some object. Beyond the previous methods, we design H2ONet to fully exploit non-occluded information from multiple frames to boost the reconstruction quality. First, we decouple hand mesh reconstruction into two branches, one to exploit finger-level non-occluded information and the other to exploit global hand orientation, with lightweight structures to promote real-time inference. Second, we propose finger-level occlusion-aware feature fusion, leveraging predicted finger-level occlusion information as guidance to fuse finger-level information across time frames. Further, we design hand-level occlusion-aware feature fusion to fetch non-occluded information from nearby time frames. We conduct experiments on the Dex-YCB and HO3D-v2 datasets with challenging hand-object occlusion cases, manifesting that H2ONet is able to run in real-time and achieves state-of-the-art performance on both the hand mesh and pose precision. The code will be released on GitHub.

3D reconstruction is a fundamental problem in computer vision, and the task is especially challenging when the object to reconstruct is partially or fully occluded. We introduce a method that uses the shadows cast by an unobserved object in order to infer the possible 3D volumes under occlusion. We create a differentiable image formation model that allows us to jointly infer the 3D shape of an object, its pose, and the position of a light source. Since the approach is end-to-end differentiable, we are able to integrate learned priors of object geometry in order to generate realistic 3D shapes of different object categories. Experiments and visualizations show that the method is able to generate multiple possible solutions that are consistent with the observation of the shadow. Our approach works even when the position of the light source and object pose are both unknown. Our approach is also robust to real-world images where ground-truth shadow mask is unknown.

In comparison with most methods focusing on 3D rigid object recognition and manipulation, deformable objects are more common in our real life but attract less attention. Generally, most existing methods for deformable object manipulation suffer two issues, 1) Massive demonstration: repeating thousands of robot-object demonstrations for model training of one specific instance; 2) Poor generalization: inevitably re-training for transferring the learned skill to a similar/new instance from the same category. Therefore, we propose a category-level deformable 3D object manipulation framework, which could manipulate deformable 3D objects with only one demonstration and generalize the learned skills to new similar instances without re-training. Specifically, our proposed framework consists of two modules. The Nocs State Transform (NST) module transfers the observed point clouds of the target to a pre-defined unified pose state (i.e., Nocs state), which is the foundation for the category-level manipulation learning; the Neural Spatial Encoding (NSE) module generalizes the learned skill to novel instances by encoding the category-level spatial information to pursue the expected grasping point without re-training. The relative motion path is then planned to achieve autonomous manipulation. Both the simulated results via our Cap40 dataset and real robotic experiments justify the effectiveness of our framework.

We propose a method for in-hand 3D scanning of an unknown object with a monocular camera. Our method relies on a neural implicit surface representation that captures both the geometry and the appearance of the object, however, by contrast with most NeRF-based methods, we do not assume that the camera-object relative poses are known. Instead, we simultaneously optimize both the object shape and the pose trajectory. As direct optimization over all shape and pose parameters is prone to fail without coarse-level initialization, we propose an incremental approach that starts by splitting the sequence into carefully selected overlapping segments within which the optimization is likely to succeed. We reconstruct the object shape and track its poses independently within each segment, then merge all the segments before performing a global optimization. We show that our method is able to reconstruct the shape and color of both textured and challenging texture-less objects, outperforms classical methods that rely only on appearance features, and that its performance is close to recent methods that assume known camera poses.

We study the problem of inferring scene affordances by presenting a method for realistically inserting people into scenes. Given a scene image with a marked region and an image of a person, we insert the person into the scene while respecting the scene affordances. Our model can infer the set of realistic poses given the scene context, re-pose the reference person, and harmonize the composition. We set up the task in a self-supervised fashion by learning to re- pose humans in video clips. We train a large-scale diffusion model on a dataset of 2.4M video clips that produces diverse plausible poses while respecting the scene context. Given the learned human-scene composition, our model can also hallucinate realistic people and scenes when prompted without conditioning and also enables interactive editing. We conduct quantitative evaluation and show that our method synthesizes more realistic human appearance and more natural human-scene interactions when compared to prior work.

Humans constantly contact objects to move and perform tasks. Thus, detecting human-object contact is important for building human-centered artificial intelligence. However, there exists no robust method to detect contact between the body and the scene from an image, and there exists no dataset to learn such a detector. We fill this gap with HOT (“Human-Object conTact”), a new dataset of human-object contacts in images. To build HOT, we use two data sources: (1) We use the PROX dataset of 3D human meshes moving in 3D scenes, and automatically annotate 2D image areas for contact via 3D mesh proximity and projection. (2) We use the V-COCO, HAKE and Watch-n-Patch datasets, and ask trained annotators to draw polygons around the 2D image areas where contact takes place. We also annotate the involved body part of the human body. We use our HOT dataset to train a new contact detector, which takes a single color image as input, and outputs 2D contact heatmaps as well as the body-part labels that are in contact. This is a new and challenging task, that extends current foot-ground or hand-object contact detectors to the full generality of the whole body. The detector uses a part-attention branch to guide contact estimation through the context of the surrounding body parts and scene. We evaluate our detector extensively, and quantitative results show that our model outperforms baselines, and that all components contribute to better performance. Results on images from an online repository show reasonable detections and generalizability. Our HOT data and model are available for research at https://hot.is.tue.mpg.de.

Inferring past human motion from RGB images is challenging due to the inherent uncertainty of the prediction problem. Thermal images, on the other hand, encode traces of past human-object interactions left in the environment via thermal radiation measurement. Based on this observation, we collect the first RGB-Thermal dataset for human motion analysis, dubbed Thermal-IM. Then we develop a three-stage neural network model for accurate past human pose estimation. Comprehensive experiments show that thermal cues significantly reduce the ambiguities of this task, and the proposed model achieves remarkable performance. The dataset is available at https://github.com/ZitianTang/Thermal-IM.

Multi-person pose forecasting remains a challenging problem, especially in modeling fine-grained human body interaction in complex crowd scenarios. Existing methods typically represent the whole pose sequence as a temporal series, yet overlook interactive influences among people based on skeletal body parts. In this paper, we propose a novel Trajectory-Aware Body Interaction Transformer (TBIFormer) for multi-person pose forecasting via effectively modeling body part interactions. Specifically, we construct a Temporal Body Partition Module that transforms all the pose sequences into a Multi-Person Body-Part sequence to retain spatial and temporal information based on body semantics. Then, we devise a Social Body Interaction Self-Attention (SBI-MSA) module, utilizing the transformed sequence to learn body part dynamics for inter- and intra-individual interactions. Furthermore, different from prior Euclidean distance-based spatial encodings, we present a novel and efficient Trajectory-Aware Relative Position Encoding for SBI-MSA to offer discriminative spatial information and additional interactive clues. On both short- and long-term horizons, we empirically evaluate our framework on CMU-Mocap, MuPoTS-3D as well as synthesized datasets (6 ~ 10 persons), and demonstrate that our method greatly outperforms the state-of-the-art methods.

Temporal modeling is crucial for multi-frame human pose estimation. Most existing methods directly employ optical flow or deformable convolution to predict full-spectrum motion fields, which might incur numerous irrelevant cues, such as a nearby person or background. Without further efforts to excavate meaningful motion priors, their results are suboptimal, especially in complicated spatiotemporal interactions. On the other hand, the temporal difference has the ability to encode representative motion information which can potentially be valuable for pose estimation but has not been fully exploited. In this paper, we present a novel multi-frame human pose estimation framework, which employs temporal differences across frames to model dynamic contexts and engages mutual information objectively to facilitate useful motion information disentanglement. To be specific, we design a multi-stage Temporal Difference Encoder that performs incremental cascaded learning conditioned on multi-stage feature difference sequences to derive informative motion representation. We further propose a Representation Disentanglement module from the mutual information perspective, which can grasp discriminative task-relevant motion signals by explicitly defining useful and noisy constituents of the raw motion features and minimizing their mutual information. These place us to rank No.1 in the Crowd Pose Estimation in Complex Events Challenge on benchmark dataset HiEve, and achieve state-of-the-art performance on three benchmarks PoseTrack2017, PoseTrack2018, and PoseTrack21.

Estimating 3D human motion from an egocentric video sequence plays a critical role in human behavior understanding and has various applications in VR/AR. However, naively learning a mapping between egocentric videos and human motions is challenging, because the user’s body is often unobserved by the front-facing camera placed on the head of the user. In addition, collecting large-scale, high-quality datasets with paired egocentric videos and 3D human motions requires accurate motion capture devices, which often limit the variety of scenes in the videos to lab-like environments. To eliminate the need for paired egocentric video and human motions, we propose a new method, Ego-Body Pose Estimation via Ego-Head Pose Estimation (EgoEgo), which decomposes the problem into two stages, connected by the head motion as an intermediate representation. EgoEgo first integrates SLAM and a learning approach to estimate accurate head motion. Subsequently, leveraging the estimated head pose as input, EgoEgo utilizes conditional diffusion to generate multiple plausible full-body motions. This disentanglement of head and body pose eliminates the need for training datasets with paired egocentric videos and 3D human motion, enabling us to leverage large-scale egocentric video datasets and motion capture datasets separately. Moreover, for systematic benchmarking, we develop a synthetic dataset, AMASS-Replica-Ego-Syn (ARES), with paired egocentric videos and human motion. On both ARES and real data, our EgoEgo model performs significantly better than the current state-of-the-art methods.

Human-Object Interaction (HOI) detection, which localizes and infers relationships between human and objects, plays an important role in scene understanding. Although two-stage HOI detectors have advantages of high efficiency in training and inference, they suffer from lower performance than one-stage methods due to the old backbone networks and the lack of considerations for the HOI perception process of humans in the interaction classifiers. In this paper, we propose Vision Transformer based Pose-Conditioned Self-Loop Graph (ViPLO) to resolve these problems. First, we propose a novel feature extraction method suitable for the Vision Transformer backbone, called masking with overlapped area (MOA) module. The MOA module utilizes the overlapped area between each patch and the given region in the attention function, which addresses the quantization problem when using the Vision Transformer backbone. In addition, we design a graph with a pose-conditioned self-loop structure, which updates the human node encoding with local features of human joints. This allows the classifier to focus on specific human joints to effectively identify the type of interaction, which is motivated by the human perception process for HOI. As a result, ViPLO achieves the state-of-the-art results on two public benchmarks, especially obtaining a +2.07 mAP performance gain on the HICO-DET dataset.

In this paper, we focus on the problem of category-level object pose estimation, which is challenging due to the large intra-category shape variation. 3D graph convolution (3D-GC) based methods have been widely used to extract local geometric features, but they have limitations for complex shaped objects and are sensitive to noise. Moreover, the scale and translation invariant properties of 3D-GC restrict the perception of an object’s size and translation information. In this paper, we propose a simple network structure, the HS-layer, which extends 3D-GC to extract hybrid scope latent features from point cloud data for category-level object pose estimation tasks. The proposed HS-layer: 1) is able to perceive local-global geometric structure and global information, 2) is robust to noise, and 3) can encode size and translation information. Our experiments show that the simple replacement of the 3D-GC layer with the proposed HS-layer on the baseline method (GPV-Pose) achieves a significant improvement, with the performance increased by 14.5% on 5d2cm metric and 10.3% on IoU75. Our method outperforms the state-of-the-art methods by a large margin (8.3% on 5d2cm, 6.9% on IoU75) on REAL275 dataset and runs in real-time (50 FPS).

Animal pose estimation is an important but under-explored task due to the lack of labeled data. In this paper, we tackle the task of animal pose estimation with scarce annotations, where only a small set of labeled data and unlabeled images are available. At the core of the solution to this problem setting is the use of the unlabeled data to compensate for the lack of well-labeled animal pose data. To this end, we propose the ScarceNet, a pseudo label-based approach to generate artificial labels for the unlabeled images. The pseudo labels, which are generated with a model trained with the small set of labeled images, are generally noisy and can hurt the performance when directly used for training. To solve this problem, we first use a small-loss trick to select reliable pseudo labels. Although effective, the selection process is improvident since numerous high-loss samples are left unused. We further propose to identify reusable samples from the high-loss samples based on an agreement check. Pseudo labels are re-generated to provide supervision for those reusable samples. Lastly, we introduce a student-teacher framework to enforce a consistency constraint since there are still samples that are neither reliable nor reusable. By combining the reliable pseudo label selection with the reusable sample re-labeling and the consistency constraint, we can make full use of the unlabeled data. We evaluate our approach on the challenging AP-10K dataset, where our approach outperforms existing semi-supervised approaches by a large margin. We also test on the TigDog dataset, where our approach can achieve better performance than domain adaptation based approaches when only very few annotations are available. Our code is available at the project website.

Recent advances in hand pose estimation have shed light on utilizing synthetic data to train neural networks, which however inevitably hinders generalization to real-world data due to domain gaps. To solve this problem, we present a framework for cross-domain semi-supervised hand pose estimation and target the challenging scenario of learning models from labelled multi-modal synthetic data and unlabelled real-world data. To that end, we propose a dual-modality network that exploits synthetic RGB and synthetic depth images. For pre-training, our network uses multi-modal contrastive learning and attention-fused supervision to learn effective representations of the RGB images. We then integrate a novel self-distillation technique during fine-tuning to reduce pseudo-label noise. Experiments show that the proposed method significantly improves 3D hand pose estimation and 2D keypoint detection on benchmarks.

In this paper, a novel virtual try-on algorithm, dubbed SAL-VTON, is proposed, which links the garment with the person via semantically associated landmarks to alleviate misalignment. The semantically associated landmarks are a series of landmark pairs with the same local semantics on the in-shop garment image and the try-on image. Based on the semantically associated landmarks, SAL-VTON effectively models the local semantic association between garment and person, making up for the misalignment in the overall deformation of the garment. The outcome is achieved with a three-stage framework: 1) the semantically associated landmarks are estimated using the landmark localization model; 2) taking the landmarks as input, the warping model explicitly associates the corresponding parts of the garment and person for obtaining the local flow, thus refining the alignment in the global flow; 3) finally, a generator consumes the landmarks to better capture local semantics and control the try-on results.Moreover, we propose a new landmark dataset with a unified labelling rule of landmarks for diverse styles of garments. Extensive experimental results on popular datasets demonstrate that SAL-VTON can handle misalignment and outperform state-of-the-art methods both qualitatively and quantitatively. The dataset is available on https://modelscope.cn/datasets/damo/SAL-HG/summary.

This paper presents a neural incremental Structure-from-Motion (SfM) approach, Level-S2fM, which estimates the camera poses and scene geometry from a set of uncalibrated images by learning coordinate MLPs for the implicit surfaces and the radiance fields from the established keypoint correspondences. Our novel formulation poses some new challenges due to inevitable two-view and few-view configurations in the incremental SfM pipeline, which complicates the optimization of coordinate MLPs for volumetric neural rendering with unknown camera poses. Nevertheless, we demonstrate that the strong inductive basis conveying in the 2D correspondences is promising to tackle those challenges by exploiting the relationship between the ray sampling schemes. Based on this, we revisit the pipeline of incremental SfM and renew the key components, including two-view geometry initialization, the camera poses registration, the 3D points triangulation, and Bundle Adjustment, with a fresh perspective based on neural implicit surfaces. By unifying the scene geometry in small MLP networks through coordinate MLPs, our Level-S2fM treats the zero-level set of the implicit surface as an informative top-down regularization to manage the reconstructed 3D points, reject the outliers in correspondences via querying SDF, and refine the estimated geometries by NBA (Neural BA). Not only does our Level-S2fM lead to promising results on camera pose estimation and scene geometry reconstruction, but it also shows a promising way for neural implicit rendering without knowing camera extrinsic beforehand.

In this paper, we revisit the rotation averaging problem applied in global Structure-from-Motion pipelines. We argue that the main problem of current methods is the minimized cost function that is only weakly connected with the input data via the estimated epipolar geometries. We propose to better model the underlying noise distributions by directly propagating the uncertainty from the point correspondences into the rotation averaging. Such uncertainties are obtained for free by considering the Jacobians of two-view refinements. Moreover, we explore integrating a variant of the MAGSAC loss into the rotation averaging problem, instead of using classical robust losses employed in current frameworks. The proposed method leads to results superior to baselines, in terms of accuracy, on large-scale public benchmarks. The code is public. https://github.com/zhangganlin/GlobalSfMpy

This work addresses cross-view camera pose estimation, i.e., determining the 3-Degrees-of-Freedom camera pose of a given ground-level image w.r.t. an aerial image of the local area. We propose SliceMatch, which consists of ground and aerial feature extractors, feature aggregators, and a pose predictor. The feature extractors extract dense features from the ground and aerial images. Given a set of candidate camera poses, the feature aggregators construct a single ground descriptor and a set of pose-dependent aerial descriptors. Notably, our novel aerial feature aggregator has a cross-view attention module for ground-view guided aerial feature selection and utilizes the geometric projection of the ground camera’s viewing frustum on the aerial image to pool features. The efficient construction of aerial descriptors is achieved using precomputed masks. SliceMatch is trained using contrastive learning and pose estimation is formulated as a similarity comparison between the ground descriptor and the aerial descriptors. Compared to the state-of-the-art, SliceMatch achieves a 19% lower median localization error on the VIGOR benchmark using the same VGG16 backbone at 150 frames per second, and a 50% lower error when using a ResNet50 backbone.

We present a new loss function for joint disparity and uncertainty estimation in deep stereo matching. Our work is motivated by the need for precise uncertainty estimates and the observation that multi-task learning often leads to improved performance in all tasks. We show that this can be achieved by requiring the distribution of uncertainty to match the distribution of disparity errors via a KL divergence term in the network’s loss function. A differentiable soft-histogramming technique is used to approximate the distributions so that they can be used in the loss. We experimentally assess the effectiveness of our approach and observe significant improvements in both disparity and uncertainty prediction on large datasets. Our code is available at https://github.com/lly00412/SEDNet.git.

Despite the remarkable advances in image matching and pose estimation, image-based localization of a camera in a temporally-varying outdoor environment is still a challenging problem due to huge appearance disparity between query and reference images caused by illumination, seasonal and structural changes. In this work, we propose to leverage additional sensors on a mobile phone, mainly GPS, compass, and gravity sensor, to solve this challenging problem. We show that these mobile sensors provide decent initial poses and effective constraints to reduce the searching space in image matching and final pose estimation. With the initial pose, we are also able to devise a direct 2D-3D matching network to efficiently establish 2D-3D correspondences instead of tedious 2D-2D matching in existing systems. As no public dataset exists for the studied problem, we collect a new dataset that provides a variety of mobile sensor data and significant scene appearance variations, and develop a system to acquire ground-truth poses for query images. We benchmark our method as well as several state-of-the-art baselines and demonstrate the effectiveness of the proposed approach. Our code and dataset are available on the project page: https://zju3dv.github.io/sensloc/.

We introduce a method that can learn to predict scene-level implicit functions for 3D reconstruction from posed RGBD data. At test time, our system maps a previously unseen RGB image to a 3D reconstruction of a scene via implicit functions. While implicit functions for 3D reconstruction have often been tied to meshes, we show that we can train one using only a set of posed RGBD images. This setting may help 3D reconstruction unlock the sea of accelerometer+RGBD data that is coming with new phones. Our system, D2-DRDF, can match and sometimes outperform current methods that use mesh supervision and shows better robustness to sparse data.

Visual localization refers to the process of recovering camera pose from input image relative to a known scene, forming a cornerstone of numerous vision and robotics systems. While many algorithms utilize sparse 3D point cloud of the scene obtained via structure-from-motion (SfM) for localization, recent studies have raised privacy concerns by successfully revealing high-fidelity appearance of the scene from such sparse 3D representation. One prominent approach for bypassing this attack was to lift 3D points to randomly oriented 3D lines thereby hiding scene geometry, but latest work have shown such random line cloud has a critical statistical flaw that can be exploited to break through protection. In this work, we present an alternative lightweight strategy called Paired-Point Lifting (PPL) for constructing 3D line clouds. Instead of drawing one randomly oriented line per 3D point, PPL splits 3D points into pairs and joins each pair to form 3D lines. This seemingly simple strategy yields 3 benefits, i) new ambiguity in feature selection, ii) increased line cloud sparsity, and iii) non-trivial distribution of 3D lines, all of which contributes to enhanced protection against privacy attacks. Extensive experimental results demonstrate the strength of PPL in concealing scene details without compromising localization accuracy, unlocking the true potential of 3D line clouds.

We introduce the ObjectFolder Benchmark, a benchmark suite of 10 tasks for multisensory object-centric learning, centered around object recognition, reconstruction, and manipulation with sight, sound, and touch. We also introduce the ObjectFolder Real dataset, including the multisensory measurements for 100 real-world household objects, building upon a newly designed pipeline for collecting the 3D meshes, videos, impact sounds, and tactile readings of real-world objects. For each task in the ObjectFolder Benchmark, we conduct systematic benchmarking on both the 1,000 multisensory neural objects from ObjectFolder, and the real multisensory data from ObjectFolder Real. Our results demonstrate the importance of multisensory perception and reveal the respective roles of vision, audio, and touch for different object-centric learning tasks. By publicly releasing our dataset and benchmark suite, we hope to catalyze and enable new research in multisensory object-centric learning in computer vision, robotics, and beyond. Project page: https://objectfolder.stanford.edu

Shape from Polarization (SfP) aims to recover surface normal using the polarization cues of light. The accuracy of existing SfP methods is affected by two main problems. First, the ambiguity of polarization cues partially results in false normal estimation. Second, the widely-used assumption about orthographic projection is too ideal. To solve these problems, we propose the first approach that combines deep learning and stereo polarization information to recover not only normal but also disparity. Specifically, for the ambiguity problem, we design a Shape Consistency-based Mask Prediction (SCMP) module. It exploits the inherent consistency between normal and disparity to identify the areas with false normal estimation. We replace the unreliable features enclosed by these areas with new features extracted by global attention mechanism. As to the orthographic projection problem, we propose a novel Viewing Direction-aided Positional Encoding (VDPE) strategy. This strategy is based on the unique pixel-viewing direction encoding, and thus enables our neural network to handle the non-orthographic projection. In addition, we establish a real-world stereo SfP dataset that contains various object categories and illumination conditions. Experiments showed that compared with existing SfP methods, our approach is more accurate. Moreover, our approach shows higher robustness to light variation.

Inferring the structure of 3D scenes from 2D observations is a fundamental challenge in computer vision. Recently popularized approaches based on neural scene representations have achieved tremendous impact and have been applied across a variety of applications. One of the major remaining challenges in this space is training a single model which can provide latent representations which effectively generalize beyond a single scene. Scene Representation Transformer (SRT) has shown promise in this direction, but scaling it to a larger set of diverse scenes is challenging and necessitates accurately posed ground truth data. To address this problem, we propose RUST (Really Unposed Scene representation Transformer), a pose-free approach to novel view synthesis trained on RGB images alone. Our main insight is that one can train a Pose Encoder that peeks at the target image and learns a latent pose embedding which is used by the decoder for view synthesis. We perform an empirical investigation into the learned latent pose structure and show that it allows meaningful test-time camera transformations and accurate explicit pose readouts. Perhaps surprisingly, RUST achieves similar quality as methods which have access to perfect camera pose, thereby unlocking the potential for large-scale training of amortized neural scene representations.

Geometric camera calibration is often required for applications that understand the perspective of the image. We propose perspective fields as a representation that models the local perspective properties of an image. Perspective Fields contain per-pixel information about the camera view, parameterized as an up vector and a latitude value. This representation has a number of advantages as it makes minimal assumptions about the camera model and is invariant or equivariant to common image editing operations like cropping, warping, and rotation. It is also more interpretable and aligned with human perception. We train a neural network to predict Perspective Fields and the predicted Perspective Fields can be converted to calibration parameters easily. We demonstrate the robustness of our approach under various scenarios compared with camera calibration-based methods and show example applications in image compositing. Project page: https://jinlinyi.github.io/PerspectiveFields/

We propose VisFusion, a visibility-aware online 3D scene reconstruction approach from posed monocular videos. In particular, we aim to reconstruct the scene from volumetric features. Unlike previous reconstruction methods which aggregate features for each voxel from input views without considering its visibility, we aim to improve the feature fusion by explicitly inferring its visibility from a similarity matrix, computed from its projected features in each image pair. Following previous works, our model is a coarse-to-fine pipeline including a volume sparsification process. Different from their works which sparsify voxels globally with a fixed occupancy threshold, we perform the sparsification on a local feature volume along each visual ray to preserve at least one voxel per ray for more fine details. The sparse local volume is then fused with a global one for online reconstruction. We further propose to predict TSDF in a coarse-to-fine manner by learning its residuals across scales leading to better TSDF predictions. Experimental results on benchmarks show that our method can achieve superior performance with more scene details. Code is available at: https://github.com/huiyu-gao/VisFusion

Line segments are ubiquitous in our human-made world and are increasingly used in vision tasks. They are complementary to feature points thanks to their spatial extent and the structural information they provide. Traditional line detectors based on the image gradient are extremely fast and accurate, but lack robustness in noisy images and challenging conditions. Their learned counterparts are more repeatable and can handle challenging images, but at the cost of a lower accuracy and a bias towards wireframe lines. We propose to combine traditional and learned approaches to get the best of both worlds: an accurate and robust line detector that can be trained in the wild without ground truth lines. Our new line segment detector, DeepLSD, processes images with a deep network to generate a line attraction field, before converting it to a surrogate image gradient magnitude and angle, which is then fed to any existing handcrafted line detector. Additionally, we propose a new optimization tool to refine line segments based on the attraction field and vanishing points. This refinement improves the accuracy of current deep detectors by a large margin. We demonstrate the performance of our method on low-level line detection metrics, as well as on several downstream tasks using multiple challenging datasets. The source code and models are available at https://github.com/cvg/DeepLSD.

Based on the Manhattan World assumption, most existing indoor layout estimation schemes focus on recovering layouts from vertically compressed 1D sequences. However, the compression procedure confuses the semantics of different planes, yielding inferior performance with ambiguous interpretability. To address this issue, we propose to disentangle this 1D representation by pre-segmenting orthogonal (vertical and horizontal) planes from a complex scene, explicitly capturing the geometric cues for indoor layout estimation. Considering the symmetry between the floor boundary and ceiling boundary, we also design a soft-flipping fusion strategy to assist the pre-segmentation. Besides, we present a feature assembling mechanism to effectively integrate shallow and deep features with distortion distribution awareness. To compensate for the potential errors in pre-segmentation, we further leverage triple attention to reconstruct the disentangled sequences for better performance. Experiments on four popular benchmarks demonstrate our superiority over existing SoTA solutions, especially on the 3DIoU metric. The code is available at https://github.com/zhijieshen-bjtu/DOPNet.

Neural-network-based single image depth prediction (SIDP) is a challenging task where the goal is to predict the scene’s per-pixel depth at test time. Since the problem, by definition, is ill-posed, the fundamental goal is to come up with an approach that can reliably model the scene depth from a set of training examples. In the pursuit of perfect depth estimation, most existing state-of-the-art learning techniques predict a single scalar depth value per-pixel. Yet, it is well-known that the trained model has accuracy limits and can predict imprecise depth. Therefore, an SIDP approach must be mindful of the expected depth variations in the model’s prediction at test time. Accordingly, we introduce an approach that performs continuous modeling of per-pixel depth, where we can predict and reason about the per-pixel depth and its distribution. To this end, we model per-pixel scene depth using a multivariate Gaussian distribution. Moreover, contrary to the existing uncertainty modeling methods---in the same spirit, where per-pixel depth is assumed to be independent, we introduce per-pixel covariance modeling that encodes its depth dependency w.r.t. all the scene points. Unfortunately, per-pixel depth covariance modeling leads to a computationally expensive continuous loss function, which we solve efficiently using the learned low-rank approximation of the overall covariance matrix. Notably, when tested on benchmark datasets such as KITTI, NYU, and SUN-RGB-D, the SIDP model obtained by optimizing our loss function shows state-of-the-art results. Our method’s accuracy (named MG) is among the top on the KITTI depth-prediction benchmark leaderboard.

Despite the proliferation of ultra wide-angle lenses on smartphone cameras, such lenses often come with severe image distortion (e.g. curved linear structure, unnaturally skewed faces). Most existing rectification methods adopt a global warping transformation to undistort the input wide-angle image, yet their performances are not entirely satisfactory, leaving many unwanted residue distortions uncorrected or at the sacrifice of the intended wide FoV (field-of-view). This paper proposes a new method to tackle these challenges. Specifically, we derive a locally-adaptive polar-domain conformal mapping to rectify a wide-angle image. Parameters of the mapping are found automatically by analyzing image contents via deep neural networks. Experiments on large number of photos have confirmed the superior performance of the proposed method compared with all available previous methods.

Traditional focal stack methods require multiple shots to capture images focused at different distances of the same scene, which cannot be applied to dynamic scenes well. Generating a high-quality all-in-focus image from a single shot is challenging, due to the highly ill-posed nature of the single-image defocus and deblurring problem. In this paper, to restore an all-in-focus image, we propose the event focal stack which is defined as event streams captured during a continuous focal sweep. Given an RGB image focused at an arbitrary distance, we explore the high temporal resolution of event streams, from which we automatically select refocusing timestamps and reconstruct corresponding refocused images with events to form a focal stack. Guided by the neighbouring events around the selected timestamps, we can merge the focal stack with proper weights and restore a sharp all-in-focus image. Experimental results on both synthetic and real datasets show superior performance over state-of-the-art methods.

Deep learning-based multi-view stereo has emerged as a powerful paradigm for reconstructing the complete geometrically-detailed objects from multi-views. Most of the existing approaches only estimate the pixel-wise depth value by minimizing the gap between the predicted point and the intersection of ray and surface, which usually ignore the surface topology. It is essential to the textureless regions and surface boundary that cannot be properly reconstructed.To address this issue, we suggest to take advantage of point-to-surface distance so that the model is able to perceive a wider range of surfaces. To this end, we predict the distance volume from cost volume to estimate the signed distance of points around the surface. Our proposed RA-MVSNet is patch-awared, since the perception range is enhanced by associating hypothetical planes with a patch of surface. Therefore, it could increase the completion of textureless regions and reduce the outliers at the boundary. Moreover, the mesh topologies with fine details can be generated by the introduced distance volume. Comparing to the conventional deep learning-based multi-view stereo methods, our proposed RA-MVSNet approach obtains more complete reconstruction results by taking advantage of signed distance supervision. The experiments on both the DTU and Tanks & Temples datasets demonstrate that our proposed approach achieves the state-of-the-art results.

In this paper, we aim to learn a semantic radiance field from multiple scenes that is accurate, efficient and generalizable. While most existing NeRFs target at the tasks of neural scene rendering, image synthesis and multi-view reconstruction, there are a few attempts such as Semantic-NeRF that explore to learn high-level semantic understanding with the NeRF structure. However, Semantic-NeRF simultaneously learns color and semantic label from a single ray with multiple heads, where the single ray fails to provide rich semantic information. As a result, Semantic NeRF relies on positional encoding and needs to train one specific model for each scene. To address this, we propose Semantic Ray (S-Ray) to fully exploit semantic information along the ray direction from its multi-view reprojections. As directly performing dense attention over multi-view reprojected rays would suffer from heavy computational cost, we design a Cross-Reprojection Attention module with consecutive intra-view radial and cross-view sparse attentions, which decomposes contextual information along reprojected rays and cross multiple views and then collects dense connections by stacking the modules. Experiments show that our S-Ray is able to learn from multiple scenes, and it presents strong generalization ability to adapt to unseen scenes.

This paper presents OmniCity, a new dataset for omnipotent city understanding from multi-level and multi-view images. More precisely, OmniCity contains multi-view satellite images as well as street-level panorama and mono-view images, constituting over 100K pixel-wise annotated images that are well-aligned and collected from 25K geo-locations in New York City. To alleviate the substantial pixel-wise annotation efforts, we propose an efficient street-view image annotation pipeline that leverages the existing label maps of satellite view and the transformation relations between different views (satellite, panorama, and mono-view). With the new OmniCity dataset, we provide benchmarks for a variety of tasks including building footprint extraction, height estimation, and building plane/instance/fine-grained segmentation. Compared with existing multi-level and multi-view benchmarks, OmniCity contains a larger number of images with richer annotation types and more views, provides more benchmark results of state-of-the-art models, and introduces a new task for fine-grained building instance segmentation on street-level panorama images. Moreover, OmniCity provides new problem settings for existing tasks, such as cross-view image matching, synthesis, segmentation, detection, etc., and facilitates the developing of new methods for large-scale city understanding, reconstruction, and simulation. The OmniCity dataset as well as the benchmarks will be released at https://city-super.github.io/omnicity/.

We present ESLAM, an efficient implicit neural representation method for Simultaneous Localization and Mapping (SLAM). ESLAM reads RGB-D frames with unknown camera poses in a sequential manner and incrementally reconstructs the scene representation while estimating the current camera position in the scene. We incorporate the latest advances in Neural Radiance Fields (NeRF) into a SLAM system, resulting in an efficient and accurate dense visual SLAM method. Our scene representation consists of multi-scale axis-aligned perpendicular feature planes and shallow decoders that, for each point in the continuous space, decode the interpolated features into Truncated Signed Distance Field (TSDF) and RGB values. Our extensive experiments on three standard datasets, Replica, ScanNet, and TUM RGB-D show that ESLAM improves the accuracy of 3D reconstruction and camera localization of state-of-the-art dense visual SLAM methods by more than 50%, while it runs up to 10 times faster and does not require any pre-training. Project page: https://www.idiap.ch/paper/eslam

Non-line-of-sight (NLOS) imaging aims at reconstructing the location, shape, albedo, and surface normal of the hidden object around the corner with measured transient data. Due to its strong potential in various fields, it has drawn much attention in recent years. However, long exposure time is not always available for applications such as auto-driving, which hinders the practical use of NLOS imaging. Although scanning fewer points can reduce the total measurement time, it also brings the problem of imaging quality degradation. This paper proposes a general learning-based pipeline for increasing imaging quality with only a few scanning points. We tailor a neural network to learn the operator that recovers a high spatial resolution signal. Experiments on synthetic and measured data indicate that the proposed method provides faithful reconstructions of the hidden scene under both confocal and non-confocal settings. Compared with original measurements, the acquisition of our approach is 16 times faster while maintaining similar reconstruction quality. Besides, the proposed pipeline can be applied directly to existing optical systems and imaging algorithms as a plug-in-and-play module. We believe the proposed pipeline is powerful in increasing the frame rate in NLOS video imaging.

Next generation cellular networks will implement radio sensing functions alongside customary communications, thereby enabling unprecedented worldwide sensing coverage outdoors. Deep learning has revolutionised computer vision but has had limited application to radio perception tasks, in part due to lack of systematic datasets and benchmarks dedicated to the study of the performance and promise of radio sensing. To address this gap, we present MaxRay: a synthetic radio-visual dataset and benchmark that facilitate precise target localisation in radio. We further propose to learn to localise targets in radio without supervision by extracting self-coordinates from radio-visual correspondence. We use such self-supervised coordinates to train a radio localiser network. We characterise our performance against a number of state-of-the-art baselines. Our results indicate that accurate radio target localisation can be automatically learned from paired radio-visual data without labels, which is important for empirical data. This opens the door for vast data scalability and may prove key to realising the promise of robust radio sensing atop a unified communication-perception cellular infrastructure. Dataset will be hosted on IEEE DataPort.

The performance of modern object detectors drops when the test distribution differs from the training one. Most of the methods that address this focus on object appearance changes caused by, e.g., different illumination conditions, or gaps between synthetic and real images. Here, by contrast, we tackle geometric shifts emerging from variations in the image capture process, or due to the constraints of the environment causing differences in the apparent geometry of the content itself. We introduce a self-training approach that learns a set of geometric transformations to minimize these shifts without leveraging any labeled data in the new domain, nor any information about the cameras. We evaluate our method on two different shifts, i.e., a camera’s field of view (FoV) change and a viewpoint change. Our results evidence that learning geometric transformations helps detectors to perform better in the target domains.

Monocular 3D lane detection is a challenging task due to its lack of depth information. A popular solution is to first transform the front-viewed (FV) images or features into the bird-eye-view (BEV) space with inverse perspective mapping (IPM) and detect lanes from BEV features. However, the reliance of IPM on flat ground assumption and loss of context information make it inaccurate to restore 3D information from BEV representations. An attempt has been made to get rid of BEV and predict 3D lanes from FV representations directly, while it still underperforms other BEV-based methods given its lack of structured representation for 3D lanes. In this paper, we define 3D lane anchors in the 3D space and propose a BEV-free method named Anchor3DLane to predict 3D lanes directly from FV representations. 3D lane anchors are projected to the FV features to extract their features which contain both good structural and context information to make accurate predictions. In addition, we also develop a global optimization method that makes use of the equal-width property between lanes to reduce the lateral error of predictions. Extensive experiments on three popular 3D lane detection benchmarks show that our Anchor3DLane outperforms previous BEV-based methods and achieves state-of-the-art performances. The code is available at: https://github.com/tusen-ai/Anchor3DLane.

Bird’s-Eye-View (BEV) 3D Object Detection is a crucial multi-view technique for autonomous driving systems. Recently, plenty of works are proposed, following a similar paradigm consisting of three essential components, i.e., camera feature extraction, BEV feature construction, and task heads. Among the three components, BEV feature construction is BEV-specific compared with 2D tasks. Existing methods aggregate the multi-view camera features to the flattened grid in order to construct the BEV feature. However, flattening the BEV space along the height dimension fails to emphasize the informative features of different heights. For example, the barrier is located at a low height while the truck is located at a high height. In this paper, we propose a novel method named BEV Slice Attention Network (BEV-SAN) for exploiting the intrinsic characteristics of different heights. Instead of flattening the BEV space, we first sample along the height dimension to build the global and local BEV slices. Then, the features of BEV slices are aggregated from the camera features and merged by the attention mechanism. Finally, we fuse the merged local and global BEV features by a transformer to generate the final feature map for task heads. The purpose of local BEV slices is to emphasize informative heights. In order to find them, we further propose a LiDAR-guided sampling strategy to leverage the statistical distribution of LiDAR to determine the heights of local slices. Compared with uniform sampling, LiDAR-guided sampling can determine more informative heights. We conduct detailed experiments to demonstrate the effectiveness of BEV-SAN. Code will be released.

Stereo-based 3D object detection, which aims at detecting 3D objects with stereo cameras, shows great potential in low-cost deployment compared to LiDAR-based methods and excellent performance compared to monocular-based algorithms. However, the impressive performance of stereo-based 3D object detection is at the huge cost of high-quality manual annotations, which are hardly attainable for any given scene. Semi-supervised learning, in which limited annotated data and numerous unannotated data are required to achieve a satisfactory model, is a promising method to address the problem of data deficiency. In this work, we propose to achieve semi-supervised learning for stereo-based 3D object detection through pseudo annotation generation from a temporal-aggregated teacher model, which temporally accumulates knowledge from a student model. To facilitate a more stable and accurate depth estimation, we introduce Temporal-Aggregation-Guided (TAG) disparity consistency, a cross-view disparity consistency constraint between the teacher model and the student model for robust and improved depth estimation. To mitigate noise in pseudo annotation generation, we propose a cross-view agreement strategy, in which pseudo annotations should attain high degree of agreements between 3D and 2D views, as well as between binocular views. We perform extensive experiments on the KITTI 3D dataset to demonstrate our proposed method’s capability in leveraging a huge amount of unannotated stereo images to attain significantly improved detection results.

Monocular 3D object detection has become a mainstream approach in automatic driving for its easy application. A prominent advantage is that it does not need LiDAR point clouds during the inference. However, most current methods still rely on 3D point cloud data for labeling the ground truths used in the training phase. This inconsistency between the training and inference makes it hard to utilize the large-scale feedback data and increases the data collection expenses. To bridge this gap, we propose a new weakly supervised monocular 3D objection detection method, which can train the model with only 2D labels marked on images. To be specific, we explore three types of consistency in this task, i.e. the projection, multi-view and direction consistency, and design a weakly-supervised architecture based on these consistencies. Moreover, we propose a new 2D direction labeling method in this task to guide the model for accurate rotation direction prediction. Experiments show that our weakly-supervised method achieves comparable performance with some fully supervised methods. When used as a pre-training method, our model can significantly outperform the corresponding fully-supervised baseline with only 1/3 3D labels.

Mobile monocular 3D object detection (Mono3D) (e.g., on a vehicle, a drone, or a robot) is an important yet challenging task. Existing transformer-based offline Mono3D models adopt grid-based vision tokens, which is suboptimal when using coarse tokens due to the limited available computational power. In this paper, we propose an online Mono3D framework, called MonoATT, which leverages a novel vision transformer with heterogeneous tokens of varying shapes and sizes to facilitate mobile Mono3D. The core idea of MonoATT is to adaptively assign finer tokens to areas of more significance before utilizing a transformer to enhance Mono3D. To this end, we first use prior knowledge to design a scoring network for selecting the most important areas of the image, and then propose a token clustering and merging network with an attention mechanism to gradually merge tokens around the selected areas in multiple stages. Finally, a pixel-level feature map is reconstructed from heterogeneous tokens before employing a SOTA Mono3D detector as the underlying detection core. Experiment results on the real-world KITTI dataset demonstrate that MonoATT can effectively improve the Mono3D accuracy for both near and far objects and guarantee low latency. MonoATT yields the best performance compared with the state-of-the-art methods by a large margin and is ranked number one on the KITTI 3D benchmark.

We tackle the task of Azimuth (angular dimension) super-resolution for Frequency Modulated Continuous Wave (FMCW) multiple-input multiple-output (MIMO) radar. FMCW MIMO radar is widely used in autonomous driving alongside Lidar and RGB cameras. However, compared to Lidar, MIMO radar is usually of low resolution due to hardware size restrictions. For example, achieving 1-degree azimuth resolution requires at least 100 receivers, but a single MIMO device usually supports at most 12 receivers. Having limitations on the number of receivers is problematic since a high-resolution measurement of azimuth angle is essential for estimating the location and velocity of objects. To improve the azimuth resolution of MIMO radar, we propose a light, yet efficient, Analog-to-Digital super-resolution model (ADC-SR) that predicts or hallucinates additional radar signals using signals from only a few receivers. Compared with the baseline models that are applied to processed radar Range-Azimuth-Doppler (RAD) maps, we show that our ADC-SR method that processes raw ADC signals achieves comparable performance with 98% (50 times) fewer parameters. We also propose a hybrid super-resolution model (Hybrid-SR) combining our ADC-SR with a standard RAD super-resolution model, and show that performance can be improved by a large margin. Experiments on our City-Radar dataset and the RADIal dataset validate the importance of leveraging raw radar ADC signals. To assess the value of our super-resolution model for autonomous driving, we also perform object detection on the results of our super-resolution model and find that our super-resolution model improves detection performance by around 4% in mAP.

Self-driving vehicles rely on urban street maps for autonomous navigation. In this paper, we introduce Pix2Map, a method for inferring urban street map topology directly from ego-view images, as needed to continually update and expand existing maps. This is a challenging task, as we need to infer a complex urban road topology directly from raw image data. The main insight of this paper is that this problem can be posed as cross-modal retrieval by learning a joint, cross-modal embedding space for images and existing maps, represented as discrete graphs that encode the topological layout of the visual surroundings. We conduct our experimental evaluation using the Argoverse dataset and show that it is indeed possible to accurately retrieve street maps corresponding to both seen and unseen roads solely from image data. Moreover, we show that our retrieved maps can be used to update or expand existing maps and even show proof-of-concept results for visual localization and image retrieval from spatial graphs.

LiDAR-camera fusion methods have shown impressive performance in 3D object detection. Recent advanced multi-modal methods mainly perform global fusion, where image features and point cloud features are fused across the whole scene. Such practice lacks fine-grained region-level information, yielding suboptimal fusion performance. In this paper, we present the novel Local-to-Global fusion network (LoGoNet), which performs LiDAR-camera fusion at both local and global levels. Concretely, the Global Fusion (GoF) of LoGoNet is built upon previous literature, while we exclusively use point centroids to more precisely represent the position of voxel features, thus achieving better cross-modal alignment. As to the Local Fusion (LoF), we first divide each proposal into uniform grids and then project these grid centers to the images. The image features around the projected grid points are sampled to be fused with position-decorated point cloud features, maximally utilizing the rich contextual information around the proposals. The Feature Dynamic Aggregation (FDA) module is further proposed to achieve information interaction between these locally and globally fused features, thus producing more informative multi-modal features. Extensive experiments on both Waymo Open Dataset (WOD) and KITTI datasets show that LoGoNet outperforms all state-of-the-art 3D detection methods. Notably, LoGoNet ranks 1st on Waymo 3D object detection leaderboard and obtains 81.02 mAPH (L2) detection performance. It is noteworthy that, for the first time, the detection performance on three classes surpasses 80 APH (L2) simultaneously. Code will be available at https://github.com/sankin97/LoGoNet.

High-definition (HD) semantic maps are a crucial component for autonomous driving on urban streets. Traditional offline HD maps are created through labor-intensive manual annotation processes, which are costly and do not accommodate timely updates. Recently, researchers have proposed to infer local maps based on online sensor observations. However, the range of online map inference is constrained by sensor perception range and is easily affected by occlusions. In this work, we propose Neural Map Prior (NMP), a neural representation of global maps that enables automatic global map updates and enhances local map inference performance. To incorporate the strong map prior into local map inference, we leverage cross-attention to dynamically capture the correlations between current features and prior features. For updating the global neural map prior, we use a learning-based fusion module to guide the network in fusing features from previous traversals. This design allows the network to capture a global neural map prior while making sequential online map predictions. Experimental results on the nuScenes dataset demonstrate that our framework is compatible with most map segmentation/detection methods, improving map prediction performance in challenging weather conditions and over an extended horizon. To the best of our knowledge, this represents the first learning-based system for constructing a global map prior.

LiDAR-based 3D point cloud recognition has benefited various applications. Without specially considering the LiDAR point distribution, most current methods suffer from information disconnection and limited receptive field, especially for the sparse distant points. In this work, we study the varying-sparsity distribution of LiDAR points and present SphereFormer to directly aggregate information from dense close points to the sparse distant ones. We design radial window self-attention that partitions the space into multiple non-overlapping narrow and long windows. It overcomes the disconnection issue and enlarges the receptive field smoothly and dramatically, which significantly boosts the performance of sparse distant points. Moreover, to fit the narrow and long windows, we propose exponential splitting to yield fine-grained position encoding and dynamic feature selection to increase model representation ability. Notably, our method ranks 1st on both nuScenes and SemanticKITTI semantic segmentation benchmarks with 81.9% and 74.8% mIoU, respectively. Also, we achieve the 3rd place on nuScenes object detection benchmark with 72.8% NDS and 68.5% mAP. Code is available at https://github.com/dvlab-research/SphereFormer.git.

3D object detection from point clouds is crucial in safety-critical autonomous driving. Although many works have made great efforts and achieved significant progress on this task, most of them suffer from expensive annotation cost and poor transferability to unknown data due to the domain gap. Recently, few works attempt to tackle the domain gap in objects, but still fail to adapt to the gap of varying beam-densities between two domains, which is critical to mitigate the characteristic differences of the LiDAR collectors. To this end, we make the attempt to propose a density-insensitive domain adaption framework to address the density-induced domain gap. In particular, we first introduce Random Beam Re-Sampling (RBRS) to enhance the robustness of 3D detectors trained on the source domain to the varying beam-density. Then, we take this pre-trained detector as the backbone model, and feed the unlabeled target domain data into our newly designed task-specific teacher-student framework for predicting its high-quality pseudo labels. To further adapt the property of density-insensitive into the target domain, we feed the teacher and student branches with the same sample of different densities, and propose an Object Graph Alignment (OGA) module to construct two object-graphs between the two branches for enforcing the consistency in both the attribute and relation of cross-density objects. Experimental results on three widely adopted 3D object detection datasets demonstrate that our proposed domain adaption method outperforms the state-of-the-art methods, especially over varying-density data. Code is available at https://github.com/WoodwindHu/DTS.

In order to deal with the sparse and unstructured raw point clouds, most LiDAR based 3D object detection research focuses on designing dedicated local point aggregators for fine-grained geometrical modeling. In this paper, we revisit the local point aggregators from the perspective of allocating computational resources. We find that the simplest pillar based models perform surprisingly well considering both accuracy and latency. Additionally, we show that minimal adaptions from the success of 2D object detection, such as enlarging receptive field, significantly boost the performance. Extensive experiments reveal that our pillar based networks with modernized designs in terms of architecture and training render the state-of-the-art performance on two popular benchmarks: Waymo Open Dataset and nuScenes. Our results challenge the common intuition that detailed geometry modeling is essential to achieve high performance for 3D object detection.

The modern machine learning-based technologies have shown considerable potential in automatic radar scene understanding. Among these efforts, radar semantic segmentation (RSS) can provide more refined and detailed information including the moving objects and background clutters within the effective receptive field of the radar. Motivated by the success of convolutional networks in various visual computing tasks, these networks have also been introduced to solve RSS task. However, neither the regular convolution operation nor the modified ones are specific to interpret radar signals. The receptive fields of existing convolutions are defined by the object presentation in optical signals, but these two signals have different perception mechanisms. In classic radar signal processing, the object signature is detected according to a local peak response, i.e., CFAR detection. Inspired by this idea, we redefine the receptive field of the convolution operation as the peak receptive field (PRF) and propose the peak convolution operation (PeakConv) to learn the object signatures in an end-to-end network. By incorporating the proposed PeakConv layers into the encoders, our RSS network can achieve better segmentation results compared with other SoTA methods on a multi-view real-measured dataset collected from an FMCW radar. Our code for PeakConv is available at https://github.com/zlw9161/PKC.

With the success of the 3D deep learning models, various perception technologies for autonomous driving have been developed in the LiDAR domain. While these models perform well in the trained source domain, they struggle in unseen domains with a domain gap. In this paper, we propose a single domain generalization method for LiDAR semantic segmentation (DGLSS) that aims to ensure good performance not only in the source domain but also in the unseen domain by learning only on the source domain. We mainly focus on generalizing from a dense source domain and target the domain shift from different LiDAR sensor configurations and scene distributions. To this end, we augment the domain to simulate the unseen domains by randomly subsampling the LiDAR scans. With the augmented domain, we introduce two constraints for generalizable representation learning: sparsity invariant feature consistency (SIFC) and semantic correlation consistency (SCC). The SIFC aligns sparse internal features of the source domain with the augmented domain based on the feature affinity. For SCC, we constrain the correlation between class prototypes to be similar for every LiDAR scan. We also establish a standardized training and evaluation setting for DGLSS. With the proposed evaluation setting, our method showed improved performance in the unseen domains compared to other baselines. Even without access to the target domain, our method performed better than the domain adaptation method. The code is available at https://github.com/gzgzys9887/DGLSS.

Understanding the motion behavior of dynamic environments is vital for autonomous driving, leading to increasing attention in class-agnostic motion prediction in LiDAR point clouds. Outdoor scenes can often be decomposed into mobile foregrounds and static backgrounds, which enables us to associate motion understanding with scene parsing. Based on this observation, we study a novel weakly supervised motion prediction paradigm, where fully or partially (1%, 0.1%) annotated foreground/background binary masks rather than expensive motion annotations are used for supervision. To this end, we propose a two-stage weakly supervised approach, where the segmentation model trained with the incomplete binary masks in Stage1 will facilitate the self-supervised learning of the motion prediction network in Stage2 by estimating possible moving foregrounds in advance. Furthermore, for robust self-supervised motion learning, we design a Consistency-aware Chamfer Distance loss by exploiting multi-frame information and explicitly suppressing potential outliers. Comprehensive experiments show that, with fully or partially binary masks as supervision, our weakly supervised models surpass the self-supervised models by a large margin and perform on par with some supervised ones. This further demonstrates that our approach achieves a good compromise between annotation effort and performance.

Methane (CH 4 ) is the chief contributor to global climate change. Recent Airborne Visible-Infrared Imaging Spectrometer-Next Generation (AVIRIS-NG) has been very useful in quantitative mapping of methane emissions. Existing methods for analyzing this data are sensitive to local terrain conditions, often require manual inspection from domain experts, prone to significant error and hence are not scalable. To address these challenges, we propose a novel end-to-end spectral absorption wavelength aware transformer network, MethaneMapper, to detect and quantify the emissions. MethaneMapper introduces two novel modules that help to locate the most relevant methane plume regions in the spectral domain and uses them to localize these accurately. Thorough evaluation shows that MethaneMapper achieves 0.63 mAP in detection and reduces the model size (by 5×) compared to the current state of the art. In addition, we also introduce a large-scale dataset of methane plume segmentation mask for over 1200 AVIRIS-NG flightlines from 2015-2022. It contains over 4000 methane plume sites. Our dataset will provide researchers the opportunity to develop and advance new methods for tackling this challenging green-house gas detection problem with significant broader social impact. Dataset and source code link.

We study the problem of 3D semantic segmentation from raw point clouds. Unlike existing methods which primarily rely on a large amount of human annotations for training neural networks, we propose the first purely unsupervised method, called GrowSP, to successfully identify complex semantic classes for every point in 3D scenes, without needing any type of human labels or pretrained models. The key to our approach is to discover 3D semantic elements via progressive growing of superpoints. Our method consists of three major components, 1) the feature extractor to learn per-point features from input point clouds, 2) the superpoint constructor to progressively grow the sizes of superpoints, and 3) the semantic primitive clustering module to group superpoints into semantic elements for the final semantic segmentation. We extensively evaluate our method on multiple datasets, demonstrating superior performance over all unsupervised baselines and approaching the classic fully supervised PointNet. We hope our work could inspire more advanced methods for unsupervised 3D semantic learning.

3D shape completion from point clouds is a challenging task, especially from scans of real-world objects. Considering the paucity of 3D shape ground truths for real scans, existing works mainly focus on benchmarking this task on synthetic data, e.g. 3D computer-aided design models. However, the domain gap between synthetic and real data limits the generalizability of these methods. Thus, we propose a new task, SCoDA, for the domain adaptation of real scan shape completion from synthetic data. A new dataset, ScanSalon, is contributed with a bunch of elaborate 3D models created by skillful artists according to scans. To address this new task, we propose a novel cross-domain feature fusion method for knowledge transfer and a novel volume-consistent self-training framework for robust learning from real data. Extensive experiments prove our method is effective to bring an improvement of 6%~7% mIoU.

Training deep models for semantic scene completion is challenging due to the sparse and incomplete input, a large quantity of objects of diverse scales as well as the inherent label noise for moving objects. To address the above-mentioned problems, we propose the following three solutions: 1) Redesigning the completion network. We design a novel completion network, which consists of several Multi-Path Blocks (MPBs) to aggregate multi-scale features and is free from the lossy downsampling operations. 2) Distilling rich knowledge from the multi-frame model. We design a novel knowledge distillation objective, dubbed Dense-to-Sparse Knowledge Distillation (DSKD). It transfers the dense, relation-based semantic knowledge from the multi-frame teacher to the single-frame student, significantly improving the representation learning of the single-frame model. 3) Completion label rectification. We propose a simple yet effective label rectification strategy, which uses off-the-shelf panoptic segmentation labels to remove the traces of dynamic objects in completion labels, greatly improving the performance of deep models especially for those moving objects. Extensive experiments are conducted in two public semantic scene completion benchmarks, i.e., SemanticKITTI and SemanticPOSS. Our SCPNet ranks 1st on SemanticKITTI semantic scene completion challenge and surpasses the competitive S3CNet by 7.2 mIoU. SCPNet also outperforms previous completion algorithms on the SemanticPOSS dataset. Besides, our method also achieves competitive results on SemanticKITTI semantic segmentation tasks, showing that knowledge learned in the scene completion is beneficial to the segmentation task.

Although different approaches have been proposed for 3D point cloud-related tasks, few-shot learning (FSL) of 3D point clouds still remains under-explored. In FSL, unlike traditional supervised learning, the classes of training and test data do not overlap, and a model needs to recognize unseen classes from only a few samples. Existing FSL methods for 3D point clouds employ point-based models as their backbone. Yet, based on our extensive experiments and analysis, we first show that using a point-based backbone is not the most suitable FSL approach, since (i) a large number of points’ features are discarded by the max pooling operation used in 3D point-based backbones, decreasing the ability of representing shape information; (ii)point-based backbones are sensitive to occlusion. To address these issues, we propose employing a projection- and 2D Convolutional Neural Network-based backbone, referred to as the ViewNet, for FSL from 3D point clouds. Our approach first projects a 3D point cloud onto six different views to alleviate the issue of missing points. Also, to generate more descriptive and distinguishing features, we propose View Pooling, which combines different projected plane combinations into five groups and performs max-pooling on each of them. The experiments performed on the ModelNet40, ScanObjectNN and ModelNet40-C datasets, with cross validation, show that our method consistently outperforms the state-of-the-art baselines. Moreover, compared to traditional image classification backbones, such as ResNet, the proposed ViewNet can extract more distinguishing features from multiple views of a point cloud. We also show that ViewNet can be used as a backbone with different FSL heads and provides improved performance compared to traditionally used backbones.

Recent work on 4D point cloud sequences has attracted a lot of attention. However, obtaining exhaustively labeled 4D datasets is often very expensive and laborious, so it is especially important to investigate how to utilize raw unlabeled data. However, most existing self-supervised point cloud representation learning methods only consider geometry from a static snapshot omitting the fact that sequential observations of dynamic scenes could reveal more comprehensive geometric details. To overcome such issues, this paper proposes a new 4D self-supervised pre-training method called Complete-to-Partial 4D Distillation. Our key idea is to formulate 4D self-supervised representation learning as a teacher-student knowledge distillation framework and let the student learn useful 4D representations with the guidance of the teacher. Experiments show that this approach significantly outperforms previous pre-training approaches on a wide range of 4D point cloud sequence understanding tasks. Code is available at: https://github.com/dongyh20/C2P.

Point cloud sampling is crucial for efficient large-scale point cloud analysis, where learning-to-sample methods have recently received increasing attention from the community for jointly training with downstream tasks. However, the above-mentioned task-specific sampling methods usually fail to explore the geometries of objects in an explicit manner. In this paper, we introduce a new skeleton-aware learning-to-sample method by learning object skeletons as the prior knowledge to preserve the object geometry and topology information during sampling. Specifically, without labor-intensive annotations per object category, we first learn category-agnostic object skeletons via the medial axis transform definition in an unsupervised manner. With object skeleton, we then evaluate the histogram of the local feature size as the prior knowledge to formulate skeleton-aware sampling from a probabilistic perspective. Additionally, the proposed skeleton-aware sampling pipeline with the task network is thus end-to-end trainable by exploring the reparameterization trick. Extensive experiments on three popular downstream tasks, point cloud classification, retrieval, and reconstruction, demonstrate the effectiveness of the proposed method for efficient point cloud analysis.

Recent advances in 3D point cloud analysis bring a diverse set of network architectures to the field. However, the lack of a unified framework to interpret those networks makes any systematic comparison, contrast, or analysis challenging, and practically limits healthy development of the field. In this paper, we take the initiative to explore and propose a unified framework called PointMeta, to which the popular 3D point cloud analysis approaches could fit. This brings three benefits. First, it allows us to compare different approaches in a fair manner, and use quick experiments to verify any empirical observations or assumptions summarized from the comparison. Second, the big picture brought by PointMeta enables us to think across different components, and revisit common beliefs and key design decisions made by the popular approaches. Third, based on the learnings from the previous two analyses, by doing simple tweaks on the existing approaches, we are able to derive a basic building block, termed PointMetaBase. It shows very strong performance in efficiency and effectiveness through extensive experiments on challenging benchmarks, and thus verifies the necessity and benefits of high-level interpretation, contrast, and comparison like PointMeta. In particular, PointMetaBase surpasses the previous state-of-the-art method by 0.7%/1.4/%2.1% mIoU with only 2%/11%/13% of the computation cost on the S3DIS datasets. Codes are available in the supplementary materials.

Deep neural networks on regular 1D lists (e.g., natural languages) and irregular 3D sets (e.g., point clouds) have made tremendous achievements. The key to natural language processing is to model words and their regular order dependency in texts. For point cloud understanding, the challenge is to understand the geometry via irregular point coordinates, in which point-feeding orders do not matter. However, there are a few kinds of data that exhibit both regular 1D list and irregular 3D set structures, such as proteins and non-coding RNAs. In this paper, we refer to them as 3D point lists and propose a Transformer-style PointListNet to model them. First, PointListNet employs non-parametric distance-based attention because we find sometimes it is the distance, instead of the feature or type, that mainly determines how much two points, e.g., amino acids, are correlated in the micro world. Second, different from the vanilla Transformer that directly performs a simple linear transformation on inputs to generate values and does not explicitly model relative relations, our PointListNet integrates the 1D order and 3D Euclidean displacements into values. We conduct experiments on protein fold classification and enzyme reaction classification. Experimental results show the effectiveness of the proposed PointListNet.

Learning distinctive point-wise features is critical for low-overlap point cloud registration. Recently, it has achieved huge success in incorporating Transformer into point cloud feature representation, which usually adopts a self-attention module to learn intra-point-cloud features first, then utilizes a cross-attention module to perform feature exchange between input point clouds. Self-attention is computed by capturing the global dependency in geometric space. However, this global dependency can be ambiguous and lacks distinctiveness, especially in indoor low-overlap scenarios, as which the dependence with an extensive range of non-overlapping points introduces ambiguity. To address this issue, we present PEAL, a Prior-embedded Explicit Attention Learning model. By incorporating prior knowledge into the learning process, the points are divided into two parts. One includes points lying in the putative overlapping region and the other includes points lying in the putative non-overlapping region. Then PEAL explicitly learns one-way attention with the putative overlapping points. This simplistic design attains surprising performance, significantly relieving the aforementioned feature ambiguity. Our method improves the Registration Recall by 6+% on the challenging 3DLoMatch benchmark and achieves state-of-the-art performance on Feature Matching Recall, Inlier Ratio, and Registration Recall on both 3DMatch and 3DLoMatch. Code will be made publicly available.

It is vital to infer signed distance functions (SDFs) from 3D point clouds. The latest methods rely on generalizing the priors learned from large scale supervision. However, the learned priors do not generalize well to various geometric variations that are unseen during training, especially for extremely sparse point clouds. To resolve this issue, we present a neural network to directly infer SDFs from single sparse point clouds without using signed distance supervision, learned priors or even normals. Our insight here is to learn surface parameterization and SDFs inference in an end-to-end manner. To make up the sparsity, we leverage parameterized surfaces as a coarse surface sampler to provide many coarse surface estimations in training iterations, according to which we mine supervision and our thin plate splines (TPS) based network infers SDFs as smooth functions in a statistical way. Our method significantly improves the generalization ability and accuracy in unseen point clouds. Our experimental results show our advantages over the state-of-the-art methods in surface reconstruction for sparse point clouds under synthetic datasets and real scans.The code is available at https://github.com/chenchao15/NeuralTPS.

Neural signed distance functions (SDFs) have shown remarkable capability in representing geometry with details. However, without signed distance supervision, it is still a challenge to infer SDFs from point clouds or multi-view images using neural networks. In this paper, we claim that gradient consistency in the field, indicated by the parallelism of level sets, is the key factor affecting the inference accuracy. Hence, we propose a level set alignment loss to evaluate the parallelism of level sets, which can be minimized to achieve better gradient consistency. Our novelty lies in that we can align all level sets to the zero level set by constraining gradients at queries and their projections on the zero level set in an adaptive way. Our insight is to propagate the zero level set to everywhere in the field through consistent gradients to eliminate uncertainty in the field that is caused by the discreteness of 3D point clouds or the lack of observations from multi-view images. Our proposed loss is a general term which can be used upon different methods to infer SDFs from 3D point clouds and multi-view images. Our numerical and visual comparisons demonstrate that our loss can significantly improve the accuracy of SDFs inferred from point clouds or multi-view images under various benchmarks. Code and data are available at https://github.com/mabaorui/TowardsBetterGradient.

The matching of 3D shapes has been extensively studied for shapes represented as surface meshes, as well as for shapes represented as point clouds. While point clouds are a common representation of raw real-world 3D data (e.g. from laser scanners), meshes encode rich and expressive topological information, but their creation typically requires some form of (often manual) curation. In turn, methods that purely rely on point clouds are unable to meet the matching quality of mesh-based methods that utilise the additional topological structure. In this work we close this gap by introducing a self-supervised multimodal learning strategy that combines mesh-based functional map regularisation with a contrastive loss that couples mesh and point cloud data. Our shape matching approach allows to obtain intramodal correspondences for triangle meshes, complete point clouds, and partially observed point clouds, as well as correspondences across these data modalities. We demonstrate that our method achieves state-of-the-art results on several challenging benchmark datasets even in comparison to recent supervised methods, and that our method reaches previously unseen cross-dataset generalisation ability.

As a fundamental problem in computer vision, 3D point cloud registration (PCR) aims to seek the optimal pose to align a point cloud pair. In this paper, we present a 3D registration method with maximal cliques (MAC). The key insight is to loosen the previous maximum clique constraint, and to mine more local consensus information in a graph for accurate pose hypotheses generation: 1) A compatibility graph is constructed to render the affinity relationship between initial correspondences. 2) We search for maximal cliques in the graph, each of which represents a consensus set. We perform node-guided clique selection then, where each node corresponds to the maximal clique with the greatest graph weight. 3) Transformation hypotheses are computed for the selected cliques by SVD algorithm and the best hypothesis is used to perform registration. Extensive experiments on U3M, 3DMatch, 3DLoMatch and KITTI demonstrate that MAC effectively increases registration accuracy, outperforms various state-of-the-art methods and boosts the performance of deep-learned methods. MAC combined with deep-learned methods achieves state-of-the-art registration recall of 95.7% / 78.9% on the 3DMatch / 3DLoMatch.

In panorama understanding, the widely used equirectangular projection (ERP) entails boundary discontinuity and spatial distortion. It severely deteriorates the conventional CNNs and vision Transformers on panoramas. In this paper, we propose a simple yet effective architecture named PanoSwin to learn panorama representations with ERP. To deal with the challenges brought by equirectangular projection, we explore a pano-style shift windowing scheme and novel pitch attention to address the boundary discontinuity and the spatial distortion, respectively. Besides, based on spherical distance and Cartesian coordinates, we adapt absolute positional encodings and relative positional biases for panoramas to enhance panoramic geometry information. Realizing that planar image understanding might share some common knowledge with panorama understanding, we devise a novel two-stage learning framework to facilitate knowledge transfer from the planar images to panoramas. We conduct experiments against the state-of-the-art on various panoramic tasks, i.e., panoramic object detection, panoramic classification, and panoramic layout estimation. The experimental results demonstrate the effectiveness of PanoSwin in panorama understanding.

Feature matching is a challenging computer vision task that involves finding correspondences between two images of a 3D scene. In this paper we consider the dense approach instead of the more common sparse paradigm, thus striving to find all correspondences. Perhaps counter-intuitively, dense methods have previously shown inferior performance to their sparse and semi-sparse counterparts for estimation of two-view geometry. This changes with our novel dense method, which outperforms both dense and sparse methods on geometry estimation. The novelty is threefold: First, we propose a kernel regression global matcher. Secondly, we propose warp refinement through stacked feature maps and depthwise convolution kernels. Thirdly, we propose learning dense confidence through consistent depth and a balanced sampling approach for dense confidence maps. Through extensive experiments we confirm that our proposed dense method, Dense Kernelized Feature Matching, sets a new state-of-the-art on multiple geometry estimation benchmarks. In particular, we achieve an improvement on MegaDepth-1500 of +4.9 and +8.9 AUC@5 compared to the best previous sparse method and dense method respectively. Our code is provided at the following repository: https://github.com/Parskatt/DKM

Local feature matching aims at establishing sparse correspondences between a pair of images. Recently, detector-free methods present generally better performance but are not satisfactory in image pairs with large scale differences. In this paper, we propose Patch Area Transportation with Subdivision (PATS) to tackle this issue. Instead of building an expensive image pyramid, we start by splitting the original image pair into equal-sized patches and gradually resizing and subdividing them into smaller patches with the same scale. However, estimating scale differences between these patches is non-trivial since the scale differences are determined by both relative camera poses and scene structures, and thus spatially varying over image pairs. Moreover, it is hard to obtain the ground truth for real scenes. To this end, we propose patch area transportation, which enables learning scale differences in a self-supervised manner. In contrast to bipartite graph matching, which only handles one-to-one matching, our patch area transportation can deal with many-to-many relationships. PATS improves both matching accuracy and coverage, and shows superior performance in downstream tasks, such as relative pose estimation, visual localization, and optical flow estimation.The source code will be released to benefit the community.

This paper solves the problem of learning dense visual correspondences between different object instances of the same category with only sparse annotations. We decompose this pixel-level semantic matching problem into two easier ones: (i) First, local feature descriptors of source and target images need to be mapped into shared semantic spaces to get coarse matching flows. (ii) Second, matching flows in low resolution should be refined to generate accurate point-to-point matching results. We propose asymmetric feature learning and matching flow super-resolution based on vision transformers to solve the above problems. The asymmetric feature learning module exploits a biased cross-attention mechanism to encode token features of source images with their target counterparts. Then matching flow in low resolutions is enhanced by a super-resolution network to get accurate correspondences. Our pipeline is built upon vision transformers and can be trained in an end-to-end manner. Extensive experimental results on several popular benchmarks, such as PF-PASCAL, PF-WILLOW, and SPair-71K, demonstrate that the proposed method can catch subtle semantic differences in pixels efficiently. Code is available on https://github.com/YXSUNMADMAX/ACTR.

Recently, event-based stereo matching has been studied due to its robustness in poor light conditions. However, existing event-based stereo networks suffer severe performance degradation when domains shift. Unsupervised domain adaptation (UDA) aims at resolving this problem without using the target domain ground-truth. However, traditional UDA still needs the input event data with ground-truth in the source domain, which is more challenging and costly to obtain than image data. To tackle this issue, we propose a novel unsupervised domain Adaptive Dense Event Stereo (ADES), which resolves gaps between the different domains and input modalities. The proposed ADES framework adapts event-based stereo networks from abundant image datasets with ground-truth on the source domain to event datasets without ground-truth on the target domain, which is a more practical setup. First, we propose a self-supervision module that trains the network on the target domain through image reconstruction, while an artifact prediction network trained on the source domain assists in removing intermittent artifacts in the reconstructed image. Secondly, we utilize the feature-level normalization scheme to align the extracted features along the epipolar line. Finally, we present the motion-invariant consistency module to impose the consistent output between the perturbed motion. Our experiments demonstrate that our approach achieves remarkable results in the adaptation ability of event-based stereo matching from the image domain.

Outlier-robust estimation involves estimating some parameters (e.g., 3D rotations) from data samples in the presence of outliers, and is typically formulated as a non-convex and non-smooth problem. For this problem, the classical method called iteratively reweighted least-squares (IRLS) and its variants have shown impressive performance. This paper makes several contributions towards understanding why these algorithms work so well. First, we incorporate majorization and graduated non-convexity (GNC) into the IRLS framework and prove that the resulting IRLS variant is a convergent method for outlier-robust estimation. Moreover, in the robust regression context with a constant fraction of outliers, we prove this IRLS variant converges to the ground truth at a global linear and local quadratic rate for a random Gaussian feature matrix with high probability. Experiments corroborate our theory and show that the proposed IRLS variant converges within 5-10 iterations for typical problem instances of outlier-robust estimation, while state-of-the-art methods need at least 30 iterations. A basic implementation of our method is provided: https://github.com/liangzu/IRLS-CVPR2023

In this paper, we propose YOSO, a real-time panoptic segmentation framework. YOSO predicts masks via dynamic convolutions between panoptic kernels and image feature maps, in which you only need to segment once for both instance and semantic segmentation tasks. To reduce the computational overhead, we design a feature pyramid aggregator for the feature map extraction, and a separable dynamic decoder for the panoptic kernel generation. The aggregator re-parameterizes interpolation-first modules in a convolution-first way, which significantly speeds up the pipeline without any additional costs. The decoder performs multi-head cross-attention via separable dynamic convolution for better efficiency and accuracy. To the best of our knowledge, YOSO is the first real-time panoptic segmentation framework that delivers competitive performance compared to state-of-the-art models. Specifically, YOSO achieves 46.4 PQ, 45.6 FPS on COCO; 52.5 PQ, 22.6 FPS on Cityscapes; 38.0 PQ, 35.4 FPS on ADE20K; and 34.1 PQ, 7.1 FPS on Mapillary Vistas. Code is available at https://github.com/hujiecpp/YOSO.

We present a novel bird’s-eye-view (BEV) detector with perspective supervision, which converges faster and better suits modern image backbones. Existing state-of-the-art BEV detectors are often tied to certain depth pre-trained backbones like VoVNet, hindering the synergy between booming image backbones and BEV detectors. To address this limitation, we prioritize easing the optimization of BEV detectors by introducing perspective space supervision. To this end, we propose a two-stage BEV detector, where proposals from the perspective head are fed into the bird’s-eye-view head for final predictions. To evaluate the effectiveness of our model, we conduct extensive ablation studies focusing on the form of supervision and the generality of the proposed detector. The proposed method is verified with a wide spectrum of traditional and modern image backbones and achieves new SoTA results on the large-scale nuScenes dataset. The code shall be released soon.

Human-centric perceptions (e.g., pose estimation, human parsing, pedestrian detection, person re-identification, etc.) play a key role in industrial applications of visual models. While specific human-centric tasks have their own relevant semantic aspect to focus on, they also share the same underlying semantic structure of the human body. However, few works have attempted to exploit such homogeneity and design a general-propose model for human-centric tasks. In this work, we revisit a broad range of human-centric tasks and unify them in a minimalist manner. We propose UniHCP, a Unified Model for Human-Centric Perceptions, which unifies a wide range of human-centric tasks in a simplified end-to-end manner with the plain vision transformer architecture. With large-scale joint training on 33 humancentric datasets, UniHCP can outperform strong baselines on several in-domain and downstream tasks by direct evaluation. When adapted to a specific task, UniHCP achieves new SOTAs on a wide range of human-centric tasks, e.g., 69.8 mIoU on CIHP for human parsing, 86.18 mA on PA-100K for attribute prediction, 90.3 mAP on Market1501 for ReID, and 85.8 JI on CrowdHuman for pedestrian detection, performing better than specialized models tailored for each task. The code and pretrained model are available at https://github.com/OpenGVLab/UniHCP.

Modern autonomous driving system is characterized as modular tasks in sequential order, i.e., perception, prediction, and planning. In order to perform a wide diversity of tasks and achieve advanced-level intelligence, contemporary approaches either deploy standalone models for individual tasks, or design a multi-task paradigm with separate heads. However, they might suffer from accumulative errors or deficient task coordination. Instead, we argue that a favorable framework should be devised and optimized in pursuit of the ultimate goal, i.e., planning of the self-driving car. Oriented at this, we revisit the key components within perception and prediction, and prioritize the tasks such that all these tasks contribute to planning. We introduce Unified Autonomous Driving (UniAD), a comprehensive framework up-to-date that incorporates full-stack driving tasks in one network. It is exquisitely devised to leverage advantages of each module, and provide complementary feature abstractions for agent interaction from a global perspective. Tasks are communicated with unified query interfaces to facilitate each other toward planning. We instantiate UniAD on the challenging nuScenes benchmark. With extensive ablations, the effectiveness of using such a philosophy is proven by substantially outperforming previous state-of-the-arts in all aspects. Code and models are public.

Predicting the future trajectories of surrounding agents is essential for autonomous vehicles to operate safely. This paper presents QCNet, a modeling framework toward pushing the boundaries of trajectory prediction. First, we identify that the agent-centric modeling scheme used by existing approaches requires re-normalizing and re-encoding the input whenever the observation window slides forward, leading to redundant computations during online prediction. To overcome this limitation and achieve faster inference, we introduce a query-centric paradigm for scene encoding, which enables the reuse of past computations by learning representations independent of the global spacetime coordinate system. Sharing the invariant scene features among all target agents further allows the parallelism of multi-agent trajectory decoding. Second, even given rich encodings of the scene, existing decoding strategies struggle to capture the multimodality inherent in agents’ future behavior, especially when the prediction horizon is long. To tackle this challenge, we first employ anchor-free queries to generate trajectory proposals in a recurrent fashion, which allows the model to utilize different scene contexts when decoding waypoints at different horizons. A refinement module then takes the trajectory proposals as anchors and leverages anchor-based queries to refine the trajectories further. By supplying adaptive and high-quality anchors to the refinement module, our query-based decoder can better deal with the multimodality in the output of trajectory prediction. Our approach ranks 1st on Argoverse 1 and Argoverse 2 motion forecasting benchmarks, outperforming all methods on all main metrics by a large margin. Meanwhile, our model can achieve streaming scene encoding and parallel multi-agent decoding thanks to the query-centric design ethos.

The indeterminate nature of human motion requires trajectory prediction systems to use a probabilistic model to formulate the multi-modality phenomenon and infer a finite set of future trajectories. However, the inference processes of most existing methods rely on Monte Carlo random sampling, which is insufficient to cover the realistic paths with finite samples, due to the long tail effect of the predicted distribution. To promote the sampling process of stochastic prediction, we propose a novel method, called BOsampler, to adaptively mine potential paths with Bayesian optimization in an unsupervised manner, as a sequential design strategy in which new prediction is dependent on the previously drawn samples. Specifically, we model the trajectory sampling as a Gaussian process and construct an acquisition function to measure the potential sampling value. This acquisition function applies the original distribution as prior and encourages exploring paths in the long-tail region. This sampling method can be integrated with existing stochastic predictive models without retraining. Experimental results on various baseline methods demonstrate the effectiveness of our method. The source code is released in this link.

Predicting future action locations is vital for applications like human-robot collaboration. While some computer vision tasks have made progress in predicting human actions, accurately localizing these actions in future frames remains an area with room for improvement. We introduce a new task called spatial action localization in the future (SALF), which aims to predict action locations in both observed and future frames. SALF is challenging because it requires understanding the underlying physics of video observations to predict future action locations accurately. To address SALF, we use the concept of NeuralODE, which models the latent dynamics of sequential data by solving ordinary differential equations (ODE) with neural networks. We propose a novel architecture, AdamsFormer, which extends observed frame features to future time horizons by modeling continuous temporal dynamics through ODE solving. Specifically, we employ the Adams method, a multi-step approach that efficiently uses information from previous steps without discarding it. Our extensive experiments on UCF101-24 and JHMDB-21 datasets demonstrate that our proposed model outperforms existing long-range temporal modeling methods by a significant margin in terms of frame-mAP.

We study ObjectGoal Navigation -- where a virtual robot situated in a new environment is asked to navigate to an object. Prior work has shown that imitation learning (IL) using behavior cloning (BC) on a dataset of human demonstrations achieves promising results. However, this has limitations -- 1) BC policies generalize poorly to new states, since the training mimics actions not their consequences, and 2) collecting demonstrations is expensive. On the other hand, reinforcement learning (RL) is trivially scalable, but requires careful reward engineering to achieve desirable behavior. We present PIRLNav, a two-stage learning scheme for BC pretraining on human demonstrations followed by RL-finetuning. This leads to a policy that achieves a success rate of 65.0% on ObjectNav (+5.0% absolute over previous state-of-the-art). Using this BC->RL training recipe, we present a rigorous empirical analysis of design choices. First, we investigate whether human demonstrations can be replaced with ‘free’ (automatically generated) sources of demonstrations, e.g. shortest paths (SP) or task-agnostic frontier exploration (FE) trajectories. We find that BC->RL on human demonstrations outperforms BC->RL on SP and FE trajectories, even when controlled for the same BC-pretraining success on train, and even on a subset of val episodes where BC-pretraining success favors the SP or FE policies. Next, we study how RL-finetuning performance scales with the size of the BC pretraining dataset. We find that as we increase the size of the BC-pretraining dataset and get to high BC accuracies, the improvements from RL-finetuning are smaller, and that 90% of the performance of our best BC->RL policy can be achieved with less than half the number of BC demonstrations. Finally, we analyze failure modes of our ObjectNav policies, and present guidelines for further improving them.

Expert demonstrations are a rich source of supervision for training visual robotic manipulation policies, but imitation learning methods often require either a large number of demonstrations or expensive online expert supervision to learn reactive closed-loop behaviors. In this work, we introduce SPARTN (Synthetic Perturbations for Augmenting Robot Trajectories via NeRF): a fully-offline data augmentation scheme for improving robot policies that use eye-in-hand cameras. Our approach leverages neural radiance fields (NeRFs) to synthetically inject corrective noise into visual demonstrations: using NeRFs to generate perturbed viewpoints while simultaneously calculating the corrective actions. This requires no additional expert supervision or environment interaction, and distills the geometric information in NeRFs into a real-time reactive RGB-only policy. In a simulated 6-DoF visual grasping benchmark, SPARTN improves offline success rates by 2.8× over imitation learning without the corrective augmentations and even outperforms some methods that use online supervision. It additionally closes the gap between RGB-only and RGB-D success rates, eliminating the previous need for depth sensors. In real-world 6-DoF robotic grasping experiments from limited human demonstrations, our method improves absolute success rates by 22.5% on average, including objects that are traditionally challenging for depth-based methods.

Camouflaged Instance Segmentation (CIS) aims at predicting the instance-level masks of camouflaged objects, which are usually the animals in the wild adapting their appearance to match the surroundings. Previous instance segmentation methods perform poorly on this task as they are easily disturbed by the deceptive camouflage. To address these challenges, we propose a novel De-camouflaging Network (DCNet) including a pixel-level camouflage decoupling module and an instance-level camouflage suppression module. The proposed DCNet enjoys several merits. First, the pixel-level camouflage decoupling module can extract camouflage characteristics based on the Fourier transformation. Then a difference attention mechanism is proposed to eliminate the camouflage characteristics while reserving target object characteristics in the pixel feature. Second, the instance-level camouflage suppression module can aggregate rich instance information from pixels by use of instance prototypes. To mitigate the effect of background noise during segmentation, we introduce some reliable reference points to build a more robust similarity measurement. With the aid of these two modules, our DCNet can effectively model de-camouflaging and achieve accurate segmentation for camouflaged instances. Extensive experimental results on two benchmarks demonstrate that our DCNet performs favorably against state-of-the-art CIS methods, e.g., with more than 5% performance gains on COD10K and NC4K datasets in average precision.

This work proposes an end-to-end multi-camera 3D multi-object tracking (MOT) framework. It emphasizes spatio-temporal continuity and integrates both past and future reasoning for tracked objects. Thus, we name it “Past-and-Future reasoning for Tracking” (PF-Track). Specifically, our method adapts the “tracking by attention” framework and represents tracked instances coherently over time with object queries. To explicitly use historical cues, our “Past Reasoning” module learns to refine the tracks and enhance the object features by cross-attending to queries from previous frames and other objects. The “Future Reasoning” module digests historical information and predicts robust future trajectories. In the case of long-term occlusions, our method maintains the object positions and enables re-association by integrating motion predictions. On the nuScenes dataset, our method improves AMOTA by a large margin and remarkably reduces ID-Switches by 90% compared to prior approaches, which is an order of magnitude less. The code and models are made available at https://github.com/TRI-ML/PF-Track.

The main challenge of Multi-Object Tracking~(MOT) lies in maintaining a continuous trajectory for each target. Existing methods often learn reliable motion patterns to match the same target between adjacent frames and discriminative appearance features to re-identify the lost targets after a long period. However, the reliability of motion prediction and the discriminability of appearances can be easily hurt by dense crowds and extreme occlusions in the tracking process. In this paper, we propose a simple yet effective multi-object tracker, i.e., MotionTrack, which learns robust short-term and long-term motions in a unified framework to associate trajectories from a short to long range. For dense crowds, we design a novel Interaction Module to learn interaction-aware motions from short-term trajectories, which can estimate the complex movement of each target. For extreme occlusions, we build a novel Refind Module to learn reliable long-term motions from the target’s history trajectory, which can link the interrupted trajectory with its corresponding detection. Our Interaction Module and Refind Module are embedded in the well-known tracking-by-detection paradigm, which can work in tandem to maintain superior performance. Extensive experimental results on MOT17 and MOT20 datasets demonstrate the superiority of our approach in challenging scenarios, and it achieves state-of-the-art performances at various MOT metrics. We will make the code and trained models publicly available.

Gait recognition is a biometric technology that identifies people by their walking patterns. The silhouettes-based method and the skeletons-based method are the two most popular approaches. However, the silhouette data are easily affected by clothing occlusion, and the skeleton data lack body shape information. To obtain a more robust and comprehensive gait representation for recognition, we propose a transformer-based gait recognition framework called MMGaitFormer, which effectively fuses and aggregates the spatial-temporal information from the skeletons and silhouettes. Specifically, a Spatial Fusion Module (SFM) and a Temporal Fusion Module (TFM) are proposed for effective spatial-level and temporal-level feature fusion, respectively. The SFM performs fine-grained body parts spatial fusion and guides the alignment of each part of the silhouette and each joint of the skeleton through the attention mechanism. The TFM performs temporal modeling through Cycle Position Embedding (CPE) and fuses temporal information of two modalities. Experiments demonstrate that our MMGaitFormer achieves state-of-the-art performance on popular gait datasets. For the most challenging “CL” (i.e., walking in different clothes) condition in CASIA-B, our method achieves a rank-1 accuracy of 94.8%, which outperforms the state-of-the-art single-modal methods by a large margin.

Dynamic Facial Expression Recognition (DFER) is a rapidly developing field that focuses on recognizing facial expressions in video format. Previous research has considered non-target frames as noisy frames, but we propose that it should be treated as a weakly supervised problem. We also identify the imbalance of short- and long-term temporal relationships in DFER. Therefore, we introduce the Multi-3D Dynamic Facial Expression Learning (M3DFEL) framework, which utilizes Multi-Instance Learning (MIL) to handle inexact labels. M3DFEL generates 3D-instances to model the strong short-term temporal relationship and utilizes 3DCNNs for feature extraction. The Dynamic Long-term Instance Aggregation Module (DLIAM) is then utilized to learn the long-term temporal relationships and dynamically aggregate the instances. Our experiments on DFEW and FERV39K datasets show that M3DFEL outperforms existing state-of-the-art approaches with a vanilla R3D18 backbone. The source code is available at https://github.com/faceeyes/M3DFEL.

Talking head generation aims to generate faces that maintain the identity information of the source image and imitate the motion of the driving image. Most pioneering methods rely primarily on 2D representations and thus will inevitably suffer from face distortion when large head rotations are encountered. Recent works instead employ explicit 3D structural representations or implicit neural rendering to improve performance under large pose changes. Nevertheless, the fidelity of identity and expression is not so desirable, especially for novel-view synthesis. In this paper, we propose HiDe-NeRF, which achieves high-fidelity and free-view talking-head synthesis. Drawing on the recently proposed Deformable Neural Radiance Fields, HiDe-NeRF represents the 3D dynamic scene into a canonical appearance field and an implicit deformation field, where the former comprises the canonical source face and the latter models the driving pose and expression. In particular, we improve fidelity from two aspects: (i) to enhance identity expressiveness, we design a generalized appearance module that leverages multi-scale volume features to preserve face shape and details; (ii) to improve expression preciseness, we propose a lightweight deformation module that explicitly decouples the pose and expression to enable precise expression modeling. Extensive experiments demonstrate that our proposed approach can generate better results than previous works. Project page: https://www.waytron.net/hidenerf/

We present a novel one-shot talking head synthesis method that achieves disentangled and fine-grained control over lip motion, eye gaze&blink, head pose, and emotional expression. We represent different motions via disentangled latent representations and leverage an image generator to synthesize talking heads from them. To effectively disentangle each motion factor, we propose a progressive disentangled representation learning strategy by separating the factors in a coarse-to-fine manner, where we first extract unified motion feature from the driving signal, and then isolate each fine-grained motion from the unified feature. We introduce motion-specific contrastive learning and regressing for non-emotional motions, and feature-level decorrelation and self-reconstruction for emotional expression, to fully utilize the inherent properties of each motion factor in unstructured video data to achieve disentanglement. Experiments show that our method provides high quality speech&lip-motion synchronization along with precise and disentangled control over multiple extra facial motions, which can hardly be achieved by previous methods.

Video-based person re-identification (Re-ID) is a prominent computer vision topic due to its wide range of video surveillance applications. Most existing methods utilize spatial and temporal correlations in frame sequences to obtain discriminative person features. However, inevitable degradations, e.g., motion blur contained in frames often cause ambiguity texture noise and temporal disturbance, leading to the loss of identity-discriminating cues. Recently, a new bio-inspired sensor called event camera, which can asynchronously record intensity changes, brings new vitality to the Re-ID task. With the microsecond resolution and low latency, event cameras can accurately capture the movements of pedestrians even in the aforementioned degraded environments. Inspired by the properties of event cameras, in this work, we propose a Sparse-Dense Complementary Learning Framework, which effectively extracts identity features by fully exploiting the complementary information of dense frames and sparse events. Specifically, for frames, we build a CNN-based module to aggregate the dense features of pedestrian appearance step-by-step, while for event streams, we design a bio-inspired spiking neural backbone, which encodes event signals into sparse feature maps in a spiking form, to present the dynamic motion cues of pedestrians. Finally, a cross feature alignment module is constructed to complementarily fuse motion information from events and appearance cues from frames to enhance identity representation learning. Experiments on several benchmarks show that by employing events and SNN into Re-ID, our method significantly outperforms competitive methods.

We study a challenging task, conditional human motion generation, which produces plausible human motion sequences according to various conditional inputs, such as action classes or textual descriptors. Since human motions are highly diverse and have a property of quite different distribution from conditional modalities, such as textual descriptors in natural languages, it is hard to learn a probabilistic mapping from the desired conditional modality to the human motion sequences. Besides, the raw motion data from the motion capture system might be redundant in sequences and contain noises; directly modeling the joint distribution over the raw motion sequences and conditional modalities would need a heavy computational overhead and might result in artifacts introduced by the captured noises. To learn a better representation of the various human motion sequences, we first design a powerful Variational AutoEncoder (VAE) and arrive at a representative and low-dimensional latent code for a human motion sequence. Then, instead of using a diffusion model to establish the connections between the raw motion sequences and the conditional inputs, we perform a diffusion process on the motion latent space. Our proposed Motion Latent-based Diffusion model (MLD) could produce vivid motion sequences conforming to the given conditional inputs and substantially reduce the computational overhead in both the training and inference stages. Extensive experiments on various human motion generation tasks demonstrate that our MLD achieves significant improvements over the state-of-the-art methods among extensive human motion generation tasks, with two orders of magnitude faster than previous diffusion models on raw motion sequences.

Current state-of-the-art approaches for few-shot action recognition achieve promising performance by conducting frame-level matching on learned visual features. However, they generally suffer from two limitations: i) the matching procedure between local frames tends to be inaccurate due to the lack of guidance to force long-range temporal perception; ii) explicit motion learning is usually ignored, leading to partial information loss. To address these issues, we develop a Motion-augmented Long-short Contrastive Learning (MoLo) method that contains two crucial components, including a long-short contrastive objective and a motion autodecoder. Specifically, the long-short contrastive objective is to endow local frame features with long-form temporal awareness by maximizing their agreement with the global token of videos belonging to the same class. The motion autodecoder is a lightweight architecture to reconstruct pixel motions from the differential features, which explicitly embeds the network with motion dynamics. By this means, MoLo can simultaneously learn long-range temporal context and motion cues for comprehensive few-shot matching. To demonstrate the effectiveness, we evaluate MoLo on five standard benchmarks, and the results show that MoLo favorably outperforms recent advanced methods. The source code is available at https://github.com/alibaba-mmai-research/MoLo.

Most of the artificial lights fluctuate in response to the grid’s alternating current and exhibit subtle variations in terms of both intensity and spectrum, providing the potential to estimate the Electric Network Frequency (ENF) from conventional frame-based videos. Nevertheless, the performance of Video-based ENF (V-ENF) estimation largely relies on the imaging quality and thus may suffer from significant interference caused by non-ideal sampling, motion, and extreme lighting conditions. In this paper, we show that the ENF can be extracted without the above limitations from a new modality provided by the so-called event camera, a neuromorphic sensor that encodes the light intensity variations and asynchronously emits events with extremely high temporal resolution and high dynamic range. Specifically, we first formulate and validate the physical mechanism for the ENF captured in events, and then propose a simple yet robust Event-based ENF (E-ENF) estimation method through mode filtering and harmonic enhancement. Furthermore, we build an Event-Video ENF Dataset (EV-ENFD) that records both events and videos in diverse scenes. Extensive experiments on EV-ENFD demonstrate that our proposed E-ENF method can extract more accurate ENF traces, outperforming the conventional V-ENF by a large margin, especially in challenging environments with object motions and extreme lighting conditions. The code and dataset are available at https://github.com/xlx-creater/E-ENF.

Video Frame Interpolation (VFI) aims to generate intermediate video frames between consecutive input frames. Since the event cameras are bio-inspired sensors that only encode brightness changes with a micro-second temporal resolution, several works utilized the event camera to enhance the performance of VFI. However, existing methods estimate bidirectional inter-frame motion fields with only events or approximations, which can not consider the complex motion in real-world scenarios. In this paper, we propose a novel event-based VFI framework with cross-modal asymmetric bidirectional motion field estimation. In detail, our EIF-BiOFNet utilizes each valuable characteristic of the events and images for direct estimation of inter-frame motion fields without any approximation methods.Moreover, we develop an interactive attention-based frame synthesis network to efficiently leverage the complementary warping-based and synthesis-based features. Finally, we build a large-scale event-based VFI dataset, ERF-X170FPS, with a high frame rate, extreme motion, and dynamic textures to overcome the limitations of previous event-based VFI datasets. Extensive experimental results validate that our method shows significant performance improvement over the state-of-the-art VFI methods on various datasets.Our project pages are available at: https://github.com/intelpro/CBMNet

The performance of video frame interpolation is inherently correlated with the ability to handle motion in the input scene. Even though previous works recognize the utility of asynchronous event information for this task, they ignore the fact that motion may or may not result in blur in the input video to be interpolated, depending on the length of the exposure time of the frames and the speed of the motion, and assume either that the input video is sharp, restricting themselves to frame interpolation, or that it is blurry, including an explicit, separate deblurring stage before interpolation in their pipeline. We instead propose a general method for event-based frame interpolation that performs deblurring ad-hoc and thus works both on sharp and blurry input videos. Our model consists in a bidirectional recurrent network that naturally incorporates the temporal dimension of interpolation and fuses information from the input frames and the events adaptively based on their temporal proximity. In addition, we introduce a novel real-world high-resolution dataset with events and color videos which provides a challenging evaluation setting for the examined task. Extensive experiments on the standard GoPro benchmark and on our dataset show that our network consistently outperforms previous state-of-the-art methods on frame interpolation, single image deblurring and the joint task of interpolation and deblurring. Our code and dataset will be available at https://github.com/AHupuJR/REFID.

Video dehazing aims to recover haze-free frames with high visibility and contrast. This paper presents a novel framework to effectively explore the physical haze priors and aggregate temporal information. Specifically, we design a memory-based physical prior guidance module to encode the prior-related features into long-range memory. Besides, we formulate a multi-range scene radiance recovery module to capture space-time dependencies in multiple space-time ranges, which helps to effectively aggregate temporal information from adjacent frames. Moreover, we construct the first large-scale outdoor video dehazing benchmark dataset, which contains videos in various real-world scenarios. Experimental results on both synthetic and real conditions show the superiority of our proposed method.

Optical flow is an indispensable building block for various important computer vision tasks, including motion estimation, object tracking, and disparity measurement. In this work, we propose TransFlow, a pure transformer architecture for optical flow estimation. Compared to dominant CNN-based methods, TransFlow demonstrates three advantages. First, it provides more accurate correlation and trustworthy matching in flow estimation by utilizing spatial self-attention and cross-attention mechanisms between adjacent frames to effectively capture global dependencies; Second, it recovers more compromised information (e.g., occlusion and motion blur) in flow estimation through long-range temporal association in dynamic scenes; Third, it enables a concise self-learning paradigm and effectively eliminate the complex and laborious multi-stage pre-training procedures. We achieve the state-of-the-art results on the Sintel, KITTI-15, as well as several downstream tasks, including video object detection, interpolation and stabilization. For its efficacy, we hope TransFlow could serve as a flexible baseline for optical flow estimation.

We present a mask-piloted Transformer which improves masked-attention in Mask2Former for image segmentation. The improvement is based on our observation that Mask2Former suffers from inconsistent mask predictions between consecutive decoder layers, which leads to inconsistent optimization goals and low utilization of decoder queries. To address this problem, we propose a mask-piloted training approach, which additionally feeds noised ground-truth masks in masked-attention and trains the model to reconstruct the original ones. Compared with the predicted masks used in mask-attention, the ground-truth masks serve as a pilot and effectively alleviate the negative impact of inaccurate mask predictions in Mask2Former. Based on this technique, our MP-Former achieves a remarkable performance improvement on all three image segmentation tasks (instance, panoptic, and semantic), yielding +2.3 AP and +1.6 mIoU on the Cityscapes instance and semantic segmentation tasks with a ResNet-50 backbone. Our method also significantly speeds up the training, outperforming Mask2Former with half of the number of training epochs on ADE20K with both a ResNet-50 and a Swin-L backbones. Moreover, our method only introduces little computation during training and no extra computation during inference. Our code will be released at https://github.com/IDEA-Research/MP-Former.

We present an approach to learning regular spatial transformations between image pairs in the context of medical image registration. Contrary to optimization-based registration techniques and many modern learning-based methods, we do not directly penalize transformation irregularities but instead promote transformation regularity via an inverse consistency penalty. We use a neural network to predict a map between a source and a target image as well as the map when swapping the source and target images. Different from existing approaches, we compose these two resulting maps and regularize deviations of the Jacobian of this composition from the identity matrix. This regularizer -- GradICON -- results in much better convergence when training registration models compared to promoting inverse consistency of the composition of maps directly while retaining the desirable implicit regularization effects of the latter. We achieve state-of-the-art registration performance on a variety of real-world medical image datasets using a single set of hyperparameters and a single non-dataset-specific training protocol. The code is available at https://github.com/uncbiag/ICON.

Markov random fields (MRFs) are the cornerstone of classical approaches to example-based texture synthesis. Yet, it is not fully valued in the deep learning era. This paper aims to re-promote the combination of MRFs and neural networks, i.e., the CNNMRF model, for texture synthesis, with two key observations made. We first propose to compute the Guided Correspondence Distance in the nearest neighbor search, based on which a Guided Correspondence loss is defined to measure the similarity of the output texture to the example. Experiments show that our approach surpasses existing neural approaches in uncontrolled and controlled texture synthesis. More importantly, the Guided Correspondence loss can function as a general textural loss in, e.g., training generative networks for real-time controlled synthesis and inversion-based single-image editing. In contrast, existing textural losses, such as the Sliced Wasserstein loss, cannot work on these challenging tasks.

Many deep learning-based solutions to blind image deblurring estimate the blur representation and reconstruct the target image from its blurry observation. However, these methods suffer from severe performance degradation in real-world scenarios because they ignore important prior information about motion blur (e.g., real-world motion blur is diverse and spatially varying). Some methods have attempted to explicitly estimate non-uniform blur kernels by CNNs, but accurate estimation is still challenging due to the lack of ground truth about spatially varying blur kernels in real-world images. To address these issues, we propose to represent the field of motion blur kernels in a latent space by normalizing flows, and design CNNs to predict the latent codes instead of motion kernels. To further improve the accuracy and robustness of non-uniform kernel estimation, we introduce uncertainty learning into the process of estimating latent codes and propose a multi-scale kernel attention module to better integrate image features with estimated kernels. Extensive experimental results, especially on real-world blur datasets, demonstrate that our method achieves state-of-the-art results in terms of both subjective and objective quality as well as excellent generalization performance for non-uniform image deblurring. The code is available at https://see.xidian.edu.cn/faculty/wsdong/Projects/UFPNet.htm.

The images captured under improper exposure conditions often suffer from contrast degradation and detail distortion. Contrast degradation will destroy the statistical properties of low-frequency components, while detail distortion will disturb the structural properties of high-frequency components, leading to the low-frequency and high-frequency components being mixed and inseparable. This will limit the statistical and structural modeling capacity for exposure correction. To address this issue, this paper proposes to decouple the contrast enhancement and detail restoration within each convolution process. It is based on the observation that, in the local regions covered by convolution kernels, the feature response of low-/high-frequency can be decoupled by addition/difference operation. To this end, we inject the addition/difference operation into the convolution process and devise a Contrast Aware (CA) unit and a Detail Aware (DA) unit to facilitate the statistical and structural regularities modeling. The proposed CA and DA can be plugged into existing CNN-based exposure correction networks to substitute the Traditional Convolution (TConv) to improve the performance. Furthermore, to maintain the computational costs of the network without changing, we aggregate two units into a single TConv kernel using structural re-parameterization. Evaluations of nine methods and five benchmark datasets demonstrate that our proposed method can comprehensively improve the performance of existing methods without introducing extra computational costs compared with the original networks. The codes will be publicly available.

Images captured in low-light conditions often suffer from significant quality degradation. Recent works have built a large variety of deep Retinex-based networks to enhance low-light images. The Retinex-based methods require decomposing the image into reflectance and illumination components, which is a highly ill-posed problem and there is no available ground truth. Previous works addressed this problem by imposing some additional priors or regularizers. However, finding an effective prior or regularizer that can be applied in various scenes is challenging, and the performance of the model suffers from too many additional constraints. We propose a contrastive learning method and a self-knowledge distillation method that allow training our Retinex-based model for Retinex decomposition without elaborate hand-crafted regularization functions. Rather than estimating reflectance and illuminance images and representing the final images as their element-wise products as in previous works, our regularizer-free Retinex decomposition and synthesis network (RFR) extracts reflectance and illuminance features and synthesizes them end-to-end. In addition, we propose a loss function for contrastive learning and a progressive learning strategy for self-knowledge distillation. Extensive experimental results demonstrate that our proposed methods can achieve superior performance compared with state-of-the-art approaches.

The exclusive properties of RAW data have shown great potential for low-light image enhancement. Nevertheless, the performance is bottlenecked by the inherent limitations of existing architectures in both single-stage and multi-stage methods. Mixed mapping across two different domains, noise-to-clean and RAW-to-sRGB, misleads the single-stage methods due to the domain ambiguity. The multi-stage methods propagate the information merely through the resulting image of each stage, neglecting the abundant features in the lossy image-level dataflow. In this paper, we probe a generalized solution to these bottlenecks and propose a Decouple aNd Feedback framework, abbreviated as DNF. To mitigate the domain ambiguity, domainspecific subtasks are decoupled, along with fully utilizing the unique properties in RAW and sRGB domains. The feature propagation across stages with a feedback mechanism avoids the information loss caused by image-level dataflow. The two key insights of our method resolve the inherent limitations of RAW data-based low-light image enhancement satisfactorily, empowering our method to outperform the previous state-of-the-art method by a large margin with only 19% parameters, achieving 0.97dB and 1.30dB PSNR improvements on the Sony and Fuji subsets of SID.

Despite the remarkable achievement of recent underwater image restoration techniques, the lack of labeled data has become a major hurdle for further progress. In this work, we propose a mean-teacher based Semi-supervised Underwater Image Restoration (Semi-UIR) framework to incorporate the unlabeled data into network training. However, the naive mean-teacher method suffers from two main problems: (1) The consistency loss used in training might become ineffective when the teacher’s prediction is wrong. (2) Using L1 distance may cause the network to overfit wrong labels, resulting in confirmation bias. To address the above problems, we first introduce a reliable bank to store the “best-ever” outputs as pseudo ground truth. To assess the quality of outputs, we conduct an empirical analysis based on the monotonicity property to select the most trustworthy NR-IQA method. Besides, in view of the confirmation bias problem, we incorporate contrastive regularization to prevent the overfitting on wrong labels. Experimental results on both full-reference and non-reference underwater benchmarks demonstrate that our algorithm has obvious improvement over SOTA methods quantitatively and qualitatively. Code has been released at https://github.com/Huang-ShiRui/Semi-UIR.

Despite the significant results on synthetic noise under simplified assumptions, most self-supervised denoising methods fail under real noise due to the strong spatial noise correlation, including the advanced self-supervised blind-spot networks (BSNs). For recent methods targeting real-world denoising, they either suffer from ignoring this spatial correlation, or are limited by the destruction of fine textures for under-considering the correlation. In this paper, we present a novel method called LG-BPN for self-supervised real-world denoising, which takes the spatial correlation statistic into our network design for local detail restoration, and also brings the long-range dependencies modeling ability to previously CNN-based BSN methods. First, based on the correlation statistic, we propose a densely-sampled patch-masked convolution module. By taking more neighbor pixels with low noise correlation into account, we enable a denser local receptive field, preserving more useful information for enhanced fine structure recovery. Second, we propose a dilated Transformer block to allow distant context exploitation in BSN. This global perception addresses the intrinsic deficiency of BSN, whose receptive field is constrained by the blind spot requirement, which can not be fully resolved by the previous CNN-based BSNs. These two designs enable LG-BPN to fully exploit both the detailed structure and the global interaction in a blind manner. Extensive results on real-world datasets demonstrate the superior performance of our method. https://github.com/Wang-XIaoDingdd/LGBPN

Recently deep learning techniques have significantly advanced image super-resolution (SR). Due to the black-box nature, quantifying reconstruction uncertainty is crucial when employing these deep SR networks. Previous approaches for SR uncertainty estimation mostly focus on capturing pixel-wise uncertainty in the spatial domain. SR uncertainty in the frequency domain which is highly related to image SR is seldom explored. In this paper, we propose to quantify spectral Bayesian uncertainty in image SR. To achieve this, a Dual-Domain Learning (DDL) framework is first proposed. Combined with Bayesian approaches, the DDL model is able to estimate spectral uncertainty accurately, enabling a reliability assessment for high frequencies reasoning from the frequency domain perspective. Extensive experiments under non-ideal premises are conducted and demonstrate the effectiveness of the proposed spectral uncertainty. Furthermore, we propose a novel Spectral Uncertainty based Decoupled Frequency (SUDF) training scheme for perceptual SR. Experimental results show the proposed SUDF can evidently boost perceptual quality of SR results without sacrificing much pixel accuracy.

Deep image prior (DIP) has shown great promise in tackling a variety of image restoration (IR) and general visual inverse problems, needing no training data. However, the resulting optimization process is often very slow, inevitably hindering DIP’s practical usage for time-sensitive scenarios. In this paper, we focus on IR, and propose two crucial modifications to DIP that help achieve substantial speedup: 1) optimizing the DIP seed while freezing randomly-initialized network weights, and 2) reducing the network depth. In addition, we reintroduce explicit priors, such as sparse gradient prior---encoded by total-variation regularization, to preserve the DIP peak performance. We evaluate the proposed method on three IR tasks, including image denoising, image super-resolution, and image inpainting, against the original DIP and variants, as well as the competing metaDIP that uses meta-learning to learn good initializers with extra data. Our method is a clear winner in obtaining competitive restoration quality in a minimal amount of time. Our code is available at https://github.com/sun-umn/Deep-Random-Projector.

In this paper, we study Blind Image Decomposition (BID), which is to uniformly remove multiple types of degradation at once without foreknowing the noise type. There remain two practical challenges: (1) Existing methods typically require massive data supervision, making them infeasible to real-world scenarios. (2) The conventional paradigm usually focuses on mining the abnormal pattern of a superimposed image to separate the noise, which de facto conflicts with the primary image restoration task. Therefore, such a pipeline compromises repairing efficiency and authenticity. In an attempt to solve the two challenges in one go, we propose an efficient and simplified paradigm, called Context-aware Pretraining (CP), with two pretext tasks: mixed image separation and masked image reconstruction. Such a paradigm reduces the annotation demands and explicitly facilitates context-aware feature learning. Assuming the restoration process follows a structure-to-texture manner, we also introduce a Context-aware Pretrained network (CPNet). In particular, CPNet contains two transformer-based parallel encoders, one information fusion module, and one multi-head prediction module. The information fusion module explicitly utilizes the mutual correlation in the spatial-channel dimension, while the multi-head prediction module facilitates texture-guided appearance flow. Moreover, a new sampling loss along with an attribute label constraint is also deployed to make use of the spatial context, leading to high-fidelity image restoration. Extensive experiments on both real and synthetic benchmarks show that our method achieves competitive performance for various BID tasks.

Many low-level computer vision tasks are desirable to utilize the unprocessed RAW image as input, which remains the linear relationship between pixel values and scene radiance. Recent works advocate to embed the RAW image samples into sRGB images at capture time, and reconstruct the RAW from sRGB by these metadata when needed. However, there still exist some limitations on taking full use of the metadata. In this paper, instead of following the perspective of sRGB-to-RAW mapping, we reformulate the problem as mapping the 2D coordinates of the metadata to its RAW values conditioned on the corresponding sRGB values. With this novel formulation, we propose to reconstruct the RAW image with an implicit neural function, which achieves significant performance improvement (more than 10dB average PSNR) only with the uniform sampling. Compared with most deep learning-based approaches, our method is trained in a self-supervised way that requiring no pre-training on different camera ISPs. We perform further experiments to demonstrate the effectiveness of our method, and show that our framework is also suitable for the task of guided super-resolution.

While raw images exhibit advantages over sRGB images (e.g. linearity and fine-grained quantization level), they are not widely used by common users due to the large storage requirements. Very recent works propose to compress raw images by designing the sampling masks in the raw image pixel space, leading to suboptimal image representations and redundant metadata. In this paper, we propose a novel framework to learn a compact representation in the latent space serving as the metadata in an end-to-end manner. Furthermore, we propose a novel sRGB-guided context model with the improved entropy estimation strategies, which leads to better reconstruction quality, smaller size of metadata, and faster speed. We illustrate how the proposed raw image compression scheme can adaptively allocate more bits to image regions that are important from a global perspective. The experimental results show that the proposed method can achieve superior raw image reconstruction results using a smaller size of the metadata on both uncompressed sRGB images and JPEG images.

Recently, deep neural networks have been successfully applied for image restoration (IR) (e.g., super-resolution, de-noising, de-blurring). Despite their promising performance, running IR networks requires heavy computation. A large body of work has been devoted to addressing this issue by designing novel neural networks or pruning their parameters. However, the common limitation is that while images are saved in a compressed format before being enhanced by IR, prior work does not consider the impact of compression on the IR quality. In this paper, we present AccelIR, a framework that optimizes image compression considering the end-to-end pipeline of IR tasks. AccelIR encodes an image through IR-aware compression that optimizes compression levels across image blocks within an image according to the impact on the IR quality. Then, it runs a lightweight IR network on the compressed image, effectively reducing IR computation, while maintaining the same IR quality and image size. Our extensive evaluation using seven IR networks shows that AccelIR can reduce the computing overhead of super-resolution, de-nosing, and de-blurring by 49%, 29%, and 32% on average, respectively

Real world images often have highly imbalanced content density. Some areas are very uniform, e.g., large patches of blue sky, while other areas are scattered with many small objects. Yet, the commonly used successive grid downsampling strategy in convolutional deep networks treats all areas equally. Hence, small objects are represented in very few spatial locations, leading to worse results in tasks such as segmentation. Intuitively, retaining more pixels representing small objects during downsampling helps to preserve important information. To achieve this, we propose AutoFocusFormer (AFF), a local-attention transformer image recognition backbone, which performs adaptive downsampling by learning to retain the most important pixels for the task. Since adaptive downsampling generates a set of pixels irregularly distributed on the image plane, we abandon the classic grid structure. Instead, we develop a novel point-based local attention block, facilitated by a balanced clustering module and a learnable neighborhood merging module, which yields representations for our point-based versions of state-of-the-art segmentation heads. Experiments show that our AutoFocusFormer (AFF) improves significantly over baseline models of similar sizes.

Performing super-resolution of a depth image using the guidance from an RGB image is a problem that concerns several fields, such as robotics, medical imaging, and remote sensing. While deep learning methods have achieved good results in this problem, recent work highlighted the value of combining modern methods with more formal frameworks. In this work we propose a novel approach which combines guided anisotropic diffusion with a deep convolutional network and advances the state of the art for guided depth super-resolution. The edge transferring/enhancing properties of the diffusion are boosted by the contextual reasoning capabilities of modern networks, and a strict adjustment step guarantees perfect adherence to the source image. We achieve unprecedented results in three commonly used benchmarks for guided depth super resolution. The performance gain compared to other methods is the largest at larger scales, such as x32 scaling. Code for the proposed method will be made available to promote reproducibility of our results.

We propose Super-resolution Neural Operator (SRNO), a deep operator learning framework that can resolve high-resolution (HR) images at arbitrary scales from the low-resolution (LR) counterparts. Treating the LR-HR image pairs as continuous functions approximated with different grid sizes, SRNO learns the mapping between the corresponding function spaces. From the perspective of approximation theory, SRNO first embeds the LR input into a higher-dimensional latent representation space, trying to capture sufficient basis functions, and then iteratively approximates the implicit image function with a kernel integral mechanism, followed by a final dimensionality reduction step to generate the RGB representation at the target coordinates. The key characteristics distinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efficiently implemented via the Galerkin-type attention, which possesses non-local properties in the spatial domain and therefore benefits the grid-free continuum; and 2) the multilayer attention architecture allows for the dynamic latent basis update, which is crucial for SR problems to “hallucinate” high-frequency information from the LR image. Experiments show that SRNO outperforms existing continuous SR methods in terms of both accuracy and running time. Our code is at https://github.com/2y7c3/Super-Resolution-Neural-Operator.

Implicit neural representation demonstrates promising ability in representing images with arbitrary resolutions recently. In this paper, we present Local Implicit Transformer (LIT) that integrates attention mechanism and frequency encoding technique into local implicit image function. We design a cross-scale local attention block to effectively aggregate local features and a local frequency encoding block to combine positional encoding with Fourier domain information for constructing high-resolution (HR) images. To further improve representative power, we propose Cascaded LIT (CLIT) exploiting multi-scale features along with cumulative training strategy that gradually increase the upsampling factors for training. We have performed extensive experiments to validate the effectiveness of these components and analyze the variants of the training strategy. The qualitative and quantitative results demonstrated that LIT and CLIT achieve favorable results and outperform the previous works within arbitrary super-resolution tasks.

Cameras and image-editing software often process images in the wide-gamut ProPhoto color space, encompassing 90% of all visible colors. However, when images are encoded for sharing, this color-rich representation is transformed and clipped to fit within the small-gamut standard RGB (sRGB) color space, representing only 30% of visible colors. Recovering the lost color information is challenging due to the clipping procedure. Inspired by neural implicit representations for 2D images, we propose a method that optimizes a lightweight multi-layer-perceptron (MLP) model during the gamut reduction step to predict the clipped values. GamutMLP takes approximately 2 seconds to optimize and requires only 23 KB of storage. The small memory footprint allows our GamutMLP model to be saved as metadata in the sRGB image---the model can be extracted when needed to restore wide-gamut color values. We demonstrate the effectiveness of our approach for color recovery and compare it with alternative strategies, including pre-trained DNN-based gamut expansion networks and other implicit neural representation methods. As part of this effort, we introduce a new color gamut dataset of 2200 wide-gamut/small-gamut images for training and testing.

The aim of this paper is to propose a mechanism to efficiently and explicitly model image hierarchies in the global, regional, and local range for image restoration. To achieve that, we start by analyzing two important properties of natural images including cross-scale similarity and anisotropic image features. Inspired by that, we propose the anchored stripe self-attention which achieves a good balance between the space and time complexity of self-attention and the modelling capacity beyond the regional range. Then we propose a new network architecture dubbed GRL to explicitly model image hierarchies in the Global, Regional, and Local range via anchored stripe self-attention, window self-attention, and channel attention enhanced convolution. Finally, the proposed network is applied to 7 image restoration types, covering both real and synthetic settings. The proposed method sets the new state-of-the-art for several of those. Code will be available at https://github.com/ofsoundof/GRL-Image-Restoration.git.

We present a simple yet effective self-supervised pretraining method for image harmonization which can leverage large-scale unannotated image datasets. To achieve this goal, we first generate pre-training data online with our Label-Efficient Masked Region Transform (LEMaRT) pipeline. Given an image, LEMaRT generates a foreground mask and then applies a set of transformations to perturb various visual attributes, e.g., defocus blur, contrast, saturation, of the region specified by the generated mask. We then pre-train image harmonization models by recovering the original image from the perturbed image. Secondly, we introduce an image harmonization model, namely SwinIH, by retrofitting the Swin Transformer [27] with a combination of local and global self-attention mechanisms. Pretraining SwinIH with LEMaRT results in a new state of the art for image harmonization, while being label-efficient, i.e., consuming less annotated data for fine-tuning than existing methods. Notably, on iHarmony4 dataset [8], SwinIH outperforms the state of the art, i.e., SCS-Co [16] by a margin of 0.4 dB when it is fine-tuned on only 50% of the training data, and by 1.0 dB when it is trained on the full training dataset.

Content affinity loss including feature and pixel affinity is a main problem which leads to artifacts in photorealistic and video style transfer. This paper proposes a new framework named CAP-VSTNet, which consists of a new reversible residual network and an unbiased linear transform module, for versatile style transfer. This reversible residual network can not only preserve content affinity but not introduce redundant information as traditional reversible networks, and hence facilitate better stylization. Empowered by Matting Laplacian training loss which can address the pixel affinity loss problem led by the linear transform, the proposed framework is applicable and effective on versatile style transfer. Extensive experiments show that CAP-VSTNet can produce better qualitative and quantitative results in comparison with the state-of-the-art methods.

Object compositing based on 2D images is a challenging problem since it typically involves multiple processing stages such as color harmonization, geometry correction and shadow generation to generate realistic results. Furthermore, annotating training data pairs for compositing requires substantial manual effort from professionals, and is hardly scalable. Thus, with the recent advances in generative models, in this work, we propose a self-supervised framework for object compositing by leveraging the power of conditional diffusion models. Our framework can hollistically address the object compositing task in a unified model, transforming the viewpoint, geometry, color and shadow of the generated object while requiring no manual labeling. To preserve the input object’s characteristics, we introduce a content adaptor that helps to maintain categorical semantics and object appearance. A data augmentation method is further adopted to improve the fidelity of the generator. Our method outperforms relevant baselines in both realism and faithfulness of the synthesized result images in a user study on various real-world images.

Vector font synthesis is a challenging and ongoing problem in the fields of Computer Vision and Computer Graphics. The recently-proposed DeepVecFont achieved state-of-the-art performance by exploiting information of both the image and sequence modalities of vector fonts. However, it has limited capability for handling long sequence data and heavily relies on an image-guided outline refinement post-processing. Thus, vector glyphs synthesized by DeepVecFont still often contain some distortions and artifacts and cannot rival human-designed results. To address the above problems, this paper proposes an enhanced version of DeepVecFont mainly by making the following three novel technical contributions. First, we adopt Transformers instead of RNNs to process sequential data and design a relaxation representation for vector outlines, markedly improving the model’s capability and stability of synthesizing long and complex outlines. Second, we propose to sample auxiliary points in addition to control points to precisely align the generated and target Bézier curves or lines. Finally, to alleviate error accumulation in the sequential generation process, we develop a context-based self-refinement module based on another Transformer-based decoder to remove artifacts in the initially synthesized glyphs. Both qualitative and quantitative results demonstrate that the proposed method effectively resolves those intrinsic problems of the original DeepVecFont and outperforms existing approaches in generating English and Chinese vector fonts with complicated structures and diverse styles.

Transformer-based models achieve favorable performance in artistic style transfer recently thanks to its global receptive field and powerful multi-head/layer attention operations. Nevertheless, the over-paramerized multi-layer structure increases parameters significantly and thus presents a heavy burden for training. Moreover, for the task of style transfer, vanilla Transformer that fuses content and style features by residual connections is prone to content-wise distortion. In this paper, we devise a novel Transformer model termed as Master specifically for style transfer. On the one hand, in the proposed model, different Transformer layers share a common group of parameters, which (1) reduces the total number of parameters, (2) leads to more robust training convergence, and (3) is readily to control the degree of stylization via tuning the number of stacked layers freely during inference. On the other hand, different from the vanilla version, we adopt a learnable scaling operation on content features before content-style feature interaction, which better preserves the original similarity between a pair of content features while ensuring the stylization quality. We also propose a novel meta learning scheme for the proposed model so that it can not only work in the typical setting of arbitrary style transfer, but also adaptable to the few-shot setting, by only fine-tuning the Transformer encoder layer in the few-shot stage for one specific style. Text-guided few-shot style transfer is firstly achieved with the proposed framework. Extensive experiments demonstrate the superiority of Master under both zero-shot and few-shot style transfer settings.

Recent works have demonstrated that natural language can be used to generate and edit 3D shapes. However, these methods generate shapes with limited fidelity and diversity. We introduce CLIP-Sculptor, a method to address these constraints by producing high-fidelity and diverse 3D shapes without the need for (text, shape) pairs during training. CLIP-Sculptor achieves this in a multi-resolution approach that first generates in a low-dimensional latent space and then upscales to a higher resolution for improved shape fidelity. For improved shape diversity, we use a discrete latent space which is modeled using a transformer conditioned on CLIP’s image-text embedding space. We also present a novel variant of classifier-free guidance, which improves the accuracy-diversity trade-off. Finally, we perform extensive experiments demonstrating that CLIP-Sculptor outperforms state-of-the-art baselines.

Automatic layout generation that can synthesize high-quality layouts is an important tool for graphic design in many applications. Though existing methods based on generative models such as Generative Adversarial Networks (GANs) and Variational Auto-Encoders (VAEs) have progressed, they still leave much room for improving the quality and diversity of the results. Inspired by the recent success of diffusion models in generating high-quality images, this paper explores their potential for conditional layout generation and proposes Transformer-based Layout Diffusion Model (LayoutDM) by instantiating the conditional denoising diffusion probabilistic model (DDPM) with a purely transformer-based architecture. Instead of using convolutional neural networks, a transformer-based conditional Layout Denoiser is proposed to learn the reverse diffusion process to generate samples from noised layout data. Benefitting from both transformer and DDPM, our LayoutDM is of desired properties such as high-quality generation, strong sample diversity, faithful distribution coverage, and stationary training in comparison to GANs and VAEs. Quantitative and qualitative experimental results show that our method outperforms state-of-the-art generative models in terms of quality and diversity.

Text-guided image editing can have a transformative impact in supporting creative applications. A key challenge is to generate edits that are faithful to the input text prompt, while consistent with the input image. We present Imagen Editor, a cascaded diffusion model, built by fine-tuning Imagen on text-guided image inpainting. Imagen Editor’s edits are faithful to the text prompts, which is accomplished by incorporating object detectors for proposing inpainting masks during training. In addition, text-guided image inpainting captures fine details in the input image by conditioning the cascaded pipeline on the original high resolution image. To improve qualitative and quantitative evaluation, we introduce EditBench, a systematic benchmark for text-guided image inpainting. EditBench evaluates inpainting edits on natural and generated images exploring objects, attributes, and scenes. Through extensive human evaluation on EditBench, we find that object-masking during training leads to across-the-board improvements in text-image alignment -- such that Imagen Editor is preferred over DALL-E 2 and Stable Diffusion -- and, as a cohort, these models are better at object-rendering than text-rendering, and handle material/color/size attributes better than count/shape attributes.

Recent text-to-image diffusion models are able to generate convincing results of unprecedented quality. However, it is nearly impossible to control the shapes of different regions/objects or their layout in a fine-grained fashion. Previous attempts to provide such controls were hindered by their reliance on a fixed set of labels. To this end, we present SpaText --- a new method for text-to-image generation using open-vocabulary scene control. In addition to a global text prompt that describes the entire scene, the user provides a segmentation map where each region of interest is annotated by a free-form natural language description. Due to lack of large-scale datasets that have a detailed textual description for each region in the image, we choose to leverage the current large-scale text-to-image datasets and base our approach on a novel CLIP-based spatio-textual representation, and show its effectiveness on two state-of-the-art diffusion models: pixel-based and latent-based. In addition, we show how to extend the classifier-free guidance method in diffusion models to the multi-conditional case and present an alternative accelerated inference algorithm. Finally, we offer several automatic evaluation metrics and use them, in addition to FID scores and a user study, to evaluate our method and show that it achieves state-of-the-art results on image generation with free-form textual scene control.

Language-guided image editing has achieved great success recently. In this paper, we investigate exemplar-guided image editing for more precise control. We achieve this goal by leveraging self-supervised training to disentangle and re-organize the source image and the exemplar. However, the naive approach will cause obvious fusing artifacts. We carefully analyze it and propose an information bottleneck and strong augmentations to avoid the trivial solution of directly copying and pasting the exemplar image. Meanwhile, to ensure the controllability of the editing process, we design an arbitrary shape mask for the exemplar image and leverage the classifier-free guidance to increase the similarity to the exemplar image. The whole framework involves a single forward of the diffusion model without any iterative optimization. We demonstrate that our method achieves an impressive performance and enables controllable editing on in-the-wild images with high fidelity.

We propose a method for editing images from human instructions: given an input image and a written instruction that tells the model what to do, our model follows these instructions to edit the image. To obtain training data for this problem, we combine the knowledge of two large pretrained models--a language model (GPT-3) and a text-to-image model (Stable Diffusion)--to generate a large dataset of image editing examples. Our conditional diffusion model, InstructPix2Pix, is trained on our generated data, and generalizes to real images and user-written instructions at inference time. Since it performs edits in the forward pass and does not require per-example fine-tuning or inversion, our model edits images quickly, in a matter of seconds. We show compelling editing results for a diverse collection of input images and written instructions.

Conditional graphic layout generation, which generates realistic layouts according to user constraints, is a challenging task that has not been well-studied yet. First, there is limited discussion about how to handle diverse user constraints flexibly and uniformly. Second, to make the layouts conform to user constraints, existing work often sacrifices generation quality significantly. In this work, we propose LayoutFormer++ to tackle the above problems. First, to flexibly handle diverse constraints, we propose a constraint serialization scheme, which represents different user constraints as sequences of tokens with a predefined format. Then, we formulate conditional layout generation as a sequence-to-sequence transformation, and leverage encoder-decoder framework with Transformer as the basic architecture. Furthermore, to make the layout better meet user requirements without harming quality, we propose a decoding space restriction strategy. Specifically, we prune the predicted distribution by ignoring the options that definitely violate user constraints and likely result in low-quality layouts, and make the model samples from the restricted distribution. Experiments demonstrate that LayoutFormer++ outperforms existing approaches on all the tasks in terms of both better generation quality and less constraint violation.

Diffusion models have demonstrated remarkable progress in image generation quality, especially when guidance is used to control the generative process. However, guidance requires a large amount of image-annotation pairs for training and is thus dependent on their availability and correctness. In this paper, we eliminate the need for such annotation by instead exploiting the flexibility of self-supervision signals to design a framework for self-guided diffusion models. By leveraging a feature extraction function and a self-annotation function, our method provides guidance signals at various image granularities: from the level of holistic images to object boxes and even segmentation masks. Our experiments on single-label and multi-label image datasets demonstrate that self-labeled guidance always outperforms diffusion models without guidance and may even surpass guidance based on ground-truth labels. When equipped with self-supervised box or mask proposals, our method further generates visually diverse yet semantically consistent images, without the need for any class, box, or segment label annotation. Self-guided diffusion is simple, flexible and expected to profit from deployment at scale.

Diffusion models have emerged as the best approach for generative modeling of 2D images. Part of their success is due to the possibility of training them on millions if not billions of images with a stable learning objective. However, extending these models to 3D remains difficult for two reasons. First, finding a large quantity of 3D training data is much more complex than for 2D images. Second, while it is conceptually trivial to extend the models to operate on 3D rather than 2D grids, the associated cubic growth in memory and compute complexity makes this infeasible. We address the first challenge by introducing a new diffusion setup that can be trained, end-to-end, with only posed 2D images for supervision; and the second challenge by proposing an image formation model that decouples model memory from spatial memory. We evaluate our method on real-world data, using the CO3D dataset which has not been used to train 3D generative models before. We show that our diffusion models are scalable, train robustly, and are competitive in terms of sample quality and fidelity to existing approaches for 3D generative modeling.

Diffusion-based models have shown the merits of generating high-quality visual data while preserving better diversity in recent studies. However, such observation is only justified with curated data distribution, where the data samples are nicely pre-processed to be uniformly distributed in terms of their labels. In practice, a long-tailed data distribution appears more common and how diffusion models perform on such class-imbalanced data remains unknown. In this work, we first investigate this problem and observe significant degradation in both diversity and fidelity when the diffusion model is trained on datasets with class-imbalanced distributions. Especially in tail classes, the generations largely lose diversity and we observe severe mode-collapse issues. To tackle this problem, we set from the hypothesis that the data distribution is not class-balanced, and propose Class-Balancing Diffusion Models (CBDM) that are trained with a distribution adjustment regularizer as a solution. Experiments show that images generated by CBDM exhibit higher diversity and quality in both quantitative and qualitative ways. Our method benchmarked the generation results on CIFAR100/CIFAR100LT dataset and shows outstanding performance on the downstream recognition task.

Conditional image-to-video (cI2V) generation aims to synthesize a new plausible video starting from an image (e.g., a person’s face) and a condition (e.g., an action class label like smile). The key challenge of the cI2V task lies in the simultaneous generation of realistic spatial appearance and temporal dynamics corresponding to the given image and condition. In this paper, we propose an approach for cI2V using novel latent flow diffusion models (LFDM) that synthesize an optical flow sequence in the latent space based on the given condition to warp the given image. Compared to previous direct-synthesis-based works, our proposed LFDM can better synthesize spatial details and temporal motion by fully utilizing the spatial content of the given image and warping it in the latent space according to the generated temporally-coherent flow. The training of LFDM consists of two separate stages: (1) an unsupervised learning stage to train a latent flow auto-encoder for spatial content generation, including a flow predictor to estimate latent flow between pairs of video frames, and (2) a conditional learning stage to train a 3D-UNet-based diffusion model (DM) for temporal latent flow generation. Unlike previous DMs operating in pixel space or latent feature space that couples spatial and temporal information, the DM in our LFDM only needs to learn a low-dimensional latent flow space for motion generation, thus being more computationally efficient. We conduct comprehensive experiments on multiple datasets, where LFDM consistently outperforms prior arts. Furthermore, we show that LFDM can be easily adapted to new domains by simply finetuning the image decoder. Our code is available at https://github.com/nihaomiao/CVPR23_LFDM.

Despite the remarkable progress in deep generative models, synthesizing high-resolution and temporally coherent videos still remains a challenge due to their high-dimensionality and complex temporal dynamics along with large spatial variations. Recent works on diffusion models have shown their potential to solve this challenge, yet they suffer from severe computation- and memory-inefficiency that limit the scalability. To handle this issue, we propose a novel generative model for videos, coined projected latent video diffusion models (PVDM), a probabilistic diffusion model which learns a video distribution in a low-dimensional latent space and thus can be efficiently trained with high-resolution videos under limited resources. Specifically, PVDM is composed of two components: (a) an autoencoder that projects a given video as 2D-shaped latent vectors that factorize the complex cubic structure of video pixels and (b) a diffusion model architecture specialized for our new factorized latent space and the training/sampling procedure to synthesize videos of arbitrary length with a single model. Experiments on popular video generation datasets demonstrate the superiority of PVDM compared with previous video synthesis methods; e.g., PVDM obtains the FVD score of 639.7 on the UCF-101 long video (128 frames) generation benchmark, which improves 1773.4 of the prior state-of-the-art.

Quantizing images into discrete representations has been a fundamental problem in unified generative modeling. Predominant approaches learn the discrete representation either in a deterministic manner by selecting the best-matching token or in a stochastic manner by sampling from a predicted distribution. However, deterministic quantization suffers from severe codebook collapse and misaligned inference stage while stochastic quantization suffers from low codebook utilization and perturbed reconstruction objective. This paper presents a regularized vector quantization framework that allows to mitigate above issues effectively by applying regularization from two perspectives. The first is a prior distribution regularization which measures the discrepancy between a prior token distribution and predicted token distribution to avoid codebook collapse and low codebook utilization. The second is a stochastic mask regularization that introduces stochasticity during quantization to strike a good balance between inference stage misalignment and unperturbed reconstruction objective. In addition, we design a probabilistic contrastive loss which serves as a calibrated metric to further mitigate the perturbed reconstruction objective. Extensive experiments show that the proposed quantization framework outperforms prevailing vector quantizers consistently across different generative models including auto-regressive models and diffusion models.

Video snapshot compressive imaging (SCI) uses a two-dimensional detector to capture consecutive video frames during a single exposure time. Following this, an efficient reconstruction algorithm needs to be designed to reconstruct the desired video frames. Although recent deep learning-based state-of-the-art (SOTA) reconstruction algorithms have achieved good results in most tasks, they still face the following challenges due to excessive model complexity and GPU memory limitations: 1) these models need high computational cost, and 2) they are usually unable to reconstruct large-scale video frames at high compression ratios. To address these issues, we develop an efficient network for video SCI by using dense connections and space-time factorization mechanism within a single residual block, dubbed EfficientSCI. The EfficientSCI network can well establish spatial-temporal correlation by using convolution in the spatial domain and Transformer in the temporal domain, respectively. We are the first time to show that an UHD color video with high compression ratio can be reconstructed from a snapshot 2D measurement using a single end-to-end deep learning model with PSNR above 32 dB. Extensive results on both simulation and real data show that our method significantly outperforms all previous SOTA algorithms with better real-time performance.

Learning-based video compression has been extensively studied over the past years, but it still has limitations in adapting to various motion patterns and entropy models. In this paper, we propose multi-mode video compression (MMVC), a block wise mode ensemble deep video compression framework that selects the optimal mode for feature domain prediction adapting to different motion patterns. Proposed multi-modes include ConvLSTM-based feature domain prediction, optical flow conditioned feature domain prediction, and feature propagation to address a wide range of cases from static scenes without apparent motions to dynamic scenes with a moving camera. We partition the feature space into blocks for temporal prediction in spatial block-based representations. For entropy coding, we consider both dense and sparse post-quantization residual blocks, and apply optional run-length coding to sparse residuals to improve the compression rate. In this sense, our method uses a dual-mode entropy coding scheme guided by a binary density map, which offers significant rate reduction surpassing the extra cost of transmitting the binary selection map. We validate our scheme with some of the most popular benchmarking datasets. Compared with state-of-the-art video compression schemes and standard codecs, our method yields better or competitive results measured with PSNR and MS-SSIM.

Encoding videos as neural networks is a recently proposed approach that allows new forms of video processing. However, traditional techniques still outperform such neural video representation (NVR) methods for the task of video compression. This performance gap can be explained by the fact that current NVR methods: i) use architectures that do not efficiently obtain a compact representation of temporal and spatial information; and ii) minimize rate and distortion disjointly (first overfitting a network on a video and then using heuristic techniques such as post-training quantization or weight pruning to compress the model). We propose a novel convolutional architecture for video representation that better represents spatio-temporal information and a training strategy capable of jointly optimizing rate and distortion. All network and quantization parameters are jointly learned end-to-end, and the post-training operations used in previous works are unnecessary. We evaluate our method on the UVG dataset, achieving new state-of-the-art results for video compression with NVRs. Moreover, we deliver the first NVR-based video compression method that improves over the typically adopted HEVC benchmark (x265, disabled b-frames, “medium” preset), closing the gap to autoencoder-based video compression techniques.

Implicit neural representations (INRs) have recently advanced numerous vision-related areas. INR performance depends strongly on the choice of activation function employed in its MLP network. A wide range of nonlinearities have been explored, but, unfortunately, current INRs designed to have high accuracy also suffer from poor robustness (to signal noise, parameter variation, etc.). Inspired by harmonic analysis, we develop a new, highly accurate and robust INR that does not exhibit this tradeoff. Our Wavelet Implicit neural REpresentation (WIRE) uses as its activation function the complex Gabor wavelet that is well-known to be optimally concentrated in space--frequency and to have excellent biases for representing images. A wide range of experiments (image denoising, image inpainting, super-resolution, computed tomography reconstruction, image overfitting, and novel view synthesis with neural radiance fields) demonstrate that WIRE defines the new state of the art in INR accuracy, training time, and robustness.

Implicit neural representation (INR) can describe the target scenes with high fidelity using a small number of parameters, and is emerging as a promising data compression technique. However, limited spectrum coverage is intrinsic to INR, and it is non-trivial to remove redundancy in diverse complex data effectively. Preliminary studies can only exploit either global or local correlation in the target data and thus of limited performance. In this paper, we propose a Tree-structured Implicit Neural Compression (TINC) to conduct compact representation for local regions and extract the shared features of these local representations in a hierarchical manner. Specifically, we use Multi-Layer Perceptrons (MLPs) to fit the partitioned local regions, and these MLPs are organized in tree structure to share parameters according to the spatial distance. The parameter sharing scheme not only ensures the continuity between adjacent regions, but also jointly removes the local and non-local redundancy. Extensive experiments show that TINC improves the compression fidelity of INR, and has shown impressive compression capabilities over commercial tools and other deep learning based methods. Besides, the approach is of high flexibility and can be tailored for different data and parameter settings. The source code can be found at https://github.com/RichealYoung/TINC.

Given sparse depths and the corresponding RGB images, depth completion aims at spatially propagating the sparse measurements throughout the whole image to get a dense depth prediction. Despite the tremendous progress of deep-learning-based depth completion methods, the locality of the convolutional layer or graph model makes it hard for the network to model the long-range relationship between pixels. While recent fully Transformer-based architecture has reported encouraging results with the global receptive field, the performance and efficiency gaps to the well-developed CNN models still exist because of its deteriorative local feature details. This paper proposes a joint convolutional attention and Transformer block (JCAT), which deeply couples the convolutional attention layer and Vision Transformer into one block, as the basic unit to construct our depth completion model in a pyramidal structure. This hybrid architecture naturally benefits both the local connectivity of convolutions and the global context of the Transformer in one single model. As a result, our CompletionFormer outperforms state-of-the-art CNNs-based methods on the outdoor KITTI Depth Completion benchmark and indoor NYUv2 dataset, achieving significantly higher efficiency (nearly 1/3 FLOPs) compared to pure Transformer-based methods. Especially when the captured depth is highly sparse, the performance gap with other methods gets much larger.

Self-supervised monocular depth estimation that does not require ground truth for training has attracted attention in recent years. It is of high interest to design lightweight but effective models so that they can be deployed on edge devices. Many existing architectures benefit from using heavier backbones at the expense of model sizes. This paper achieves comparable results with a lightweight architecture. Specifically, the efficient combination of CNNs and Transformers is investigated, and a hybrid architecture called Lite-Mono is presented. A Consecutive Dilated Convolutions (CDC) module and a Local-Global Features Interaction (LGFI) module are proposed. The former is used to extract rich multi-scale local features, and the latter takes advantage of the self-attention mechanism to encode long-range global information into the features. Experiments demonstrate that Lite-Mono outperforms Monodepth2 by a large margin in accuracy, with about 80% fewer trainable parameters. Our codes and models are available at https://github.com/noahzn/Lite-Mono.

Transformers yield state-of-the-art results across many tasks. However, their heuristically designed architecture impose huge computational costs during inference. This work aims on challenging the common design philosophy of the Vision Transformer (ViT) model with uniform dimension across all the stacked blocks in a model stage, where we redistribute the parameters both across transformer blocks and between different structures within the block via the first systematic attempt on global structural pruning. Dealing with diverse ViT structural components, we derive a novel Hessian-based structural pruning criteria comparable across all layers and structures, with latency-aware regularization for direct latency reduction. Performing iterative pruning on the DeiT-Base model leads to a new architecture family called NViT (Novel ViT), with a novel parameter redistribution that utilizes parameters more efficiently. On ImageNet-1K, NViT-Base achieves a 2.6x FLOPs reduction, 5.1x parameter reduction, and 1.9x run-time speedup over the DeiT-Base model in a near lossless manner. Smaller NViT variants achieve more than 1% accuracy gain at the same throughput of the DeiT Small/Tiny variants, as well as a lossless 3.3x parameter reduction over the SWIN-Small model. These results outperform prior art by a large margin. Further analysis is provided on the parameter redistribution insight of NViT, where we show the high prunability of ViT models, distinct sensitivity within ViT block, and unique parameter distribution trend across stacked ViT blocks. Our insights provide viability for a simple yet effective parameter redistribution rule towards more efficient ViTs for off-the-shelf performance boost.

Recent DEtection TRansformer-based (DETR) models have obtained remarkable performance. Its success cannot be achieved without the re-introduction of multi-scale feature fusion in the encoder. However, the excessively increased tokens in multi-scale features, especially for about 75% of low-level features, are quite computationally inefficient, which hinders real applications of DETR models. In this paper, we present Lite DETR, a simple yet efficient end-to-end object detection framework that can effectively reduce the GFLOPs of the detection head by 60% while keeping 99% of the original performance. Specifically, we design an efficient encoder block to update high-level features (corresponding to small-resolution feature maps) and low-level features (corresponding to large-resolution feature maps) in an interleaved way. In addition, to better fuse cross-scale features, we develop a key-aware deformable attention to predict more reliable attention weights. Comprehensive experiments validate the effectiveness and efficiency of the proposed Lite DETR, and the efficient encoder strategy can generalize well across existing DETR-based models. The code will be released after the blind review.

Vision Transformers (ViTs) are built on the assumption of treating image patches as “visual tokens” and learn patch-to-patch attention. The patch embedding based tokenizer has a semantic gap with respect to its counterpart, the textual tokenizer. The patch-to-patch attention suffers from the quadratic complexity issue, and also makes it non-trivial to explain learned ViTs. To address these issues in ViT, this paper proposes to learn Patch-to-Cluster attention (PaCa) in ViT. Queries in our PaCa-ViT starts with patches, while keys and values are directly based on clustering (with a predefined small number of clusters). The clusters are learned end-to-end, leading to better tokenizers and inducing joint clustering-for-attention and attention-for-clustering for better and interpretable models. The quadratic complexity is relaxed to linear complexity. The proposed PaCa module is used in designing efficient and interpretable ViT backbones and semantic segmentation head networks. In experiments, the proposed methods are tested on ImageNet-1k image classification, MS-COCO object detection and instance segmentation and MIT-ADE20k semantic segmentation. Compared with the prior art, it obtains better performance in all the three benchmarks than the SWin and the PVTs by significant margins in ImageNet-1k and MIT-ADE20k. It is also significantly more efficient than PVT models in MS-COCO and MIT-ADE20k due to the linear complexity. The learned clusters are semantically meaningful. Code and model checkpoints are available at https://github.com/iVMCL/PaCaViT.

Formula-driven supervised learning (FDSL) has been shown to be an effective method for pre-training vision transformers, where ExFractalDB-21k was shown to exceed the pre-training effect of ImageNet-21k. These studies also indicate that contours mattered more than textures when pre-training vision transformers. However, the lack of a systematic investigation as to why these contour-oriented synthetic datasets can achieve the same accuracy as real datasets leaves much room for skepticism. In the present work, we develop a novel methodology based on circular harmonics for systematically investigating the design space of contour-oriented synthetic datasets. This allows us to efficiently search the optimal range of FDSL parameters and maximize the variety of synthetic images in the dataset, which we found to be a critical factor. When the resulting new dataset VisualAtom-21k is used for pre-training ViT-Base, the top-1 accuracy reached 83.7% when fine-tuning on ImageNet-1k. This is only 0.5% difference from the top-1 accuracy (84.2%) achieved by the JFT-300M pre-training, even though the scale of images is 1/14. Unlike JFT-300M which is a static dataset, the quality of synthetic datasets will continue to improve, and the current work is a testament to this possibility. FDSL is also free of the common issues associated with real images, e.g. privacy/copyright issues, labeling costs/errors, and ethical biases.

Remote photoplethysmography (rPPG) technology has drawn increasing attention in recent years. It can extract Blood Volume Pulse (BVP) from facial videos, making many applications like health monitoring and emotional analysis more accessible. However, as the BVP signal is easily affected by environmental changes, existing methods struggle to generalize well for unseen domains. In this paper, we systematically address the domain shift problem in the rPPG measurement task. We show that most domain generalization methods do not work well in this problem, as domain labels are ambiguous in complicated environmental changes. In light of this, we propose a domain-label-free approach called NEuron STructure modeling (NEST). NEST improves the generalization capacity by maximizing the coverage of feature space during training, which reduces the chance for under-optimized feature activation during inference. Besides, NEST can also enrich and enhance domain invariant features across multi-domain. We create and benchmark a large-scale domain generalization protocol for the rPPG measurement task. Extensive experiments show that our approach outperforms the state-of-the-art methods on both cross-dataset and intra-dataset settings.

Image classifiers are known to be difficult to interpret and therefore require explanation methods to understand their decisions. We present ShearletX, a novel mask explanation method for image classifiers based on the shearlet transform -- a multiscale directional image representation. Current mask explanation methods are regularized by smoothness constraints that protect against undesirable fine-grained explanation artifacts. However, the smoothness of a mask limits its ability to separate fine-detail patterns, that are relevant for the classifier, from nearby nuisance patterns, that do not affect the classifier. ShearletX solves this problem by avoiding smoothness regularization all together, replacing it by shearlet sparsity constraints. The resulting explanations consist of a few edges, textures, and smooth parts of the original image, that are the most relevant for the decision of the classifier. To support our method, we propose a mathematical definition for explanation artifacts and an information theoretic score to evaluate the quality of mask explanations. We demonstrate the superiority of ShearletX over previous mask based explanation methods using these new metrics, and present exemplary situations where separating fine-detail patterns allows explaining phenomena that were not explainable before.

In this paper, we present an integral pre-training framework based on masked image modeling (MIM). We advocate for pre-training the backbone and neck jointly so that the transfer gap between MIM and downstream recognition tasks is minimal. We make two technical contributions. First, we unify the reconstruction and recognition necks by inserting a feature pyramid into the pre-training stage. Second, we complement mask image modeling (MIM) with masked feature modeling (MFM) that offers multi-stage supervision to the feature pyramid. The pre-trained models, termed integrally pre-trained transformer pyramid networks (iTPNs), serve as powerful foundation models for visual recognition. In particular, the base/large-level iTPN achieves an 86.2%/87.8% top-1 accuracy on ImageNet-1K, a 53.2%/55.6% box AP on COCO object detection with 1x training schedule using Mask-RCNN, and a 54.7%/57.7% mIoU on ADE20K semantic segmentation using UPerHead -- all these results set new records. Our work inspires the community to work on unifying upstream pre-training and downstream fine-tuning tasks. Code is available at https://github.com/sunsmarterjie/iTPN.

Modern data augmentation using a mixture-based technique can regularize the models from overfitting to the training data in various computer vision applications, but a proper data augmentation technique tailored for the part-based Visible-Infrared person Re-IDentification (VI-ReID) models remains unexplored. In this paper, we present a novel data augmentation technique, dubbed PartMix, that synthesizes the augmented samples by mixing the part descriptors across the modalities to improve the performance of part-based VI-ReID models. Especially, we synthesize the positive and negative samples within the same and across different identities and regularize the backbone model through contrastive learning. In addition, we also present an entropy-based mining strategy to weaken the adverse impact of unreliable positive and negative samples. When incorporated into existing part-based VI-ReID model, PartMix consistently boosts the performance. We conduct experiments to demonstrate the effectiveness of our PartMix over the existing VI-ReID methods and provide ablation studies.

Knowledge distillation facilitates the training of a compact student network by using a deep teacher one. While this has achieved great success in many tasks, it remains completely unstudied for image-based 6D object pose estimation. In this work, we introduce the first knowledge distillation method driven by the 6D pose estimation task. To this end, we observe that most modern 6D pose estimation frameworks output local predictions, such as sparse 2D keypoints or dense representations, and that the compact student network typically struggles to predict such local quantities precisely. Therefore, instead of imposing prediction-to-prediction supervision from the teacher to the student, we propose to distill the teacher’s distribution of local predictions into the student network, facilitating its training. Our experiments on several benchmarks show that our distillation method yields state-of-the-art results with different compact student models and for both keypoint-based and dense prediction-based architectures.

Interactive image segmentation aims at obtaining a segmentation mask for an image using simple user annotations. During each round of interaction, the segmentation result from the previous round serves as feedback to guide the user’s annotation and provides dense prior information for the segmentation model, effectively acting as a bridge between interactions. Existing methods overlook the importance of feedback or simply concatenate it with the original input, leading to underutilization of feedback and an increase in the number of required annotations. To address this, we propose an approach called Focused and Collaborative Feedback Integration (FCFI) to fully exploit the feedback for click-based interactive image segmentation. FCFI first focuses on a local area around the new click and corrects the feedback based on the similarities of high-level features. It then alternately and collaboratively updates the feedback and deep features to integrate the feedback into the features. The efficacy and efficiency of FCFI were validated on four benchmarks, namely GrabCut, Berkeley, SBD, and DAVIS. Experimental results show that FCFI achieved new state-of-the-art performance with less computational overhead than previous methods. The source code is available at https://github.com/veizgyauzgyauz/FCFI.

In this work, instead of directly predicting the pixel-level segmentation masks, the problem of referring image segmentation is formulated as sequential polygon generation, and the predicted polygons can be later converted into segmentation masks. This is enabled by a new sequence-to-sequence framework, Polygon Transformer (PolyFormer), which takes a sequence of image patches and text query tokens as input, and outputs a sequence of polygon vertices autoregressively. For more accurate geometric localization, we propose a regression-based decoder, which predicts the precise floating-point coordinates directly, without any coordinate quantization error. In the experiments, PolyFormer outperforms the prior art by a clear margin, e.g., 5.40% and 4.52% absolute improvements on the challenging RefCOCO+ and RefCOCOg datasets. It also shows strong generalization ability when evaluated on the referring video segmentation task without fine-tuning, e.g., achieving competitive 61.5% J&F on the Ref-DAVIS17 dataset.

Scene graph generation aims to construct a semantic graph structure from an image such that its nodes and edges respectively represent objects and their relationships. One of the major challenges for the task lies in the presence of distracting objects and relationships in images; contextual reasoning is strongly distracted by irrelevant objects or backgrounds and, more importantly, a vast number of irrelevant candidate relations. To tackle the issue, we propose the Selective Quad Attention Network (SQUAT) that learns to select relevant object pairs and disambiguate them via diverse contextual interactions. SQUAT consists of two main components: edge selection and quad attention. The edge selection module selects relevant object pairs, i.e., edges in the scene graph, which helps contextual reasoning, and the quad attention module then updates the edge features using both edge-to-node and edge-to-edge cross-attentions to capture contextual information between objects and object pairs. Experiments demonstrate the strong performance and robustness of SQUAT, achieving the state of the art on the Visual Genome and Open Images v6 benchmarks.

Towards building comprehensive real-world visual perception systems, we propose and study a new problem called panoptic scene graph generation (PVSG). PVSG is related to the existing video scene graph generation (VidSGG) problem, which focuses on temporal interactions between humans and objects localized with bounding boxes in videos. However, the limitation of bounding boxes in detecting non-rigid objects and backgrounds often causes VidSGG systems to miss key details that are crucial for comprehensive video understanding. In contrast, PVSG requires nodes in scene graphs to be grounded by more precise, pixel-level segmentation masks, which facilitate holistic scene understanding. To advance research in this new area, we contribute a high-quality PVSG dataset, which consists of 400 videos (289 third-person + 111 egocentric videos) with totally 150K frames labeled with panoptic segmentation masks as well as fine, temporal scene graphs. We also provide a variety of baseline methods and share useful design practices for future work.

Compared with previous two-stream trackers, the recent one-stream tracking pipeline, which allows earlier interaction between the template and search region, has achieved a remarkable performance gain. However, existing one-stream trackers always let the template interact with all parts inside the search region throughout all the encoder layers. This could potentially lead to target-background confusion when the extracted feature representations are not sufficiently discriminative. To alleviate this issue, we propose a generalized relation modeling method based on adaptive token division. The proposed method is a generalized formulation of attention-based relation modeling for Transformer tracking, which inherits the merits of both previous two-stream and one-stream pipelines whilst enabling more flexible relation modeling by selecting appropriate search tokens to interact with template tokens. An attention masking strategy and the Gumbel-Softmax technique are introduced to facilitate the parallel computation and end-to-end learning of the token division module. Extensive experiments show that our method is superior to the two-stream and one-stream pipelines and achieves state-of-the-art performance on six challenging benchmarks with a real-time running speed.

Visual representation plays an important role in visual object tracking. However, few works study the tracking-specified representation learning method. Most trackers directly use ImageNet pre-trained representations. In this paper, we propose masked appearance transfer, a simple but effective representation learning method for tracking, based on an encoder-decoder architecture. First, we encode the visual appearances of the template and search region jointly, and then we decode them separately. During decoding, the original search region image is reconstructed. However, for the template, we make the decoder reconstruct the target appearance within the search region. By this target appearance transfer, the tracking-specified representations are learned. We randomly mask out the inputs, thereby making the learned representations more discriminative. For sufficient evaluation, we design a simple and lightweight tracker that can evaluate the representation for both target localization and box regression. Extensive experiments show that the proposed method is effective, and the learned representations can enable the simple tracker to obtain state-of-the-art performance on six datasets.

The objective of this paper is self-supervised learning of video object segmentation. We develop a unified framework which simultaneously models cross-frame dense correspondence for locally discriminative feature learning and embeds object-level context for target-mask decoding. As a result, it is able to directly learn to perform mask-guided sequential segmentation from unlabeled videos, in contrast to previous efforts usually relying on an oblique solution --- cheaply “copying” labels according to pixel-wise correlations. Concretely, our algorithm alternates between i) clustering video pixels for creating pseudo segmentation labels ex nihilo; and ii) utilizing the pseudo labels to learn mask encoding and decoding for VOS. Unsupervised correspondence learning is further incorporated into this self-taught, mask embedding scheme, so as to ensure the generic nature of the learnt representation and avoid cluster degeneracy. Our algorithm sets state-of-the-arts on two standard benchmarks (i.e., DAVIS17 and YouTube-VOS), narrowing the gap between self- and fully-supervised VOS, in terms of both performance and network architecture design. Our full code will be released.

We develop a novel framework for single-scene video anomaly localization that allows for human-understandable reasons for the decisions the system makes. We first learn general representations of objects and their motions (using deep networks) and then use these representations to build a high-level, location-dependent model of any particular scene. This model can be used to detect anomalies in new videos of the same scene. Importantly, our approach is explainable -- our high-level appearance and motion features can provide human-understandable reasons for why any part of a video is classified as normal or anomalous. We conduct experiments on standard video anomaly detection datasets (Street Scene, CUHK Avenue, ShanghaiTech and UCSD Ped1, Ped2) and show significant improvements over the previous state-of-the-art. All of our code and extra datasets will be made publicly available.

Motion, scene and object are three primary visual components of a video. In particular, objects represent the foreground, scenes represent the background, and motion traces their dynamics. Based on this insight, we propose a two-stage MOtion, Scene and Object decomposition framework (MOSO) for video prediction, consisting of MOSO-VQVAE and MOSO-Transformer. In the first stage, MOSO-VQVAE decomposes a previous video clip into the motion, scene and object components, and represents them as distinct groups of discrete tokens. Then, in the second stage, MOSO-Transformer predicts the object and scene tokens of the subsequent video clip based on the previous tokens and adds dynamic motion at the token level to the generated object and scene tokens. Our framework can be easily extended to unconditional video generation and video frame interpolation tasks. Experimental results demonstrate that our method achieves new state-of-the-art performance on five challenging benchmarks for video prediction and unconditional video generation: BAIR, RoboNet, KTH, KITTI and UCF101. In addition, MOSO can produce realistic videos by combining objects and scenes from different videos.

The general domain of video segmentation is currently fragmented into different tasks spanning multiple benchmarks. Despite rapid progress in the state-of-the-art, current methods are overwhelmingly task-specific and cannot conceptually generalize to other tasks. Inspired by recent approaches with multi-task capability, we propose TarViS: a novel, unified network architecture that can be applied to any task that requires segmenting a set of arbitrarily defined ‘targets’ in video. Our approach is flexible with respect to how tasks define these targets, since it models the latter as abstract ‘queries’ which are then used to predict pixel-precise target masks. A single TarViS model can be trained jointly on a collection of datasets spanning different tasks, and can hot-swap between tasks during inference without any task-specific retraining. To demonstrate its effectiveness, we apply TarViS to four different tasks, namely Video Instance Segmentation (VIS), Video Panoptic Segmentation (VPS), Video Object Segmentation (VOS) and Point Exemplar-guided Tracking (PET). Our unified, jointly trained model achieves state-of-the-art performance on 5/7 benchmarks spanning these four tasks, and competitive performance on the remaining two. Code and model weights are available at: https://github.com/Ali2500/TarViS

The ability to distinguish between different movie scenes is critical for understanding the storyline of a movie. However, accurately detecting movie scenes is often challenging as it requires the ability to reason over very long movie segments. This is in contrast to most existing video recognition models, which are typically designed for short-range video analysis. This work proposes a State-Space Transformer model that can efficiently capture dependencies in long movie videos for accurate movie scene detection. Our model, dubbed TranS4mer, is built using a novel S4A building block, which combines the strengths of structured state-space sequence (S4) and self-attention (A) layers. Given a sequence of frames divided into movie shots (uninterrupted periods where the camera position does not change), the S4A block first applies self-attention to capture short-range intra-shot dependencies. Afterward, the state-space operation in the S4A block is used to aggregate long-range inter-shot cues. The final TranS4mer model, which can be trained end-to-end, is obtained by stacking the S4A blocks one after the other multiple times. Our proposed TranS4mer outperforms all prior methods in three movie scene detection datasets, including MovieNet, BBC, and OVSD, while also being 2x faster and requiring 3x less GPU memory than standard Transformer models. We will release our code and models.

We present RAFTformer, a real-time action forecasting transformer for latency aware real-world action forecasting applications. RAFTformer is a two-stage fully transformer based architecture which consists of a video transformer backbone that operates on high resolution, short range clips and a head transformer encoder that temporally aggregates information from multiple short range clips to span a long-term horizon. Additionally, we propose a self-supervised shuffled causal masking scheme to improve model generalization during training. Finally, we also propose a real-time evaluation setting that directly couples model inference latency to overall forecasting performance and brings forth an hitherto overlooked trade-off between latency and action forecasting performance. Our parsimonious network design facilitates RAFTformer inference latency to be 9x smaller than prior works at the same forecasting accuracy. Owing to its two-staged design, RAFTformer uses 94% less training compute and 90% lesser training parameters to outperform prior state-of-the-art baselines by 4.9 points on EGTEA Gaze+ and by 1.4 points on EPIC-Kitchens-100 dataset, as measured by Top-5 recall (T5R) in the offline setting. In the real-time setting, RAFTformer outperforms prior works by an even greater margin of upto 4.4 T5R points on the EPIC-Kitchens-100 dataset. Project Webpage: https://karttikeya.github.io/publication/RAFTformer/

Spatiotemporal predictive learning aims to generate future frames by learning from historical frames. In this paper, we investigate existing methods and present a general framework of spatiotemporal predictive learning, in which the spatial encoder and decoder capture intra-frame features and the middle temporal module catches inter-frame correlations. While the mainstream methods employ recurrent units to capture long-term temporal dependencies, they suffer from low computational efficiency due to their unparallelizable architectures. To parallelize the temporal module, we propose the Temporal Attention Unit (TAU), which decomposes temporal attention into intra-frame statical attention and inter-frame dynamical attention. Moreover, while the mean squared error loss focuses on intra-frame errors, we introduce a novel differential divergence regularization to take inter-frame variations into account. Extensive experiments demonstrate that the proposed method enables the derived model to achieve competitive performance on various spatiotemporal prediction benchmarks.

This paper deals with Audio-Visual Speech Recognition (AVSR) under multimodal input corruption situation where audio inputs and visual inputs are both corrupted, which is not well addressed in previous research directions. Previous studies have focused on how to complement the corrupted audio inputs with the clean visual inputs with the assumption of the availability of clean visual inputs. However, in real life, the clean visual inputs are not always accessible and can even be corrupted by occluded lip region or with noises. Thus, we firstly analyze that the previous AVSR models are not indeed robust to the corruption of multimodal input streams, the audio and the visual inputs, compared to uni-modal models. Then, we design multimodal input corruption modeling to develop robust AVSR models. Lastly, we propose a novel AVSR framework, namely Audio-Visual Reliability Scoring module (AV-RelScore), that is robust to the corrupted multimodal inputs. The AV-RelScore can determine which input modal stream is reliable or not for the prediction and also can exploit the more reliable streams in prediction. The effectiveness of the proposed method is evaluated with comprehensive experiments on popular benchmark databases, LRS2 and LRS3. We also show that the reliability scores obtained by AV-RelScore well reflect the degree of corruption and make the proposed model focus on the reliable multimodal representations.

Prior works on improving speech quality with visual input typically study each type of auditory distortion separately (e.g., separation, inpainting, video-to-speech) and present tailored algorithms. This paper proposes to unify these subjects and study Generalized Speech Regeneration, where the goal is not to reconstruct the exact reference clean signal, but to focus on improving certain aspects of speech while not necessarily preserving the rest such as voice. In particular, this paper concerns intelligibility, quality, and video synchronization. We cast the problem as audio-visual speech resynthesis, which is composed of two steps: pseudo audio-visual speech recognition (P-AVSR) and pseudo text-to-speech synthesis (P-TTS). P-AVSR and P-TTS are connected by discrete units derived from a self-supervised speech model. Moreover, we utilize self-supervised audio-visual speech model to initialize P-AVSR. The proposed model is coined ReVISE. ReVISE is the first high-quality model for in-the-wild video-to-speech synthesis and achieves superior performance on all LRS3 audio-visual regeneration tasks with a single model. To demonstrates its applicability in the real world, ReVISE is also evaluated on EasyCom, an audio-visual benchmark collected under challenging acoustic conditions with only 1.6 hours of training data. Similarly, ReVISE greatly suppresses noise and improves quality. Project page: https://wnhsu.github.io/ReVISE.

Recently reported state-of-the-art results in visual speech recognition (VSR) often rely on increasingly large amounts of video data, while the publicly available transcribed video datasets are limited in size. In this paper, for the first time, we study the potential of leveraging synthetic visual data for VSR. Our method, termed SynthVSR, substantially improves the performance of VSR systems with synthetic lip movements. The key idea behind SynthVSR is to leverage a speech-driven lip animation model that generates lip movements conditioned on the input speech. The speech-driven lip animation model is trained on an unlabeled audio-visual dataset and could be further optimized towards a pre-trained VSR model when labeled videos are available. As plenty of transcribed acoustic data and face images are available, we are able to generate large-scale synthetic data using the proposed lip animation model for semi-supervised VSR training. We evaluate the performance of our approach on the largest public VSR benchmark - Lip Reading Sentences 3 (LRS3). SynthVSR achieves a WER of 43.3% with only 30 hours of real labeled data, outperforming off-the-shelf approaches using thousands of hours of video. The WER is further reduced to 27.9% when using all 438 hours of labeled data from LRS3, which is on par with the state-of-the-art self-supervised AV-HuBERT method. Furthermore, when combined with large-scale pseudo-labeled audio-visual data SynthVSR yields a new state-of-the-art VSR WER of 16.9% using publicly available data only, surpassing the recent state-of-the-art approaches trained with 29 times more non-public machine-transcribed video data (90,000 hours). Finally, we perform extensive ablation studies to understand the effect of each component in our proposed method.

Semi-supervised action recognition is a challenging but critical task due to the high cost of video annotations. Existing approaches mainly use convolutional neural networks, yet current revolutionary vision transformer models have been less explored. In this paper, we investigate the use of transformer models under the SSL setting for action recognition. To this end, we introduce SVFormer, which adopts a steady pseudo-labeling framework (ie, EMA-Teacher) to cope with unlabeled video samples. While a wide range of data augmentations have been shown effective for semi-supervised image classification, they generally produce limited results for video recognition. We therefore introduce a novel augmentation strategy, Tube TokenMix, tailored for video data where video clips are mixed via a mask with consistent masked tokens over the temporal axis. In addition, we propose a temporal warping augmentation to cover the complex temporal variation in videos, which stretches selected frames to various temporal durations in the clip. Extensive experiments on three datasets Kinetics-400, UCF-101, and HMDB-51 verify the advantage of SVFormer. In particular, SVFormer outperforms the state-of-the-art by 31.5% with fewer training epochs under the 1% labeling rate of Kinetics-400. Our method can hopefully serve as a strong benchmark and encourage future search on semi-supervised action recognition with Transformer networks.

With only video-level event labels, this paper targets at the task of weakly-supervised audio-visual event perception (WS-AVEP), which aims to temporally localize and categorize events belonging to each modality. Despite the recent progress, most existing approaches either ignore the unsynchronized property of audio-visual tracks or discount the complementary modality for explicit enhancement. We argue that, for an event residing in one modality, the modality itself should provide ample presence evidence of this event, while the other complementary modality is encouraged to afford the absence evidence as a reference signal. To this end, we propose to collect Cross-Modal Presence-Absence Evidence (CMPAE) in a unified framework. Specifically, by leveraging uni-modal and cross-modal representations, a presence-absence evidence collector (PAEC) is designed under Subjective Logic theory. To learn the evidence in a reliable range, we propose a joint-modal mutual learning (JML) process, which calibrates the evidence of diverse audible, visible, and audi-visible events adaptively and dynamically. Extensive experiments show that our method surpasses state-of-the-arts (e.g., absolute gains of 3.6% and 6.1% in terms of event-level visual and audio metrics). Code is available in github.com/MengyuanChen21/CVPR2023-CMPAE.

Existing Temporal Action Detection (TAD) methods typically take a pre-processing step in converting an input varying-length video into a fixed-length snippet representation sequence, before temporal boundary estimation and action classification. This pre-processing step would temporally downsample the video, reducing the inference resolution and hampering the detection performance in the original temporal resolution. In essence, this is due to a temporal quantization error introduced during the resolution downsampling and recovery. This could negatively impact the TAD performance, but is largely ignored by existing methods. To address this problem, in this work we introduce a novel model-agnostic post-processing method without model redesign and retraining. Specifically, we model the start and end points of action instances with a Gaussian distribution for enabling temporal boundary inference at a sub-snippet level. We further introduce an efficient Taylor-expansion based approximation, dubbed as Gaussian Approximated Post-processing (GAP). Extensive experiments demonstrate that our GAP can consistently improve a wide variety of pre-trained off-the-shelf TAD models on the challenging ActivityNet (+0.2%~0.7% in average mAP) and THUMOS (+0.2%~0.5% in average mAP) benchmarks. Such performance gains are already significant and highly comparable to those achieved by novel model designs. Also, GAP can be integrated with model training for further performance gain. Importantly, GAP enables lower temporal resolutions for more efficient inference, facilitating low-resource applications. The code is available in https://github.com/sauradip/GAP.

Supervised learning of skeleton sequence encoders for action recognition has received significant attention in recent times. However, learning such encoders without labels continues to be a challenging problem. While prior works have shown promising results by applying contrastive learning to pose sequences, the quality of the learned representations is often observed to be closely tied to data augmentations that are used to craft the positives. However, augmenting pose sequences is a difficult task as the geometric constraints among the skeleton joints need to be enforced to make the augmentations realistic for that action. In this work, we propose a new contrastive learning approach to train models for skeleton-based action recognition without labels. Our key contribution is a simple module, HaLP - to Hallucinate Latent Positives for contrastive learning. Specifically, HaLP explores the latent space of poses in suitable directions to generate new positives. To this end, we present a novel optimization formulation to solve for the synthetic positives with an explicit control on their hardness. We propose approximations to the objective, making them solvable in closed form with minimal overhead. We show via experiments that using these generated positives within a standard contrastive learning framework leads to consistent improvements across benchmarks such as NTU-60, NTU-120, and PKU-II on tasks like linear evaluation, transfer learning, and kNN evaluation. Our code can be found at https://github.com/anshulbshah/HaLP.

In this paper, we present a one-stage framework TriDet for temporal action detection. Existing methods often suffer from imprecise boundary predictions due to the ambiguous action boundaries in videos. To alleviate this problem, we propose a novel Trident-head to model the action boundary via an estimated relative probability distribution around the boundary. In the feature pyramid of TriDet, we propose a Scalable-Granularity Perception (SGP) layer to aggregate information across different temporal granularities, which is much more efficient than the recent transformer-based feature pyramid. Benefiting from the Trident-head and the SGP-based feature pyramid, TriDet achieves state-of-the-art performance on three challenging benchmarks: THUMOS14, HACS and EPIC-KITCHEN 100, with lower computational costs, compared to previous methods. For example, TriDet hits an average mAP of 69.3% on THUMOS14, outperforming the previous best by 2.5%, but with only 74.6% of its latency.

In this work, we focus on reducing the annotation cost for video action detection which requires costly frame-wise dense annotations. We study a novel hybrid active learning (AL) strategy which performs efficient labeling using both intra-sample and inter-sample selection. The intra-sample selection leads to labeling of fewer frames in a video as opposed to inter-sample selection which operates at video level. This hybrid strategy reduces the annotation cost from two different aspects leading to significant labeling cost reduction. The proposed approach utilize Clustering-Aware Uncertainty Scoring (CLAUS), a novel label acquisition strategy which relies on both informativeness and diversity for sample selection. We also propose a novel Spatio-Temporal Weighted (STeW) loss formulation, which helps in model training under limited annotations. The proposed approach is evaluated on UCF-101-24 and J-HMDB-21 datasets demonstrating its effectiveness in significantly reducing the annotation cost where it consistently outperforms other baselines. Project details available at https://sites.google.com/view/activesparselabeling/home

Weakly-supervised temporal action localization aims to detect action boundaries in untrimmed videos with only video-level annotations. Most existing schemes detect temporal regions that are most responsive to video-level classification, but they overlook the semantic consistency between frames. In this paper, we hypothesize that snippets with similar representations should be considered as the same action class despite the absence of supervision signals on each snippet. To this end, we devise a learnable dictionary where entries are the class centroids of the corresponding action categories. The representations of snippets identified as the same action category are induced to be close to the same class centroid, which guides the network to perceive the semantics of frames and avoid unreasonable localization. Besides, we propose a two-stream framework that integrates the attention mechanism and the multiple-instance learning strategy to extract fine-grained clues and salient features respectively. Their complementarity enables the model to refine temporal boundaries. Finally, the developed model is validated on the publicly available THUMOS-14 and ActivityNet-1.3 datasets, where substantial experiments and analyses demonstrate that our model achieves remarkable advances over existing methods.

Automatically predicting the emotions of user-generated videos (UGVs) receives increasing interest recently. However, existing methods mainly focus on a few key visual frames, which may limit their capacity to encode the context that depicts the intended emotions. To tackle that, in this paper, we propose a cross-modal temporal erasing network that locates not only keyframes but also context and audio-related information in a weakly-supervised manner. In specific, we first leverage the intra- and inter-modal relationship among different segments to accurately select keyframes. Then, we iteratively erase keyframes to encourage the model to concentrate on the contexts that include complementary information. Extensive experiments on three challenging video emotion benchmarks demonstrate that our method performs favorably against state-of-the-art approaches. The code is released on https://github.com/nku-zhichengzhang/WECL.

Movie highlights stand out of the screenplay for efficient browsing and play a crucial role on social media platforms. Based on existing efforts, this work has two observations: (1) For different annotators, labeling highlight has uncertainty, which leads to inaccurate and time-consuming annotations. (2) Besides previous supervised or unsupervised settings, some existing video corpora can be useful, e.g., trailers, but they are often noisy and incomplete to cover the full highlights. In this work, we study a more practical and promising setting, i.e., reformulating highlight detection as “learning with noisy labels”. This setting does not require time-consuming manual annotations and can fully utilize existing abundant video corpora. First, based on movie trailers, we leverage scene segmentation to obtain complete shots, which are regarded as noisy labels. Then, we propose a Collaborative noisy Label Cleaner (CLC) framework to learn from noisy highlight moments. CLC consists of two modules: augmented cross-propagation (ACP) and multi-modality cleaning (MMC). The former aims to exploit the closely related audio-visual signals and fuse them to learn unified multi-modal representations. The latter aims to achieve cleaner highlight labels by observing the changes in losses among different modalities. To verify the effectiveness of CLC, we further collect a large-scale highlight dataset named MovieLights. Comprehensive experiments on MovieLights and YouTube Highlights datasets demonstrate the effectiveness of our approach. Code has been made available at: https://github.com/TencentYoutuResearch/HighlightDetection-CLC

The task of weakly supervised temporal sentence grounding aims at finding the corresponding temporal moments of a language description in the video, given video-language correspondence only at video-level. Most existing works select mismatched video-language pairs as negative samples and train the model to generate better positive proposals that are distinct from the negative ones. However, due to the complex temporal structure of videos, proposals distinct from the negative ones may correspond to several video segments but not necessarily the correct ground truth. To alleviate this problem, we propose an uncertainty-guided self-training technique to provide extra self-supervision signals to guide the weakly-supervised learning. The self-training process is based on teacher-student mutual learning with weak-strong augmentation, which enables the teacher network to generate relatively more reliable outputs compared to the student network, so that the student network can learn from the teacher’s output. Since directly applying existing self-training methods in this task easily causes error accumulation, we specifically design two techniques in our self-training method: (1) we construct a Bayesian teacher network, leveraging its uncertainty as a weight to suppress the noisy teacher supervisory signals; (2) we leverage the cycle consistency brought by temporal data augmentation to perform mutual learning between the two networks. Experiments demonstrate our method’s superiority on Charades-STA and ActivityNet Captions datasets. We also show in the experiment that our self-training method can be applied to improve the performance of multiple backbone methods.

Do video-text transformers learn to model temporal relationships across frames? Despite their immense capacity and the abundance of multimodal training data, recent work has revealed the strong tendency of video-text models towards frame-based spatial representations, while temporal reasoning remains largely unsolved. In this work, we identify several key challenges in temporal learning of video-text transformers: the spatiotemporal trade-off from limited network size; the curse of dimensionality for multi-frame modeling; and the diminishing returns of semantic information by extending clip length. Guided by these findings, we propose SViTT, a sparse video-text architecture that performs multi-frame reasoning with significantly lower cost than naive transformers with dense attention. Analogous to graph-based networks, SViTT employs two forms of sparsity: edge sparsity that limits the query-key communications between tokens in self-attention, and node sparsity that discards uninformative visual tokens. Trained with a curriculum which increases model sparsity with the clip length, SViTT outperforms dense transformer baselines on multiple video-text retrieval and question answering benchmarks, with a fraction of computational cost. Project page: http://svcl.ucsd.edu/projects/svitt.

The objective of this paper is an automatic Audio Description (AD) model that ingests movies and outputs AD in text form. Generating high-quality movie AD is challenging due to the dependency of the descriptions on context, and the limited amount of training data available. In this work, we leverage the power of pretrained foundation models, such as GPT and CLIP, and only train a mapping network that bridges the two models for visually-conditioned text generation. In order to obtain high-quality AD, we make the following four contributions: (i) we incorporate context from the movie clip, AD from previous clips, as well as the subtitles; (ii) we address the lack of training data by pretraining on large-scale datasets, where visual or contextual information is unavailable, e.g. text-only AD without movies or visual captioning datasets without context; (iii) we improve on the currently available AD datasets, by removing label noise in the MAD dataset, and adding character naming information; and (iv) we obtain strong results on the movie AD task compared with previous methods.

Video captioning aims to describe the content of videos using natural language. Although significant progress has been made, there is still much room to improve the performance for real-world applications, mainly due to the long-tail and open set issues of words. In this paper, we propose a text with knowledge graph augmented transformer (TextKG) for video captioning. Notably, TextKG is a two-stream transformer, formed by the external stream and internal stream. The external stream is designed to absorb external knowledge, which models the interactions between the external knowledge, e.g., pre-built knowledge graph, and the built-in information of videos, e.g., the salient object regions, speech transcripts, and video captions, to mitigate the open set of words challenge. Meanwhile, the internal stream is designed to exploit the multi-modality information in original videos (e.g., the appearance of video frames, speech transcripts, and video captions) to deal with the long-tail issue. In addition, the cross attention mechanism is also used in both streams to share information. In this way, the two streams can help each other for more accurate results. Extensive experiments conducted on four challenging video captioning datasets, i.e., YouCookII, ActivityNet Captions, MSR-VTT, and MSVD, demonstrate that the proposed method performs favorably against the state-of-the-art methods. Specifically, the proposed TextKG method outperforms the best published results by improving 18.7% absolute CIDEr scores on the YouCookII dataset.

Instructional videos are an important resource to learn procedural tasks from human demonstrations. However, the instruction steps in such videos are typically short and sparse, with most of the video being irrelevant to the procedure. This motivates the need to temporally localize the instruction steps in such videos, i.e. the task called key-step localization. Traditional methods for key-step localization require video-level human annotations and thus do not scale to large datasets. In this work, we tackle the problem with no human supervision and introduce StepFormer, a self-supervised model that discovers and localizes instruction steps in a video. StepFormer is a transformer decoder that attends to the video with learnable queries, and produces a sequence of slots capturing the key-steps in the video. We train our system on a large dataset of instructional videos, using their automatically-generated subtitles as the only source of supervision. In particular, we supervise our system with a sequence of text narrations using an order-aware loss function that filters out irrelevant phrases. We show that our model outperforms all previous unsupervised and weakly-supervised approaches on step detection and localization by a large margin on three challenging benchmarks. Moreover, our model demonstrates an emergent property to solve zero-shot multi-step localization and outperforms all relevant baselines at this task.

Video-text retrieval is an emerging stream in both computer vision and natural language processing communities, which aims to find relevant videos given text queries. In this paper, we study the notoriously challenging task, i.e., Unsupervised Domain Adaptation Video-text Retrieval (UDAVR), wherein training and testing data come from different distributions. Previous works merely alleviate the domain shift, which however overlook the pairwise misalignment issue in target domain, i.e., there exist no semantic relationships between target videos and texts. To tackle this, we propose a novel method named Dual Alignment Domain Adaptation (DADA). Specifically, we first introduce the cross-modal semantic embedding to generate discriminative source features in a joint embedding space. Besides, we utilize the video and text domain adaptations to smoothly balance the minimization of the domain shifts. To tackle the pairwise misalignment in target domain, we introduce the Dual Alignment Consistency (DAC) to fully exploit the semantic information of both modalities in target domain. The proposed DAC adaptively aligns the video-text pairs which are more likely to be relevant in target domain, enabling that positive pairs are increasing progressively and the noisy ones will potentially be aligned in the later stages. To that end, our method can generate more truly aligned target pairs and ensure the discriminality of target features.Compared with the state-of-the-art methods, DADA achieves 20.18% and 18.61% relative improvements on R@1 under the setting of TGIF->MSRVTT and TGIF->MSVD respectively, demonstrating the superiority of our method.

Video Paragraph Grounding (VPG) is an essential yet challenging task in vision-language understanding, which aims to jointly localize multiple events from an untrimmed video with a paragraph query description. One of the critical challenges in addressing this problem is to comprehend the complex semantic relations between visual and textual modalities. Previous methods focus on modeling the contextual information between the video and text from a single-level perspective (i.e., the sentence level), ignoring rich visual-textual correspondence relations at different semantic levels, e.g., the video-word and video-paragraph correspondence. To this end, we propose a novel Hierarchical Semantic Correspondence Network (HSCNet), which explores multi-level visual-textual correspondence by learning hierarchical semantic alignment and utilizes dense supervision by grounding diverse levels of queries. Specifically, we develop a hierarchical encoder that encodes the multi-modal inputs into semantics-aligned representations at different levels. To exploit the hierarchical semantic correspondence learned in the encoder for multi-level supervision, we further design a hierarchical decoder that progressively performs finer grounding for lower-level queries conditioned on higher-level semantics. Extensive experiments demonstrate the effectiveness of HSCNet and our method significantly outstrips the state-of-the-arts on two challenging benchmarks, i.e., ActivityNet-Captions and TACoS.

Pre-training a vison-language model and then fine-tuning it on downstream tasks have become a popular paradigm. However, pre-trained vison-language models with the Transformer architecture usually take long inference time. Knowledge distillation has been an efficient technique to transfer the capability of a large model to a small one while maintaining the accuracy, which has achieved remarkable success in natural language processing. However, it faces many problems when applying KD to the multi-modality applications. In this paper, we propose a novel knowledge distillation method, named CLIPPING, where the plentiful knowledge of a large teacher model that has been fine-tuned for video-language tasks with the powerful pre-trained CLIP can be effectively transferred to a small student only at the fine-tuning stage. Especially, a new layer-wise alignment with the student as the base is proposed for knowledge distillation of the intermediate layers in CLIPPING, which enables the student’s layers to be the bases of the teacher, and thus allows the student to fully absorb the knowledge of the teacher. CLIPPING with MobileViT-v2 as the vison encoder without any vison-language pre-training achieves 88.1%-95.3% of the performance of its teacher on three video-language retrieval benchmarks, with its vison encoder being 19.5x smaller. CLIPPING also significantly outperforms a state-of-the-art small baseline (ALL-in-one-B) on the MSR-VTT dataset, obtaining relatively 7.4% performance gain, with 29% fewer parameters and 86.9% fewer flops. Moreover, CLIPPING is comparable or even superior to many large pre-training models.

Emotion understanding is an essential but highly challenging component of artificial general intelligence. The absence of extensive annotated datasets has significantly impeded advancements in this field. We present EmotionCLIP, the first pre-training paradigm to extract visual emotion representations from verbal and nonverbal communication using only uncurated data. Compared to numerical labels or descriptions used in previous methods, communication naturally contains emotion information. Furthermore, acquiring emotion representations from communication is more congruent with the human learning process. We guide EmotionCLIP to attend to nonverbal emotion cues through subject-aware context encoding and verbal emotion cues using sentiment-guided contrastive learning. Extensive experiments validate the effectiveness and transferability of EmotionCLIP. Using merely linear-probe evaluation protocol, EmotionCLIP outperforms the state-of-the-art supervised visual emotion recognition methods and rivals many multimodal approaches across various benchmarks. We anticipate that the advent of EmotionCLIP will address the prevailing issue of data scarcity in emotion understanding, thereby fostering progress in related domains. The code and pre-trained models are available at https://github.com/Xeaver/EmotionCLIP.

Context-Aware Emotion Recognition (CAER) is a crucial and challenging task that aims to perceive the emotional states of the target person with contextual information. Recent approaches invariably focus on designing sophisticated architectures or mechanisms to extract seemingly meaningful representations from subjects and contexts. However, a long-overlooked issue is that a context bias in existing datasets leads to a significantly unbalanced distribution of emotional states among different context scenarios. Concretely, the harmful bias is a confounder that misleads existing models to learn spurious correlations based on conventional likelihood estimation, significantly limiting the models’ performance. To tackle the issue, this paper provides a causality-based perspective to disentangle the models from the impact of such bias, and formulate the causalities among variables in the CAER task via a tailored causal graph. Then, we propose a Contextual Causal Intervention Module (CCIM) based on the backdoor adjustment to de-confound the confounder and exploit the true causal effect for model training. CCIM is plug-in and model-agnostic, which improves diverse state-of-the-art approaches by considerable margins. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our CCIM and the significance of causal insight.

This work focuses on sign language retrieval--a recently proposed task for sign language understanding. Sign language retrieval consists of two sub-tasks: text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval. Different from traditional video-text retrieval, sign language videos, not only contain visual signals but also carry abundant semantic meanings by themselves due to the fact that sign languages are also natural languages. Considering this character, we formulate sign language retrieval as a cross-lingual retrieval problem as well as a video-text retrieval task. Concretely, we take into account the linguistic properties of both sign languages and natural languages, and simultaneously identify the fine-grained cross-lingual (i.e., sign-to-word) mappings while contrasting the texts and the sign videos in a joint embedding space. This process is termed as cross-lingual contrastive learning. Another challenge is raised by the data scarcity issue--sign language datasets are orders of magnitude smaller in scale than that of speech recognition. We alleviate this issue by adopting a domain-agnostic sign encoder pre-trained on large-scale sign videos into the target domain via pseudo-labeling. Our framework, termed as domain-aware sign language retrieval via Cross-lingual Contrastive learning or CiCo for short, outperforms the pioneering method by large margins on various datasets, e.g., +22.4 T2V and +28.0 V2T R@1 improvements on How2Sign dataset, and +13.7 T2V and +17.1 V2T R@1 improvements on PHOENIX-2014T dataset. Code and models are available at: https://github.com/FangyunWei/SLRT.

Video Question Answering (VideoQA) is challenging as it requires capturing accurate correlations between modalities from redundant information. Recent methods focus on the explicit challenges of the task, e.g. multimodal feature extraction, video-text alignment and fusion. Their frameworks reason the answer relying on statistical evidence causes, which ignores potential bias in the multimodal data. In our work, we investigate relational structure from a causal representation perspective on multimodal data and propose a novel inference framework. For visual data, question-irrelevant objects may establish simple matching associations with the answer. For textual data, the model prefers the local phrase semantics which may deviate from the global semantics in long sentences. Therefore, to enhance the generalization of the model, we discover the real association by explicitly capturing visual features that are causally related to the question semantics and weakening the impact of local language semantics on question answering. The experimental results on two large causal VideoQA datasets verify that our proposed framework 1) improves the accuracy of the existing VideoQA backbone, 2) demonstrates robustness on complex scenes and questions.

Humans universally dislike the task of cleaning up a messy room. If machines were to help us with this task, they must understand human criteria for regular arrangements, such as several types of symmetry, co-linearity or co-circularity, spacing uniformity in linear or circular patterns, and further inter-object relationships that relate to style and functionality. Previous approaches for this task relied on human input to explicitly specify goal state, or synthesized scenes from scratch--but such methods do not address the rearrangement of existing messy scenes without providing a goal state. In this paper, we present LEGO-Net, a data-driven transformer-based iterative method for LEarning reGular rearrangement of Objects in messy rooms. LEGO-Net is partly inspired by diffusion models--it starts with an initial messy state and iteratively “de-noises” the position and orientation of objects to a regular state while reducing distance traveled. Given randomly perturbed object positions and orientations in an existing dataset of professionally-arranged scenes, our method is trained to recover a regular re-arrangement. Results demonstrate that our method is able to reliably rearrange room scenes and outperform other methods. We additionally propose a metric for evaluating regularity in room arrangements using number-theoretic machinery.

Recently, visual-language navigation (VLN) -- entailing robot agents to follow navigation instructions -- has shown great advance. However, existing literature put most emphasis on interpreting instructions into actions, only delivering “dumb” wayfinding agents. In this article, we devise LANA, a language-capable navigation agent which is able to not only execute human-written navigation commands, but also provide route descriptions to humans. This is achieved by simultaneously learning instruction following and generation with only one single model. More specifically, two encoders, respectively for route and language encoding, are built and shared by two decoders, respectively, for action prediction and instruction generation, so as to exploit cross-task knowledge and capture task-specific characteristics. Throughout pretraining and fine-tuning, both instruction following and generation are set as optimization objectives. We empirically verify that, compared with recent advanced task-specific solutions, LANA attains better performances on both instruction following and route description, with nearly half complexity. In addition, endowed with language generation capability, LANA can explain to humans its behaviors and assist human’s wayfinding. This work is expected to foster future efforts towards building more trustworthy and socially-intelligent navigation robots. Our code will be released.

Recent progress on vision-language foundation models have brought significant advancement to building general-purpose robots. By using the pre-trained models to encode the scene and instructions as inputs for decision making, the instruction-conditioned policy can generalize across different objects and tasks. While this is encouraging, the policy still fails in most cases given an unseen task or environment. In this work, we propose Policy Adaptation from Foundation model Feedback (PAFF). When deploying the trained policy to a new task or a new environment, we first let the policy play with randomly generated instructions to record the demonstrations. While the execution could be wrong, we can use the pre-trained foundation models to provide feedback to relabel the demonstrations. This automatically provides new pairs of demonstration-instruction data for policy fine-tuning. We evaluate our method on a broad range of experiments with the focus on generalization on unseen objects, unseen tasks, unseen environments, and sim-to-real transfer. We show PAFF improves baselines by a large margin in all cases.

We propose Token Turing Machines (TTM), a sequential, autoregressive Transformer model with memory for real-world sequential visual understanding. Our model is inspired by the seminal Neural Turing Machine, and has an external memory consisting of a set of tokens which summarise the previous history (i.e., frames). This memory is efficiently addressed, read and written using a Transformer as the processing unit/controller at each step. The model’s memory module ensures that a new observation will only be processed with the contents of the memory (and not the entire history), meaning that it can efficiently process long sequences with a bounded computational cost at each step. We show that TTM outperforms other alternatives, such as other Transformer models designed for long sequences and recurrent neural networks, on two real-world sequential visual understanding tasks: online temporal activity detection from videos and vision-based robot action policy learning. Code is publicly available at: https://github.com/google-research/scenic/tree/main/scenic/projects/token_turing.

Analogical reasoning enables agents to extract relevant information from scenes, and efficiently navigate them in familiar ways. While progressive-matrix problems (PMPs) are becoming popular for the development and evaluation of analogical reasoning in computer vision, we argue that the dominant methodology in this area struggles to expose the lack of meaningful generalisation in solvers, and reinforces an objectivist stance on perception -- that objects can only be seen one way -- which we believe to be counter-productive. In this paper, we introduce the Unicode Analogies challenge, consisting of polysemic, character-based PMPs to benchmark fluid conceptualisation ability in vision systems. Writing systems have evolved characters at multiple levels of abstraction, from iconic through to symbolic representations, producing both visually interrelated yet exceptionally diverse images when compared to those exhibited by existing PMP datasets. Our framework has been designed to challenge models by presenting tasks much harder to complete without robust feature extraction, while remaining largely solvable by human participants. We therefore argue that Unicode Analogies elegantly captures and tests for a facet of human visual reasoning that is severely lacking in current-generation AI.

Compositionality is one of the fundamental properties of human cognition (Fodor & Pylyshyn, 1988). Compositional generalization is critical to simulate the compositional capability of humans, and has received much attention in the vision-and-language (V&L) community. It is essential to understand the effect of the primitives, including words, image regions, and video frames, to improve the compositional generalization capability. In this paper, we explore the effect of primitives for compositional generalization in V&L. Specifically, we present a self-supervised learning based framework that equips V&L methods with two characteristics: semantic equivariance and semantic invariance. With the two characteristics, the methods understand primitives by perceiving the effect of primitive changes on sample semantics and ground-truth. Experimental results on two tasks: temporal video grounding and visual question answering, demonstrate the effectiveness of our framework.

Research on continual learning has recently led to a variety of work in unimodal community, however little attention has been paid to multimodal tasks like visual question answering (VQA). In this paper, we establish a novel VQA Continual Learning setting named VQACL, which contains two key components: a dual-level task sequence where visual and linguistic data are nested, and a novel composition testing containing new skill-concept combinations. The former devotes to simulating the ever-changing multimodal datastream in real world and the latter aims at measuring models’ generalizability for cognitive reasoning. Based on our VQACL, we perform in-depth evaluations of five well-established continual learning methods, and observe that they suffer from catastrophic forgetting and have weak generalizability. To address above issues, we propose a novel representation learning method, which leverages a sample-specific and a sample-invariant feature to learn representations that are both discriminative and generalizable for VQA. Furthermore, by respectively extracting such representation for visual and textual input, our method can explicitly disentangle the skill and concept. Extensive experimental results illustrate that our method significantly outperforms existing models, demonstrating the effectiveness and compositionality of the proposed approach.

Pre-trained vision-language (V-L) models such as CLIP have shown excellent generalization ability to downstream tasks. However, they are sensitive to the choice of input text prompts and require careful selection of prompt templates to perform well. Inspired by the Natural Language Processing (NLP) literature, recent CLIP adaptation approaches learn prompts as the textual inputs to fine-tune CLIP for downstream tasks. We note that using prompting to adapt representations in a single branch of CLIP (language or vision) is sub-optimal since it does not allow the flexibility to dynamically adjust both representation spaces on a downstream task. In this work, we propose Multi-modal Prompt Learning (MaPLe) for both vision and language branches to improve alignment between the vision and language representations. Our design promotes strong coupling between the vision-language prompts to ensure mutual synergy and discourages learning independent uni-modal solutions. Further, we learn separate prompts across different early stages to progressively model the stage-wise feature relationships to allow rich context learning. We evaluate the effectiveness of our approach on three representative tasks of generalization to novel classes, new target datasets and unseen domain shifts. Compared with the state-of-the-art method Co-CoOp, MaPLe exhibits favorable performance and achieves an absolute gain of 3.45% on novel classes and 2.72% on overall harmonic-mean, averaged over 11 diverse image recognition datasets. Our code and pre-trained models are available at https://github.com/muzairkhattak/multimodal-prompt-learning.

Large-scale vision-language models (VLM) have shown impressive results for language-guided search applications. While these models allow category-level queries, they currently struggle with personalized searches for moments in a video where a specific object instance such as “My dog Biscuit” appears. We present the following three contributions to address this problem. First, we describe a method to meta-personalize a pre-trained VLM, i.e., learning how to learn to personalize a VLM at test time to search in video. Our method extends the VLM’s token vocabulary by learning novel word embeddings specific to each instance. To capture only instance-specific features, we represent each instance embedding as a combination of shared and learned global category features. Second, we propose to learn such personalization without explicit human supervision. Our approach automatically identifies moments of named visual instances in video using transcripts and vision-language similarity in the VLM’s embedding space. Finally, we introduce This-Is-My, a personal video instance retrieval benchmark. We evaluate our approach on This-Is-My and DeepFashion2 and show that we obtain a 15% relative improvement over the state of the art on the latter dataset.

We revisit and advance visual prompting (VP), an input prompting technique for vision tasks. VP can reprogram a fixed, pre-trained source model to accomplish downstream tasks in the target domain by simply incorporating universal prompts (in terms of input perturbation patterns) into downstream data points. Yet, it remains elusive why VP stays effective even given a ruleless label mapping (LM) between the source classes and the target classes. Inspired by the above, we ask: How is LM interrelated with VP? And how to exploit such a relationship to improve its accuracy on target tasks? We peer into the influence of LM on VP and provide an affirmative answer that a better ‘quality’ of LM (assessed by mapping precision and explanation) can consistently improve the effectiveness of VP. This is in contrast to the prior art where the factor of LM was missing. To optimize LM, we propose a new VP framework, termed ILM-VP (iterative label mapping-based visual prompting), which automatically re-maps the source labels to the target labels and progressively improves the target task accuracy of VP. Further, when using a contrastive language-image pretrained (CLIP) model, we propose to integrate an LM process to assist the text prompt selection of CLIP and to improve the target task accuracy. Extensive experiments demonstrate that our proposal significantly outperforms state-of-the-art VP methods. As highlighted below, we show that when reprogramming an ImageNet-pretrained ResNet-18 to 13 target tasks, our method outperforms baselines by a substantial margin, e.g., 7.9% and 6.7% accuracy improvements in transfer learning to the target Flowers102 and CIFAR100 datasets. Besides, our proposal on CLIP-based VP provides 13.7% and 7.1% accuracy improvements on Flowers102 and DTD respectively.

Referring expression comprehension (REC) often requires a large number of instance-level annotations for fully supervised learning, which are laborious and expensive. In this paper, we present the first attempt of semi-supervised learning for REC and propose a strong baseline method called RefTeacher. Inspired by the recent progress in computer vision, RefTeacher adopts a teacher-student learning paradigm, where the teacher REC network predicts pseudo-labels for optimizing the student one. This paradigm allows REC models to exploit massive unlabeled data based on a small fraction of labeled. In particular, we also identify two key challenges in semi-supervised REC, namely, sparse supervision signals and worse pseudo-label noise. To address these issues, we equip RefTeacher with two novel designs called Attention-based Imitation Learning (AIL) and Adaptive Pseudo-label Weighting (APW). AIL can help the student network imitate the recognition behaviors of the teacher, thereby obtaining sufficient supervision signals. APW can help the model adaptively adjust the contributions of pseudo-labels with varying qualities, thus avoiding confirmation bias. To validate RefTeacher, we conduct extensive experiments on three REC benchmark datasets. Experimental results show that RefTeacher obtains obvious gains over the fully supervised methods. More importantly, using only 10% labeled data, our approach allows the model to achieve near 100% fully supervised performance, e.g., only -2.78% on RefCOCO.

Most existing vision-language pre-training (VLP) approaches adopt cross-modal masked language modeling (CMLM) to learn vision-language associations. However, we find that CMLM is insufficient for this purpose according to our observations: (1) Modality bias: a considerable amount of masked tokens in CMLM can be recovered with only the language information, ignoring the visual inputs. (2) Under-utilization of the unmasked tokens: CMLM primarily focuses on the masked token but it cannot simultaneously leverage other tokens to learn vision-language associations. To handle those limitations, we propose EPIC (lEveraging Per Image-Token Consistency for vision-language pre-training). In EPIC, for each image-sentence pair, we mask tokens that are salient to the image (i.e., Saliency-based Masking Strategy) and replace them with alternatives sampled from a language model (i.e., Inconsistent Token Generation Procedure), and then the model is required to determine for each token in the sentence whether it is consistent with the image (i.e., Image-Token Consistency Task). The proposed EPIC method is easily combined with pre-training methods. Extensive experiments show that the combination of the EPIC method and state-of-the-art pre-training approaches, including ViLT, ALBEF, METER, and X-VLM, leads to significant improvements on downstream tasks. Our coude is released at https://github.com/gyhdog99/epic

We propose a margin-based loss for tuning joint vision-language models so that their gradient-based explanations are consistent with region-level annotations provided by humans for relatively smaller grounding datasets. We refer to this objective as Attention Mask Consistency (AMC) and demonstrate that it produces superior visual grounding results than previous methods that rely on using vision-language models to score the outputs of object detectors. Particularly, a model trained with AMC on top of standard vision-language modeling objectives obtains a state-of-the-art accuracy of 86.49% in the Flickr30k visual grounding benchmark, an absolute improvement of 5.38% when compared to the best previous model trained under the same level of supervision. Our approach also performs exceedingly well on established benchmarks for referring expression comprehension where it obtains 80.34% accuracy in the easy test of RefCOCO+, and 64.55% in the difficult split. AMC is effective, easy to implement, and is general as it can be adopted by any vision-language model, and can use any type of region annotations.

A big convergence of language, vision, and multimodal pretraining is emerging. In this work, we introduce a general-purpose multimodal foundation model BEiT-3, which achieves excellent transfer performance on both vision and vision-language tasks. Specifically, we advance the big convergence from three aspects: backbone architecture, pretraining task, and model scaling up. We use Multiway Transformers for general-purpose modeling, where the modular architecture enables both deep fusion and modality-specific encoding. Based on the shared backbone, we perform masked “language” modeling on images (Imglish), texts (English), and image-text pairs (“parallel sentences”) in a unified manner. Experimental results show that BEiT-3 obtains remarkable performance on object detection (COCO), semantic segmentation (ADE20K), image classification (ImageNet), visual reasoning (NLVR2), visual question answering (VQAv2), image captioning (COCO), and cross-modal retrieval (Flickr30K, COCO).

Concept Bottleneck Models (CBM) are inherently interpretable models that factor model decisions into human-readable concepts. They allow people to easily understand why a model is failing, a critical feature for high-stakes applications. CBMs require manually specified concepts and often under-perform their black box counterparts, preventing their broad adoption. We address these shortcomings and are first to show how to construct high-performance CBMs without manual specification of similar accuracy to black box models. Our approach, Language Guided Bottlenecks (LaBo), leverages a language model, GPT-3, to define a large space of possible bottlenecks. Given a problem domain, LaBo uses GPT-3 to produce factual sentences about categories to form candidate concepts. LaBo efficiently searches possible bottlenecks through a novel submodular utility that promotes the selection of discriminative and diverse information. Ultimately, GPT-3’s sentential concepts can be aligned to images using CLIP, to form a bottleneck layer. Experiments demonstrate that LaBo is a highly effective prior for concepts important to visual recognition. In the evaluation with 11 diverse datasets, LaBo bottlenecks excel at few-shot classification: they are 11.7% more accurate than black box linear probes at 1 shot and comparable with more data. Overall, LaBo demonstrates that inherently interpretable models can be widely applied at similar, or better, performance than black box approaches.

Object-centric learning (OCL) aspires general and com- positional understanding of scenes by representing a scene as a collection of object-centric representations. OCL has also been extended to multi-view image and video datasets to apply various data-driven inductive biases by utilizing geometric or temporal information in the multi-image data. Single-view images carry less information about how to disentangle a given scene than videos or multi-view im- ages do. Hence, owing to the difficulty of applying induc- tive biases, OCL for single-view images still remains chal- lenging, resulting in inconsistent learning of object-centric representation. To this end, we introduce a novel OCL framework for single-view images, SLot Attention via SHep- herding (SLASH), which consists of two simple-yet-effective modules on top of Slot Attention. The new modules, At- tention Refining Kernel (ARK) and Intermediate Point Pre- dictor and Encoder (IPPE), respectively, prevent slots from being distracted by the background noise and indicate lo- cations for slots to focus on to facilitate learning of object- centric representation. We also propose a weak- and semi- supervision approach for OCL, whilst our proposed frame- work can be used without any assistant annotation during the inference. Experiments show that our proposed method enables consistent learning of object-centric representa- tion and achieves strong performance across four datasets. Code is available at https://github.com/object- understanding/SLASH.

Although an object may appear in numerous contexts, we often describe it in a limited number of ways. Language allows us to abstract away visual variation to represent and communicate concepts. Building on this intuition, we propose an alternative approach to visual representation learning: using language similarity to sample semantically similar image pairs for contrastive learning. Our approach diverges from image-based contrastive learning by sampling view pairs using language similarity instead of hand-crafted augmentations or learned clusters. Our approach also differs from image-text contrastive learning by relying on pre-trained language models to guide the learning rather than directly minimizing a cross-modal loss. Through a series of experiments, we show that language-guided learning yields better features than image-based and image-text representation learning approaches.

Language-based colorization produces plausible colors consistent with the language description provided by the user. Recent studies introduce additional annotation to prevent color-object coupling and mismatch issues, but they still have difficulty in distinguishing instances corresponding to the same object words. In this paper, we propose a transformer-based framework to automatically aggregate similar image patches and achieve instance awareness without any additional knowledge. By applying our presented luminance augmentation and counter-color loss to break down the statistical correlation between luminance and color words, our model is driven to synthesize colors with better descriptive consistency. We further collect a dataset to provide distinctive visual characteristics and detailed language descriptions for multiple instances in the same image. Extensive experiments demonstrate our advantages of synthesizing visually pleasing and description-consistent results of instance-aware colorization.

3D visual grounding aims to find the object within point clouds mentioned by free-form natural language descriptions with rich semantic cues. However, existing methods either extract the sentence-level features coupling all words or focus more on object names, which would lose the word-level information or neglect other attributes. To alleviate these issues, we present EDA that Explicitly Decouples the textual attributes in a sentence and conducts Dense Alignment between such fine-grained language and point cloud objects. Specifically, we first propose a text decoupling module to produce textual features for every semantic component. Then, we design two losses to supervise the dense matching between two modalities: position alignment loss and semantic alignment loss. On top of that, we further introduce a new visual grounding task, locating objects without object names, which can thoroughly evaluate the model’s dense alignment capacity. Through experiments, we achieve state-of-the-art performance on two widely-adopted 3D visual grounding datasets, ScanRefer and SR3D/NR3D, and obtain absolute leadership on our newly-proposed task. The source code is available at https://github.com/yanmin-wu/EDA.

Neural Architecture Search (NAS) has been increasingly appealing to the society of object Re-Identification (ReID), for that task-specific architectures significantly improve the retrieval performance. Previous works explore new optimizing targets and search spaces for NAS ReID, yet they neglect the difference of training schemes between image classification and ReID. In this work, we propose a novel Twins Contrastive Mechanism (TCM) to provide more appropriate supervision for ReID architecture search. TCM reduces the category overlaps between the training and validation data, and assists NAS in simulating real-world ReID training schemes. We then design a Multi-Scale Interaction (MSI) search space to search for rational interaction operations between multi-scale features. In addition, we introduce a Spatial Alignment Module (SAM) to further enhance the attention consistency confronted with images from different sources. Under the proposed NAS scheme, a specific architecture is automatically searched, named as MSINet. Extensive experiments demonstrate that our method surpasses state-of-the-art ReID methods on both in-domain and cross-domain scenarios.

We propose Universal Document Processing (UDOP), a foundation Document AI model which unifies text, image, and layout modalities together with varied task formats, including document understanding and generation. UDOP leverages the spatial correlation between textual content and document image to model image, text, and layout modalities with one uniform representation. With a novel Vision-Text-Layout Transformer, UDOP unifies pretraining and multi-domain downstream tasks into a prompt-based sequence generation scheme. UDOP is pretrained on both large-scale unlabeled document corpora using innovative self-supervised objectives and diverse labeled data. UDOP also learns to generate document images from text and layout modalities via masked image reconstruction. To the best of our knowledge, this is the first time in the field of document AI that one model simultaneously achieves high-quality neural document editing and content customization. Our method sets the state-of-the-art on 8 Document AI tasks, e.g., document understanding and QA, across diverse data domains like finance reports, academic papers, and websites. UDOP ranks first on the leaderboard of the Document Understanding Benchmark.

Contrastive Language-Image Pre-training (CLIP) is attracting increasing attention for its impressive zero-shot recognition performance on different down-stream tasks. However, training CLIP is data-hungry and requires lots of image-text pairs to memorize various semantic concepts. In this paper, we propose a novel and efficient framework: Retrieval Augmented Contrastive Language-Image Pre-training (RA-CLIP) to augment embeddings by online retrieval. Specifically, we sample part of image-text data as a hold-out reference set. Given an input image, relevant image-text pairs are retrieved from the reference set to enrich the representation of input image. This process can be considered as an open-book exam: with the reference set as a cheat sheet, the proposed method doesn’t need to memorize all visual concepts in the training data. It explores how to recognize visual concepts by exploiting correspondence between images and texts in the cheat sheet. The proposed RA-CLIP implements this idea and comprehensive experiments are conducted to show how RA-CLIP works. Performances on 10 image classification datasets and 2 object detection datasets show that RA-CLIP outperforms vanilla CLIP baseline by a large margin on zero-shot image classification task (+12.7%), linear probe image classification task (+6.9%) and zero-shot ROI classification task (+2.8%).

Current state-of-the-art image-text matching methods implicitly align the visual-semantic fragments, like regions in images and words in sentences, and adopt cross-attention mechanism to discover fine-grained cross-modal semantic correspondence. However, the cross-attention mechanism may bring redundant or irrelevant region-word alignments, degenerating retrieval accuracy and limiting efficiency. Although many researchers have made progress in mining meaningful alignments and thus improving accuracy, the problem of poor efficiency remains unresolved. In this work, we propose to learn fine-grained image-text matching from the perspective of information coding. Specifically, we suggest a coding framework to explain the fragments aligning process, which provides a novel view to reexamine the cross-attention mechanism and analyze the problem of redundant alignments. Based on this framework, a Cross-modal Hard Aligning Network (CHAN) is designed, which comprehensively exploits the most relevant region-word pairs and eliminates all other alignments. Extensive experiments conducted on two public datasets, MS-COCO and Flickr30K, verify that the relevance of the most associated word-region pairs is discriminative enough as an indicator of the image-text similarity, with superior accuracy and efficiency over the state-of-the-art approaches on the bidirectional image and text retrieval tasks. Our code will be available at https://github.com/ppanzx/CHAN.

Great progress has been made in StyleGAN-based image editing. To associate with preset attributes, most existing approaches focus on supervised learning for semantically meaningful latent space traversal directions, and each manipulation step is typically determined for an individual attribute. To address this limitation, we propose a Text-guided Unsupervised StyleGAN Latent Transformation (TUSLT) model, which adaptively infers a single transformation step in the latent space of StyleGAN to simultaneously manipulate multiple attributes on a given input image. Specifically, we adopt a two-stage architecture for a latent mapping network to break down the transformation process into two manageable steps. Our network first learns a diverse set of semantic directions tailored to an input image, and later nonlinearly fuses the ones associated with the target attributes to infer a residual vector. The resulting tightly interlinked two-stage architecture delivers the flexibility to handle diverse attribute combinations. By leveraging the cross-modal text-image representation of CLIP, we can perform pseudo annotations based on the semantic similarity between preset attribute text descriptions and training images, and further jointly train an auxiliary attribute classifier with the latent mapping network to provide semantic guidance. We perform extensive experiments to demonstrate that the adopted strategies contribute to the superior performance of TUSLT.

Retrieval augmented models are becoming increasingly popular for computer vision tasks after their recent success in NLP problems. The goal is to enhance the recognition capabilities of the model by retrieving similar examples for the visual input from an external memory set. In this work, we introduce an attention-based memory module, which learns the importance of each retrieved example from the memory. Compared to existing approaches, our method removes the influence of the irrelevant retrieved examples, and retains those that are beneficial to the input query. We also thoroughly study various ways of constructing the memory dataset. Our experiments show the benefit of using a massive-scale memory dataset of 1B image-text pairs, and demonstrate the performance of different memory representations. We evaluate our method in three different classification tasks, namely long-tailed recognition, learning with noisy labels, and fine-grained classification, and show that it achieves state-of-the-art accuracies in ImageNet-LT, Places-LT and Webvision datasets.

In Composed Image Retrieval (CIR), a user combines a query image with text to describe their intended target. Existing methods rely on supervised learning of CIR models using labeled triplets consisting of the query image, text specification, and the target image. Labeling such triplets is expensive and hinders broad applicability of CIR. In this work, we propose to study an important task, Zero-Shot Composed Image Retrieval (ZS-CIR), whose goal is to build a CIR model without requiring labeled triplets for training. To this end, we propose a novel method, called Pic2Word, that requires only weakly labeled image-caption pairs and unlabeled image datasets to train. Unlike existing supervised CIR models, our model trained on weakly labeled or unlabeled datasets shows strong generalization across diverse ZS-CIR tasks, e.g., attribute editing, object composition, and domain conversion. Our approach outperforms several supervised CIR methods on the common CIR benchmark, CIRR and Fashion-IQ.

Product Retrieval (PR) and Grounding (PG), aiming to seek image and object-level products respectively according to a textual query, have attracted great interest recently for better shopping experience. Owing to the lack of relevant datasets, we collect two large-scale benchmark datasets from Taobao Mall and Live domains with about 474k and 101k image-query pairs for PR, and manually annotate the object bounding boxes in each image for PG. As annotating boxes is expensive and time-consuming, we attempt to transfer knowledge from annotated domain to unannotated for PG to achieve un-supervised Domain Adaptation (PG-DA). We propose a Domain Adaptive producT sEeker (DATE) framework, regarding PR and PG as Product Seeking problem at different levels, to assist the query date the product. Concretely, we first design a semantics-aggregated feature extractor for each modality to obtain concentrated and comprehensive features for following efficient retrieval and fine-grained grounding tasks. Then, we present two cooperative seekers to simultaneously search the image for PR and localize the product for PG. Besides, we devise a domain aligner for PG-DA to alleviate uni-modal marginal and multi-modal conditional distribution shift between source and target domains, and design a pseudo box generator to dynamically select reliable instances and generate bounding boxes for further knowledge transfer. Extensive experiments show that our DATE achieves satisfactory performance in fully-supervised PR, PG and un-supervised PG-DA. Our desensitized datasets will be publicly available here https://github.com/Taobao-live/Product-Seeking.

The ability to quickly learn a new task with minimal instruction - known as few-shot learning - is a central aspect of intelligent agents. Classical few-shot benchmarks make use of few-shot samples from a single modality, but such samples may not be sufficient to characterize an entire concept class. In contrast, humans use cross-modal information to learn new concepts efficiently. In this work, we demonstrate that one can indeed build a better visual dog classifier by reading about dogs and listening to them bark. To do so, we exploit the fact that recent multimodal foundation models such as CLIP are inherently cross-modal, mapping different modalities to the same representation space. Specifically, we propose a simple cross-modal adaptation approach that learns from few-shot examples spanning different modalities. By repurposing class names as additional one-shot training samples, we achieve SOTA results with an embarrassingly simple linear classifier for vision-language adaptation. Furthermore, we show that our approach can benefit existing methods such as prefix tuning and classifier ensembling. Finally, to explore other modalities beyond vision and language, we construct the first (to our knowledge) audiovisual few-shot benchmark and use cross-modal training to improve the performance of both image and audio classification. We hope our success can inspire future works to embrace cross-modality for even broader domains and tasks.

Finetuning image-text models such as CLIP achieves state-of-the-art accuracies on a variety of benchmarks. However, recent works (Kumar et al., 2022; Wortsman et al., 2021) have shown that even subtle differences in the finetuning process can lead to surprisingly large differences in the final performance, both for in-distribution (ID) and out-of-distribution (OOD) data. In this work, we show that a natural and simple approach of mimicking contrastive pretraining consistently outperforms alternative finetuning approaches. Specifically, we cast downstream class labels as text prompts and continue optimizing the contrastive loss between image embeddings and class-descriptive prompt embeddings (contrastive finetuning). Our method consistently outperforms baselines across 7 distribution shift, 6 transfer learning, and 3 few-shot learning benchmarks. On WILDS-iWILDCam, our proposed approach FLYP outperforms the top of the leaderboard by 2.3% ID and 2.7% OOD, giving the highest reported accuracy. Averaged across 7 OOD datasets (2 WILDS and 5 ImageNet associated shifts), FLYP gives gains of 4.2% OOD over standard finetuning and outperforms current state-ofthe-art (LP-FT) by more than 1% both ID and OOD. Similarly, on 3 few-shot learning benchmarks, FLYP gives gains up to 4.6% over standard finetuning and 4.4% over the state-of-the-art. Thus we establish our proposed method of contrastive finetuning as a simple and intuitive state-ofthe-art for supervised finetuning of image-text models like CLIP. Code is available at https://github.com/locuslab/FLYP.

End-to-end text spotting aims to integrate scene text detection and recognition into a unified framework. Dealing with the relationship between the two sub-tasks plays a pivotal role in designing effective spotters. Although Transformer-based methods eliminate the heuristic post-processing, they still suffer from the synergy issue between the sub-tasks and low training efficiency. In this paper, we present DeepSolo, a simple DETR-like baseline that lets a single Decoder with Explicit Points Solo for text detection and recognition simultaneously. Technically, for each text instance, we represent the character sequence as ordered points and model them with learnable explicit point queries. After passing a single decoder, the point queries have encoded requisite text semantics and locations, thus can be further decoded to the center line, boundary, script, and confidence of text via very simple prediction heads in parallel. Besides, we also introduce a text-matching criterion to deliver more accurate supervisory signals, thus enabling more efficient training. Quantitative experiments on public benchmarks demonstrate that DeepSolo outperforms previous state-of-the-art methods and achieves better training efficiency. In addition, DeepSolo is also compatible with line annotations, which require much less annotation cost than polygons. The code is available at https://github.com/ViTAE-Transformer/DeepSolo.

We launch EVA, a vision-centric foundation model to explore the limits of visual representation at scale using only publicly accessible data. EVA is a vanilla ViT pre-trained to reconstruct the masked out image-text aligned vision features conditioned on visible image patches. Via this pretext task, we can efficiently scale up EVA to one billion parameters, and sets new records on a broad range of representative vision downstream tasks, such as image recognition, video action recognition, object detection, instance segmentation and semantic segmentation without heavy supervised training. Moreover, we observe quantitative changes in scaling EVA result in qualitative changes in transfer learning performance that are not present in other models. For instance, EVA takes a great leap in the challenging large vocabulary instance segmentation task: our model achieves almost the same state-of-the-art performance on LVIS dataset with over a thousand categories and COCO dataset with only eighty categories. Beyond a pure vision encoder, EVA can also serve as a vision-centric, multi-modal pivot to connect images and text. We find initializing the vision tower of a giant CLIP from EVA can greatly stabilize the training and outperform the training from scratch counterpart with much fewer samples and less compute, providing a new direction for scaling up and accelerating the costly training of multi-modal foundation models.

Visual Place Recognition (VPR) estimates the location of query images by matching them with images in a reference database. Conventional methods generally adopt aggregated CNN features for global retrieval and RANSAC-based geometric verification for reranking. However, RANSAC only employs geometric information but ignores other possible information that could be useful for reranking, e.g. local feature correlations, and attention values. In this paper, we propose a unified place recognition framework that handles both retrieval and reranking with a novel transformer model, named R2Former. The proposed reranking module takes feature correlation, attention value, and xy coordinates into account, and learns to determine whether the image pair is from the same location. The whole pipeline is end-to-end trainable and the reranking module alone can also be adopted on other CNN or transformer backbones as a generic component. Remarkably, R2Former significantly outperforms state-of-the-art methods on major VPR datasets with much less inference time and memory consumption. It also achieves the state-of-the-art on the hold-out MSLS challenge set and could serve as a simple yet strong solution for real-world large-scale applications. Experiments also show vision transformer tokens are comparable and sometimes better than CNN local features on local matching. The code is released at https://github.com/Jeff-Zilence/R2Former.

Open-set fine-grained retrieval is an emerging challenge that requires an extra capability to retrieve unknown subcategories during evaluation. However, current works are rooted in the close-set scenarios, where all the subcategories are pre-defined, and make it hard to capture discriminative knowledge from unknown subcategories, consequently failing to handle the inevitable unknown subcategories in open-world scenarios. In this work, we propose a novel Prompting vision-Language Evaluator (PLEor) framework based on the recently introduced contrastive language-image pretraining (CLIP) model, for open-set fine-grained retrieval. PLEor could leverage pre-trained CLIP model to infer the discrepancies encompassing both pre-defined and unknown subcategories, called category-specific discrepancies, and transfer them to the backbone network trained in the close-set scenarios. To make pre-trained CLIP model sensitive to category-specific discrepancies, we design a dual prompt scheme to learn a vision prompt specifying the category-specific discrepancies, and turn random vectors with category names in a text prompt into category-specific discrepancy descriptions. Moreover, a vision-language evaluator is proposed to semantically align the vision and text prompts based on CLIP model, and reinforce each other. In addition, we propose an open-set knowledge transfer to transfer the category-specific discrepancies into the backbone network using knowledge distillation mechanism. A variety of quantitative and qualitative experiments show that our PLEor achieves promising performance on open-set fine-grained retrieval datasets.

Human-object interaction (HOI) has long been plagued by the conflict between limited supervised data and a vast number of possible interaction combinations in real life. Current methods trained from closed-set data predict HOIs as fixed-dimension logits, which restricts their scalability to open-set categories. To address this issue, we introduce OpenCat, a language modeling framework that reformulates HOI prediction as sequence generation. By converting HOI triplets into a token sequence through a serialization scheme, our model is able to exploit the open-set vocabulary of the language modeling framework to predict novel interaction classes with a high degree of freedom. In addition, inspired by the great success of vision-language pre-training, we collect a large amount of weakly-supervised data related to HOI from image-caption pairs, and devise several auxiliary proxy tasks, including soft relational matching and human-object relation prediction, to pre-train our model. Extensive experiments show that our OpenCat significantly boosts HOI performance, particularly on a broad range of rare and unseen categories.

We present Neural Congealing -- a zero-shot self-supervised framework for detecting and jointly aligning semantically-common content across a given set of images. Our approach harnesses the power of pre-trained DINO-ViT features to learn: (i) a joint semantic atlas -- a 2D grid that captures the mode of DINO-ViT features in the input set, and (ii) dense mappings from the unified atlas to each of the input images. We derive a new robust self-supervised framework that optimizes the atlas representation and mappings per image set, requiring only a few real-world images as input without any additional input information (e.g., segmentation masks). Notably, we design our losses and training paradigm to account only for the shared content under severe variations in appearance, pose, background clutter or other distracting objects. We demonstrate results on a plethora of challenging image sets including sets of mixed domains (e.g., aligning images depicting sculpture and artwork of cats), sets depicting related yet different object categories (e.g., dogs and tigers), or domains for which large-scale training data is scarce (e.g., coffee mugs). We thoroughly evaluate our method and show that our test-time optimization approach performs favorably compared to a state-of-the-art method that requires extensive training on large-scale datasets.

We introduce Patch Aligned Contrastive Learning (PACL), a modified compatibility function for CLIP’s contrastive loss, intending to train an alignment between the patch tokens of the vision encoder and the CLS token of the text encoder. With such an alignment, a model can identify regions of an image corresponding to a given text input, and therefore transfer seamlessly to the task of open vocabulary semantic segmentation without requiring any segmentation annotations during training. Using pre-trained CLIP encoders with PACL, we are able to set the state-of-the-art on the task of open vocabulary zero-shot segmentation on 4 different segmentation benchmarks: Pascal VOC, Pascal Context, COCO Stuff and ADE20K. Furthermore, we show that PACL is also applicable to image-level predictions and when used with a CLIP backbone, provides a general improvement in zero-shot classification accuracy compared to CLIP, across a suite of 12 image classification datasets.

This paper presents Scalable Semantic Transfer (SST), a novel training paradigm, to explore how to leverage the mutual benefits of the data from different label domains (i.e. various levels of label granularity) to train a powerful human parsing network. In practice, two common application scenarios are addressed, termed universal parsing and dedicated parsing, where the former aims to learn homogeneous human representations from multiple label domains and switch predictions by only using different segmentation heads, and the latter aims to learn a specific domain prediction while distilling the semantic knowledge from other domains. The proposed SST has the following appealing benefits: (1) it can capably serve as an effective training scheme to embed semantic associations of human body parts from multiple label domains into the human representation learning process; (2) it is an extensible semantic transfer framework without predetermining the overall relations of multiple label domains, which allows continuously adding human parsing datasets to promote the training. (3) the relevant modules are only used for auxiliary training and can be removed during inference, eliminating the extra reasoning cost. Experimental results demonstrate SST can effectively achieve promising universal human parsing performance as well as impressive improvements compared to its counterparts on three human parsing benchmarks (i.e., PASCAL-Person-Part, ATR, and CIHP). Code is available at https://github.com/yangjie-cv/SST.

We consider the generic problem of detecting low-level structures in images, which includes segmenting the manipulated parts, identifying out-of-focus pixels, separating shadow regions, and detecting concealed objects. Whereas each such topic has been typically addressed with a domain-specific solution, we show that a unified approach performs well across all of them. We take inspiration from the widely-used pre-training and then prompt tuning protocols in NLP and propose a new visual prompting model, named Explicit Visual Prompting (EVP). Different from the previous visual prompting which is typically a dataset-level implicit embedding, our key insight is to enforce the tunable parameters focusing on the explicit visual content from each individual image, i.e., the features from frozen patch embeddings and the input’s high-frequency components. The proposed EVP significantly outperforms other parameter-efficient tuning protocols under the same amount of tunable parameters (5.7% extra trainable parameters of each task). EVP also achieves state-of-the-art performances on diverse low-level structure segmentation tasks compared to task-specific solutions. Our code is available at: https://github.com/NiFangBaAGe/Explicit-Visual-Prompt.

Recently, open-vocabulary learning has emerged to accomplish segmentation for arbitrary categories of text-based descriptions, which popularizes the segmentation system to more general-purpose application scenarios. However, existing methods devote to designing specialized architectures or parameters for specific segmentation tasks. These customized design paradigms lead to fragmentation between various segmentation tasks, thus hindering the uniformity of segmentation models. Hence in this paper, we propose FreeSeg, a generic framework to accomplish Unified, Universal and Open-Vocabulary Image Segmentation. FreeSeg optimizes an all-in-one network via one-shot training and employs the same architecture and parameters to handle diverse segmentation tasks seamlessly in the inference procedure. Additionally, adaptive prompt learning facilitates the unified model to capture task-aware and category-sensitive concepts, improving model robustness in multi-task and varied scenarios. Extensive experimental results demonstrate that FreeSeg establishes new state-of-the-art results in performance and generalization on three segmentation tasks, which outperforms the best task-specific architectures by a large margin: 5.5% mIoU on semantic segmentation, 17.6% mAP on instance segmentation, 20.1% PQ on panoptic segmentation for the unseen class on COCO. Project page: https://FreeSeg.github.io.

Referring image segmentation (RIS) aims to find a segmentation mask given a referring expression grounded to a region of the input image. Collecting labelled datasets for this task, however, is notoriously costly and labor-intensive. To overcome this issue, we propose a simple yet effective zero-shot referring image segmentation method by leveraging the pre-trained cross-modal knowledge from CLIP. In order to obtain segmentation masks grounded to the input text, we propose a mask-guided visual encoder that captures global and local contextual information of an input image. By utilizing instance masks obtained from off-the-shelf mask proposal techniques, our method is able to segment fine-detailed instance-level groundings. We also introduce a global-local text encoder where the global feature captures complex sentence-level semantics of the entire input expression while the local feature focuses on the target noun phrase extracted by a dependency parser. In our experiments, the proposed method outperforms several zero-shot baselines of the task and even the weakly supervised referring expression segmentation method with substantial margins. Our code is available at https://github.com/Seonghoon-Yu/Zero-shot-RIS.

We present DejaVu, a novel framework which leverages conditional image regeneration as additional supervision during training to improve deep networks for dense prediction tasks such as segmentation, depth estimation, and surface normal prediction. First, we apply redaction to the input image, which removes certain structural information by sparse sampling or selective frequency removal. Next, we use a conditional regenerator, which takes the redacted image and the dense predictions as inputs, and reconstructs the original image by filling in the missing structural information. In the redacted image, structural attributes like boundaries are broken while semantic context is largely preserved. In order to make the regeneration feasible, the conditional generator will then require the structure information from the other input source, i.e., the dense predictions. As such, by including this conditional regeneration objective during training, DejaVu encourages the base network to learn to embed accurate scene structure in its dense prediction. This leads to more accurate predictions with clearer boundaries and better spatial consistency. When it is feasible to leverage additional computation, DejaVu can be extended to incorporate an attention-based regeneration module within the dense prediction network, which further improves accuracy. Through extensive experiments on multiple dense prediction benchmarks such as Cityscapes, COCO, ADE20K, NYUD-v2, and KITTI, we demonstrate the efficacy of employing DejaVu during training, as it outperforms SOTA methods at no added computation cost.

Referring expression segmentation aims to segment an object described by a language expression from an image. Despite the recent progress on this task, existing models tackling this task may not be able to fully capture semantics and visual representations of individual concepts, which limits their generalization capability, especially when handling novel compositions of learned concepts. In this work, through the lens of meta learning, we propose a Meta Compositional Referring Expression Segmentation (MCRES) framework to enhance model compositional generalization performance. Specifically, to handle various levels of novel compositions, our framework first uses training data to construct a virtual training set and multiple virtual testing sets, where data samples in each virtual testing set contain a level of novel compositions w.r.t. the support set. Then, following a novel meta optimization scheme to optimize the model to obtain good testing performance on the virtual testing sets after training on the virtual training set, our framework can effectively drive the model to better capture semantics and visual representations of individual concepts, and thus obtain robust generalization performance even when handling novel compositions. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our framework.

Click-based interactive segmentation (IS) aims to extract the target objects under user interaction. For this task, most of the current deep learning (DL)-based methods mainly follow the general pipelines of semantic segmentation. Albeit achieving promising performance, they do not fully and explicitly utilize and propagate the click information, inevitably leading to unsatisfactory segmentation results, even at clicked points. Against this issue, in this paper, we propose to formulate the IS task as a Gaussian process (GP)-based pixel-wise binary classification model on each image. To solve this model, we utilize amortized variational inference to approximate the intractable GP posterior in a data-driven manner and then decouple the approximated GP posterior into double space forms for efficient sampling with linear complexity. Then, we correspondingly construct a GP classification framework, named GPCIS, which is integrated with the deep kernel learning mechanism for more flexibility. The main specificities of the proposed GPCIS lie in: 1) Under the explicit guidance of the derived GP posterior, the information contained in clicks can be finely propagated to the entire image and then boost the segmentation; 2) The accuracy of predictions at clicks has good theoretical support. These merits of GPCIS as well as its good generality and high efficiency are substantiated by comprehensive experiments on several benchmarks, as compared with representative methods both quantitatively and qualitatively. Codes will be released at https://github.com/zmhhmz/GPCIS_CVPR2023.

Zero-shot instance segmentation aims to detect and precisely segment objects of unseen categories without any training samples. Since the model is trained on seen categories, there is a strong bias that the model tends to classify all the objects into seen categories. Besides, there is a natural confusion between background and novel objects that have never shown up in training. These two challenges make novel objects hard to be raised in the final instance segmentation results. It is desired to rescue novel objects from background and dominated seen categories. To this end, we propose D^2Zero with Semantic-Promoted Debiasing and Background Disambiguation to enhance the performance of Zero-shot instance segmentation. Semantic-promoted debiasing utilizes inter-class semantic relationships to involve unseen categories in visual feature training and learns an input-conditional classifier to conduct dynamical classification based on the input image. Background disambiguation produces image-adaptive background representation to avoid mistaking novel objects for background. Extensive experiments show that we significantly outperform previous state-of-the-art methods by a large margin, e.g., 16.86% improvement on COCO.

Deep neural networks for scene perception in automated vehicles achieve excellent results for the domains they were trained on. However, in real-world conditions, the domain of operation and its underlying data distribution are subject to change. Adverse weather conditions, in particular, can significantly decrease model performance when such data are not available during training. Additionally, when a model is incrementally adapted to a new domain, it suffers from catastrophic forgetting, causing a significant drop in performance on previously observed domains. Despite recent progress in reducing catastrophic forgetting, its causes and effects remain obscure. Therefore, we study how the representations of semantic segmentation models are affected during domain-incremental learning in adverse weather conditions. Our experiments and representational analyses indicate that catastrophic forgetting is primarily caused by changes to low-level features in domain-incremental learning and that learning more general features on the source domain using pre-training and image augmentations leads to efficient feature reuse in subsequent tasks, which drastically reduces catastrophic forgetting. These findings highlight the importance of methods that facilitate generalized features for effective continual learning algorithms.

Pointly supervised instance segmentation (PSIS) learns to segment objects using a single point within the object extent as supervision. Challenged by the non-negligible semantic variance between object parts, however, the single supervision point causes semantic bias and false segmentation. In this study, we propose an AttentionShift method, to solve the semantic bias issue by iteratively decomposing the instance attention map to parts and estimating fine-grained semantics of each part. AttentionShift consists of two modules plugged on the vision transformer backbone: (i) token querying for pointly supervised attention map generation, and (ii) key-point shift, which re-estimates part-based attention maps by key-point filtering in the feature space. These two steps are iteratively performed so that the part-based attention maps are optimized spatially as well as in the feature space to cover full object extent. Experiments on PASCAL VOC and MS COCO 2017 datasets show that AttentionShift respectively improves the state-of-the-art of by 7.7% and 4.8% under mAP@0.5, setting a solid PSIS baseline using vision transformer. Code is enclosed in the supplementary material.

Two-branch network architecture has shown its efficiency and effectiveness in real-time semantic segmentation tasks. However, direct fusion of high-resolution details and low-frequency context has the drawback of detailed features being easily overwhelmed by surrounding contextual information. This overshoot phenomenon limits the improvement of the segmentation accuracy of existing two-branch models. In this paper, we make a connection between Convolutional Neural Networks (CNN) and Proportional-Integral-Derivative (PID) controllers and reveal that a two-branch network is equivalent to a Proportional-Integral (PI) controller, which inherently suffers from similar overshoot issues. To alleviate this problem, we propose a novel three-branch network architecture: PIDNet, which contains three branches to parse detailed, context and boundary information, respectively, and employs boundary attention to guide the fusion of detailed and context branches. Our family of PIDNets achieve the best trade-off between inference speed and accuracy and their accuracy surpasses all the existing models with similar inference speed on the Cityscapes and CamVid datasets. Specifically, PIDNet-S achieves 78.6 mIOU with inference speed of 93.2 FPS on Cityscapes and 80.1 mIOU with speed of 153.7 FPS on CamVid.

Dramatic demand for manpower to label pixel-level annotations triggered the advent of unsupervised semantic segmentation. Although the recent work employing the vision transformer (ViT) backbone shows exceptional performance, there is still a lack of consideration for task-specific training guidance and local semantic consistency. To tackle these issues, we leverage contrastive learning by excavating hidden positives to learn rich semantic relationships and ensure semantic consistency in local regions. Specifically, we first discover two types of global hidden positives, task-agnostic and task-specific ones for each anchor based on the feature similarities defined by a fixed pre-trained backbone and a segmentation head-in-training, respectively. A gradual increase in the contribution of the latter induces the model to capture task-specific semantic features. In addition, we introduce a gradient propagation strategy to learn semantic consistency between adjacent patches, under the inherent premise that nearby patches are highly likely to possess the same semantics. Specifically, we add the loss propagating to local hidden positives, semantically similar nearby patches, in proportion to the predefined similarity scores. With these training schemes, our proposed method achieves new state-of-the-art (SOTA) results in COCO-stuff, Cityscapes, and Potsdam-3 datasets. Our code is available at: https://github.com/hynnsk/HP.

A recent study has shown a phenomenon called neural collapse in that the within-class means of features and the classifier weight vectors converge to the vertices of a simplex equiangular tight frame at the terminal phase of training for classification. In this paper, we explore the corresponding structures of the last-layer feature centers and classifiers in semantic segmentation. Based on our empirical and theoretical analysis, we point out that semantic segmentation naturally brings contextual correlation and imbalanced distribution among classes, which breaks the equiangular and maximally separated structure of neural collapse for both feature centers and classifiers. However, such a symmetric structure is beneficial to discrimination for the minor classes. To preserve these advantages, we introduce a regularizer on feature centers to encourage the network to learn features closer to the appealing structure in imbalanced semantic segmentation. Experimental results show that our method can bring significant improvements on both 2D and 3D semantic segmentation benchmarks. Moreover, our method ranks first and sets a new record (+6.8% mIoU) on the ScanNet200 test leaderboard.

Semantic segmentation usually suffers from a long tail data distribution. Due to the imbalanced number of samples across categories, the features of those tail classes may get squeezed into a narrow area in the feature space. Towards a balanced feature distribution, we introduce category-wise variation into the network predictions in the training phase such that an instance is no longer projected to a feature point, but a small region instead. Such a perturbation is highly dependent on the category scale, which appears as assigning smaller variation to head classes and larger variation to tail classes. In this way, we manage to close the gap between the feature areas of different categories, resulting in a more balanced representation. It is noteworthy that the introduced variation is discarded at the inference stage to facilitate a confident prediction. Although with an embarrassingly simple implementation, our method manifests itself in strong generalizability to various datasets and task settings. Extensive experiments suggest that our plug-in design lends itself well to a range of state-of-the-art approaches and boosts the performance on top of them.

The existing weakly supervised semantic segmentation (WSSS) methods pay much attention to generating accurate and complete class activation maps (CAMs) as pseudo-labels, while ignoring the importance of training the segmentation networks. In this work, we observe that there is an inconsistency between the quality of the pseudo-labels in CAMs and the performance of the final segmentation model, and the mislabeled pixels mainly lie on the boundary areas. Inspired by these findings, we argue that the focus of WSSS should be shifted to robust learning given the noisy pseudo-labels, and further propose a boundary-enhanced co-training (BECO) method for training the segmentation model. To be specific, we first propose to use a co-training paradigm with two interactive networks to improve the learning of uncertain pixels. Then we propose a boundary-enhanced strategy to boost the prediction of difficult boundary areas, which utilizes reliable predictions to construct artificial boundaries. Benefiting from the design of co-training and boundary enhancement, our method can achieve promising segmentation performance for different CAMs. Extensive experiments on PASCAL VOC 2012 and MS COCO 2014 validate the superiority of our BECO over other state-of-the-art methods.

Semi-supervised semantic segmentation (SSS) has recently gained increasing research interest as it can reduce the requirement for large-scale fully-annotated training data. The current methods often suffer from the confirmation bias from the pseudo-labelling process, which can be alleviated by the co-training framework. The current co-training-based SSS methods rely on hand-crafted perturbations to prevent the different sub-nets from collapsing into each other, but these artificial perturbations cannot lead to the optimal solution. In this work, we propose a new conflict-based cross-view consistency (CCVC) method based on a two-branch co-training framework which aims at enforcing the two sub-nets to learn informative features from irrelevant views. In particular, we first propose a new cross-view consistency (CVC) strategy that encourages the two sub-nets to learn distinct features from the same input by introducing a feature discrepancy loss, while these distinct features are expected to generate consistent prediction scores of the input. The CVC strategy helps to prevent the two sub-nets from stepping into the collapse. In addition, we further propose a conflict-based pseudo-labelling (CPL) method to guarantee the model will learn more useful information from conflicting predictions, which will lead to a stable training process. We validate our new CCVC approach on the SSS benchmark datasets where our method achieves new state-of-the-art performance. Our code is available at https://github.com/xiaoyao3302/CCVC.

Weakly supervised dense object localization (WSDOL) relies generally on Class Activation Mapping (CAM), which exploits the correlation between the class weights of the image classifier and the pixel-level features. Due to the limited ability to address intra-class variations, the image classifier cannot properly associate the pixel features, leading to inaccurate dense localization maps. In this paper, we propose to explicitly construct multi-modal class representations by leveraging the Contrastive Language-Image Pre-training (CLIP), to guide dense localization. More specifically, we propose a unified transformer framework to learn two-modalities of class-specific tokens, i.e., class-specific visual and textual tokens. The former captures semantics from the target visual data while the latter exploits the class-related language priors from CLIP, providing complementary information to better perceive the intra-class diversities. In addition, we propose to enrich the multi-modal class-specific tokens with sample-specific contexts comprising visual context and image-language context. This enables more adaptive class representation learning, which further facilitates dense localization. Extensive experiments show the superiority of the proposed method for WSDOL on two multi-label datasets, i.e., PASCAL VOC and MS COCO, and one single-label dataset, i.e., OpenImages. Our dense localization maps also lead to the state-of-the-art weakly supervised semantic segmentation (WSSS) results on PASCAL VOC and MS COCO.

Visual anomaly classification and segmentation are vital for automating industrial quality inspection. The focus of prior research in the field has been on training custom models for each quality inspection task, which requires task-specific images and annotation. In this paper we move away from this regime, addressing zero-shot and few-normal-shot anomaly classification and segmentation. Recently CLIP, a vision-language model, has shown revolutionary generality with competitive zero/few-shot performance in comparison to full-supervision. But CLIP falls short on anomaly classification and segmentation tasks. Hence, we propose window-based CLIP (WinCLIP) with (1) a compositional ensemble on state words and prompt templates and (2) efficient extraction and aggregation of window/patch/image-level features aligned with text. We also propose its few-normal-shot extension WinCLIP+, which uses complementary information from normal images. In MVTec-AD (and VisA), without further tuning, WinCLIP achieves 91.8%/85.1% (78.1%/79.6%) AUROC in zero-shot anomaly classification and segmentation while WinCLIP+ does 93.1%/95.2% (83.8%/96.4%) in 1-normal-shot, surpassing state-of-the-art by large margins.

Automatic mitochondria segmentation enjoys great popularity with the development of deep learning. However, existing methods rely heavily on the labor-intensive manual gathering by experienced domain experts. And naively applying semi-supervised segmentation methods in the natural image field to mitigate the labeling cost is undesirable. In this work, we analyze the gap between mitochondrial images and natural images and rethink how to achieve effective semi-supervised mitochondria segmentation, from the perspective of reliable prototype-level supervision. We propose a novel end-to-end dual-reliable (DualRel) network, including a reliable pixel aggregation module and a reliable prototype selection module. The proposed DualRel enjoys several merits. First, to learn the prototypes well without any explicit supervision, we carefully design the referential correlation to rectify the direct pairwise correlation. Second, the reliable prototype selection module is responsible for further evaluating the reliability of prototypes in constructing prototype-level consistency regularization. Extensive experimental results on three challenging benchmarks demonstrate that our method performs favorably against state-of-the-art semi-supervised segmentation methods. Importantly, with extremely few samples used for training, DualRel is also on par with current state-of-the-art fully supervised methods.

We address the task of weakly-supervised few-shot image classification and segmentation, by leveraging a Vision Transformer (ViT) pretrained with self-supervision. Our proposed method takes token representations from the self-supervised ViT and leverages their correlations, via self-attention, to produce classification and segmentation predictions through separate task heads. Our model is able to effectively learn to perform classification and segmentation in the absence of pixel-level labels during training, using only image-level labels. To do this it uses attention maps, created from tokens generated by the self-supervised ViT backbone, as pixel-level pseudo-labels. We also explore a practical setup with “mixed” supervision, where a small number of training images contains ground-truth pixel-level labels and the remaining images have only image-level labels. For this mixed setup, we propose to improve the pseudo-labels using a pseudo-label enhancer that was trained using the available ground-truth pixel-level labels. Experiments on Pascal-5i and COCO-20i demonstrate significant performance gains in a variety of supervision settings, and in particular when little-to-no pixel-level labels are available.

The traditional definition of co-salient object detection (CoSOD) task is to segment the common salient objects in a group of relevant images. Existing CoSOD models by default adopt the group consensus assumption. This brings about model robustness defect under the condition of irrelevant images in the testing image group, which hinders the use of CoSOD models in real-world applications. To address this issue, this paper presents a group exchange-masking (GEM) strategy for robust CoSOD model learning. With two group of image containing different types of salient object as input, the GEM first selects a set of images from each group by the proposed learning based strategy, then these images are exchanged. The proposed feature extraction module considers both the uncertainty caused by the irrelevant images and group consensus in the remaining relevant images. We design a latent variable generator branch which is made of conditional variational autoencoder to generate uncertainly-based global stochastic features. A CoSOD transformer branch is devised to capture the correlation-based local features that contain the group consistency information. At last, the output of two branches are concatenated and fed into a transformer-based decoder, producing robust co-saliency prediction. Extensive evaluations on co-saliency detection with and without irrelevant images demonstrate the superiority of our method over a variety of state-of-the-art methods.

Vision Transformers (ViTs) emerge to achieve impressive performance on many data-abundant computer vision tasks by capturing long-range dependencies among local features. However, under few-shot learning (FSL) settings on small datasets with only a few labeled data, ViT tends to overfit and suffers from severe performance degradation due to its absence of CNN-alike inductive bias. Previous works in FSL avoid such problem either through the help of self-supervised auxiliary losses, or through the dextile uses of label information under supervised settings. But the gap between self-supervised and supervised few-shot Transformers is still unfilled. Inspired by recent advances in self-supervised knowledge distillation and masked image modeling (MIM), we propose a novel Supervised Masked Knowledge Distillation model (SMKD) for few-shot Transformers which incorporates label information into self-distillation frameworks. Compared with previous self-supervised methods, we allow intra-class knowledge distillation on both class and patch tokens, and introduce the challenging task of masked patch tokens reconstruction across intra-class images. Experimental results on four few-shot classification benchmark datasets show that our method with simple design outperforms previous methods by a large margin and achieves a new start-of-the-art. Detailed ablation studies confirm the effectiveness of each component of our model. Code for this paper is available here: https://github.com/HL-hanlin/SMKD.

Most of the existing salient object detection (SOD) models focus on improving the overall model performance, without explicitly explaining the discrepancy between the training and testing distributions. In this paper, we investigate a particular type of epistemic uncertainty, namely distributional uncertainty, for salient object detection. Specifically, for the first time, we explore the existing class-aware distribution gap exploration techniques, i.e. long-tail learning, single-model uncertainty modeling and test-time strategies, and adapt them to model the distributional uncertainty for our class-agnostic task. We define test sample that is dissimilar to the training dataset as being “out-of-distribution” (OOD) samples. Different from the conventional OOD definition, where OOD samples are those not belonging to the closed-world training categories, OOD samples for SOD are those break the basic priors of saliency, i.e. center prior, color contrast prior, compactness prior and etc., indicating OOD as being “continuous” instead of being discrete for our task. We’ve carried out extensive experimental results to verify effectiveness of existing distribution gap modeling techniques for SOD, and conclude that both train-time single-model uncertainty estimation techniques and weight-regularization solutions that preventing model activation from drifting too much are promising directions for modeling distributional uncertainty for SOD.

Weak-shot Object Detection methods exploit a fully-annotated source dataset to facilitate the detection performance on the target dataset which only contains image-level labels for novel categories. To bridge the gap between these two datasets, we aim to transfer the object knowledge between the source (S) and target (T) datasets in a bi-directional manner. We propose a novel Knowledge Transfer (KT) loss which simultaneously distills the knowledge of objectness and class entropy from a proposal generator trained on the S dataset to optimize a multiple instance learning module on the T dataset. By jointly optimizing the classification loss and the proposed KT loss, the multiple instance learning module effectively learns to classify object proposals into novel categories in the T dataset with the transferred knowledge from base categories in the S dataset. Noticing the predicted boxes on the T dataset can be regarded as an extension for the original annotations on the S dataset to refine the proposal generator in return, we further propose a novel Consistency Filtering (CF) method to reliably remove inaccurate pseudo labels by evaluating the stability of the multiple instance learning module upon noise injections. Via mutually transferring knowledge between the S and T datasets in an iterative manner, the detection performance on the target dataset is significantly improved. Extensive experiments on public benchmarks validate that the proposed method performs favourably against the state-of-the-art methods without increasing the model parameters or inference computational complexity.

Open-world object detection (OWOD), as a more general and challenging goal, requires the model trained from data on known objects to detect both known and unknown objects and incrementally learn to identify these unknown objects. The existing works which employ standard detection framework and fixed pseudo-labelling mechanism (PLM) have the following problems: (i) The inclusion of detecting unknown objects substantially reduces the model’s ability to detect known ones. (ii) The PLM does not adequately utilize the priori knowledge of inputs. (iii) The fixed selection manner of PLM cannot guarantee that the model is trained in the right direction. We observe that humans subconsciously prefer to focus on all foreground objects and then identify each one in detail, rather than localize and identify a single object simultaneously, for alleviating the confusion. This motivates us to propose a novel solution called CAT: LoCalization and IdentificAtion Cascade Detection Transformer which decouples the detection process via the shared decoder in the cascade decoding way. In the meanwhile, we propose the self-adaptive pseudo-labelling mechanism which combines the model-driven with input-driven PLM and self-adaptively generates robust pseudo-labels for unknown objects, significantly improving the ability of CAT to retrieve unknown objects. Comprehensive experiments on two benchmark datasets, i.e., MS-COCO and PASCAL VOC, show that our model outperforms the state-of-the-art in terms of all metrics in the task of OWOD, incremental object detection (IOD) and open-set detection.

Object re-identification (ReID) aims to find instances with the same identity as the given probe from a large gallery. Pairwise losses play an important role in training a strong ReID network. Existing pairwise losses densely exploit each instance as an anchor and sample its triplets in a mini-batch. This dense sampling mechanism inevitably introduces positive pairs that share few visual similarities, which can be harmful to the training. To address this problem, we propose a novel loss paradigm termed Sparse Pairwise (SP) loss that only leverages few appropriate pairs for each class in a mini-batch, and empirically demonstrate that it is sufficient for the ReID tasks. Based on the proposed loss framework, we propose an adaptive positive mining strategy that can dynamically adapt to diverse intra-class variations. Extensive experiments show that SP loss and its adaptive variant AdaSP loss outperform other pairwise losses, and achieve state-of-the-art performance across several ReID benchmarks. Code is available at https://github.com/Astaxanthin/AdaSP.

One-to-one set matching is a key design for DETR to establish its end-to-end capability, so that object detection does not require a hand-crafted NMS (non-maximum suppression) to remove duplicate detections. This end-to-end signature is important for the versatility of DETR, and it has been generalized to broader vision tasks. However, we note that there are few queries assigned as positive samples and the one-to-one set matching significantly reduces the training efficacy of positive samples. We propose a simple yet effective method based on a hybrid matching scheme that combines the original one-to-one matching branch with an auxiliary one-to-many matching branch during training. Our hybrid strategy has been shown to significantly improve accuracy. In inference, only the original one-to-one match branch is used, thus maintaining the end-to-end merit and the same inference efficiency of DETR. The method is named H-DETR, and it shows that a wide range of representative DETR methods can be consistently improved across a wide range of visual tasks, including Deformable-DETR, PETRv2, PETR, and TransTrack, among others.

Two-stage object detectors generate object proposals and classify them to detect objects in images. These proposals often do not perfectly contain the objects but overlap with them in many possible ways, exhibiting great variability in the difficulty levels of the proposals. Training a robust classifier against this crop-related variability requires abundant training data, which is not available in few-shot settings. To mitigate this issue, we propose a novel variational autoencoder (VAE) based data generation model, which is capable of generating data with increased crop-related diversity. The main idea is to transform the latent space such the latent codes with different norms represent different crop-related variations. This allows us to generate features with increased crop-related diversity in difficulty levels by simply varying the latent norm. In particular, each latent code is rescaled such that its norm linearly correlates with the IoU score of the input crop w.r.t. the ground-truth box. Here the IoU score is a proxy that represents the difficulty level of the crop. We train this VAE model on base classes conditioned on the semantic code of each class and then use the trained model to generate features for novel classes. Our experimental results show that our generated features consistently improve state-of-the-art few-shot object detection methods on PASCAL VOC and MS COCO datasets.

Despite the prominent success of general object detection, the performance and efficiency of Small Object Detection (SOD) are still unsatisfactory. Unlike existing works that struggle to balance the trade-off between inference speed and SOD performance, in this paper, we propose a novel Scale-aware Knowledge Distillation (ScaleKD), which transfers knowledge of a complex teacher model to a compact student model. We design two novel modules to boost the quality of knowledge transfer in distillation for SOD: 1) a scale-decoupled feature distillation module that disentangled teacher’s feature representation into multi-scale embedding that enables explicit feature mimicking of the student model on small objects. 2) a cross-scale assistant to refine the noisy and uninformative bounding boxes prediction student models, which can mislead the student model and impair the efficacy of knowledge distillation. A multi-scale cross-attention layer is established to capture the multi-scale semantic information to improve the student model. We conduct experiments on COCO and VisDrone datasets with diverse types of models, i.e., two-stage and one-stage detectors, to evaluate our proposed method. Our ScaleKD achieves superior performance on general detection performance and obtains spectacular improvement regarding the SOD performance.

Albeit achieving high predictive accuracy across many challenging computer vision problems, recent studies suggest that deep neural networks (DNNs) tend to make overconfident predictions, rendering them poorly calibrated. Most of the existing attempts for improving DNN calibration are limited to classification tasks and restricted to calibrating in-domain predictions. Surprisingly, very little to no attempts have been made in studying the calibration of object detection methods, which occupy a pivotal space in vision-based security-sensitive, and safety-critical applications. In this paper, we propose a new train-time technique for calibrating modern object detection methods. It is capable of jointly calibrating multiclass confidence and box localization by leveraging their predictive uncertainties. We perform extensive experiments on several in-domain and out-of-domain detection benchmarks. Results demonstrate that our proposed train-time calibration method consistently outperforms several baselines in reducing calibration error for both in-domain and out-of-domain predictions. Our code and models are available at https://github.com/bimsarapathiraja/MCCL

Visual recognition tasks are often limited to dealing with a small subset of classes simply because the labels for the remaining classes are unavailable. We are interested in identifying novel concepts in a dataset through representation learning based on both labeled and unlabeled examples, and extending the horizon of recognition to both known and novel classes. To address this challenging task, we propose a combinatorial learning approach, which naturally clusters the examples in unseen classes using the compositional knowledge given by multiple supervised meta-classifiers on heterogeneous label spaces. The representations given by the combinatorial embedding are made more robust by unsupervised pairwise relation learning. The proposed algorithm discovers novel concepts via a joint optimization for enhancing the discrimitiveness of unseen classes as well as learning the representations of known classes generalizable to novel ones. Our extensive experiments demonstrate remarkable performance gains by the proposed approach on public datasets for image retrieval and image categorization with novel class discovery.

This paper considers few-shot image classification under the cross-domain scenario, where the train-to-test domain gap compromises classification accuracy. To mitigate the domain gap, we propose a prompting-to-disentangle (ProD) method through a novel exploration with the prompting mechanism. ProD adopts the popular multi-domain training scheme and extracts the backbone feature with a standard Convolutional Neural Network. Based on these two common practices, the key point of ProD is using the prompting mechanism in the transformer to disentangle the domain-general (DG) and domain-specific (DS) knowledge from the backbone feature. Specifically, ProD concatenates a DG and a DS prompt to the backbone feature and feeds them into a lightweight transformer. The DG prompt is learnable and shared by all the training domains, while the DS prompt is generated from the domain-of-interest on the fly. As a result, the transformer outputs DG and DS features in parallel with the two prompts, yielding the disentangling effect. We show that: 1) Simply sharing a single DG prompt for all the training domains already improves generalization towards the novel test domain. 2) The cross-domain generalization can be further reinforced by making the DG prompt neutral towards the training domains. 3) When inference, the DS prompt is generated from the support samples and can capture test domain knowledge through the prompting mechanism. Combining all three benefits, ProD significantly improves cross-domain few-shot classification. For instance, on CUB, ProD improves the 5-way 5-shot accuracy from 73.56% (baseline) to 79.19%, setting a new state of the art.

Contrastive visual language pretraining has emerged as a powerful method for either training new language-aware image encoders or augmenting existing pretrained models with zero-shot visual recognition capabilities. However, existing works typically train on large datasets of image-text pairs and have been designed to perform downstream tasks involving only small to medium sized-images, neither of which are applicable to the emerging field of computational pathology where there are limited publicly available paired image-text datasets and each image can span up to 100,000 x 100,000 pixels in dimensions. In this paper we present MI-Zero, a simple and intuitive framework for unleashing the zero-shot transfer capabilities of contrastively aligned image and text models to gigapixel histopathology whole slide images, enabling multiple downstream diagnostic tasks to be carried out by pretrained encoders without requiring any additional labels. MI-Zero reformulates zero-shot transfer under the framework of multiple instance learning to overcome the computational challenge of inference on extremely large images. We used over 550k pathology reports and other available in-domain text corpora to pretrain our text encoder. By effectively leveraging strong pretrained encoders, our best model pretrained on over 33k histopathology image-caption pairs achieves an average median zero-shot accuracy of 70.2% across three different real-world cancer subtyping tasks. Our code is available at: https://github.com/mahmoodlab/MI-Zero.

Cryo-electron microscopy (cryo-EM) is a technique for reconstructing the 3-dimensional (3D) structure of biomolecules (especially large protein complexes and molecular assemblies). As the resolution increases to the near-atomic scale, building protein structures de novo from cryo-EM maps becomes possible. Recently, recognition-based de novo building methods have shown the potential to streamline this process. However, it cannot build a complete structure due to the low signal-to-noise ratio (SNR) problem. At the same time, AlphaFold has led to a great breakthrough in predicting protein structures. This has inspired us to combine fragment recognition and structure prediction methods to build a complete structure. In this paper, we propose a new method named FFF that bridges protein structure prediction and protein structure recognition with flexible fitting. First, a multi-level recognition network is used to capture various structural features from the input 3D cryo-EM map. Next, protein structural fragments are generated using pseudo peptide vectors and a protein sequence alignment method based on these extracted features. Finally, a complete structural model is constructed using the predicted protein fragments via flexible fitting. Based on our benchmark tests, FFF outperforms the baseline meth- ods for building complete protein structures.

Although recent works in semi-supervised learning (SemiSL) have accomplished significant success in natural image segmentation, the task of learning discriminative representations from limited annotations has been an open problem in medical images. Contrastive Learning (CL) frameworks use the notion of similarity measure which is useful for classification problems, however, they fail to transfer these quality representations for accurate pixel-level segmentation. To this end, we propose a novel semi-supervised patch-based CL framework for medical image segmentation without using any explicit pretext task. We harness the power of both CL and SemiSL, where the pseudo-labels generated from SemiSL aid CL by providing additional guidance, whereas discriminative class information learned in CL leads to accurate multi-class segmentation. Additionally, we formulate a novel loss that synergistically encourages inter-class separability and intra-class compactness among the learned representations. A new inter-patch semantic disparity mapping using average patch entropy is employed for a guided sampling of positives and negatives in the proposed CL framework. Experimental analysis on three publicly available datasets of multiple modalities reveals the superiority of our proposed method as compared to the state-of-the-art methods. Code is available at: https://github.com/hritam-98/PatchCL-MedSeg.

Learning high-quality, self-supervised, visual representations is essential to advance the role of computer vision in biomedical microscopy and clinical medicine. Previous work has focused on self-supervised representation learning (SSL) methods developed for instance discrimination and applied them directly to image patches, or fields-of-view, sampled from gigapixel whole-slide images (WSIs) used for cancer diagnosis. However, this strategy is limited because it (1) assumes patches from the same patient are independent, (2) neglects the patient-slide-patch hierarchy of clinical biomedical microscopy, and (3) requires strong data augmentations that can degrade downstream performance. Importantly, sampled patches from WSIs of a patient’s tumor are a diverse set of image examples that capture the same underlying cancer diagnosis. This motivated HiDisc, a data-driven method that leverages the inherent patient-slide-patch hierarchy of clinical biomedical microscopy to define a hierarchical discriminative learning task that implicitly learns features of the underlying diagnosis. HiDisc uses a self-supervised contrastive learning framework in which positive patch pairs are defined based on a common ancestry in the data hierarchy, and a unified patch, slide, and patient discriminative learning objective is used for visual SSL. We benchmark HiDisc visual representations on two vision tasks using two biomedical microscopy datasets, and demonstrate that (1) HiDisc pretraining outperforms current state-of-the-art self-supervised pretraining methods for cancer diagnosis and genetic mutation prediction, and (2) HiDisc learns high-quality visual representations using natural patch diversity without strong data augmentations.

Radiology report generation aims to automatically generate a clinically accurate and coherent paragraph from the X-ray image, which could relieve radiologists from the heavy burden of report writing. Although various image caption methods have shown remarkable performance in the natural image field, generating accurate reports for medical images requires knowledge of multiple modalities, including vision, language, and medical terminology. We propose a Knowledge-injected U-Transformer (KiUT) to learn multi-level visual representation and adaptively distill the information with contextual and clinical knowledge for word prediction. In detail, a U-connection schema between the encoder and decoder is designed to model interactions between different modalities. And a symptom graph and an injected knowledge distiller are developed to assist the report generation. Experimentally, we outperform state-of-the-art methods on two widely used benchmark datasets: IU-Xray and MIMIC-CXR. Further experimental results prove the advantages of our architecture and the complementary benefits of the injected knowledge.

Medical images usually suffer from image degradation in clinical practice, leading to decreased performance of deep learning-based models. To resolve this problem, most previous works have focused on filtering out degradation-causing low-quality images while ignoring their potential value for models. Through effectively learning and leveraging the knowledge of degradations, models can better resist their adverse effects and avoid misdiagnosis. In this paper, we raise the problem of image quality-aware diagnosis, which aims to take advantage of low-quality images and image quality labels to achieve a more accurate and robust diagnosis. However, the diversity of degradations and superficially unrelated targets between image quality assessment and disease diagnosis makes it still quite challenging to effectively leverage quality labels to assist diagnosis. Thus, to tackle these issues, we propose a novel meta-knowledge co-embedding network, consisting of two subnets: Task Net and Meta Learner. Task Net constructs an explicit quality information utilization mechanism to enhance diagnosis via knowledge co-embedding features, while Meta Learner ensures the effectiveness and constrains the semantics of these features via meta-learning and joint-encoding masking. Superior performance on five datasets with four widely-used medical imaging modalities demonstrates the effectiveness and generalizability of our method.

Multi-instance learning (MIL) is an effective paradigm for whole-slide pathological images (WSIs) classification to handle the gigapixel resolution and slide-level label. Prevailing MIL methods primarily focus on improving the feature extractor and aggregator. However, one deficiency of these methods is that the bag contextual prior may trick the model into capturing spurious correlations between bags and labels. This deficiency is a confounder that limits the performance of existing MIL methods. In this paper, we propose a novel scheme, Interventional Bag Multi-Instance Learning (IBMIL), to achieve deconfounded bag-level prediction. Unlike traditional likelihood-based strategies, the proposed scheme is based on the backdoor adjustment to achieve the interventional training, thus is capable of suppressing the bias caused by the bag contextual prior. Note that the principle of IBMIL is orthogonal to existing bag MIL methods. Therefore, IBMIL is able to bring consistent performance boosting to existing schemes, achieving new state-of-the-art performance. Code is available at https://github.com/HHHedo/IBMIL.

Learning generative image models from various domains efficiently needs transferring knowledge from an image synthesis model trained on a large dataset. We present a recipe for learning vision transformers by generative knowledge transfer. We base our framework on generative vision transformers representing an image as a sequence of visual tokens with the autoregressive or non-autoregressive transformers. To adapt to a new domain, we employ prompt tuning, which prepends learnable tokens called prompts to the image token sequence and introduces a new prompt design for our task. We study on a variety of visual domains with varying amounts of training images. We show the effectiveness of knowledge transfer and a significantly better image generation quality. Code is available at https://github.com/google-research/generative_transfer.

It is important to quantify the uncertainty of input samples, especially in mission-critical domains such as autonomous driving and healthcare, where failure predictions on out-of-distribution (OOD) data are likely to cause big problems. OOD detection problem fundamentally begins in that the model cannot express what it is not aware of. Post-hoc OOD detection approaches are widely explored because they do not require an additional re-training process which might degrade the model’s performance and increase the training cost. In this study, from the perspective of neurons in the deep layer of the model representing high-level features, we introduce a new aspect for analyzing the difference in model outputs between in-distribution data and OOD data. We propose a novel method, Leveraging Important Neurons (LINe), for post-hoc Out of distribution detection. Shapley value-based pruning reduces the effects of noisy outputs by selecting only high-contribution neurons for predicting specific classes of input data and masking the rest. Activation clipping fixes all values above a certain threshold into the same value, allowing LINe to treat all the class-specific features equally and just consider the difference between the number of activated feature differences between in-distribution and OOD data. Comprehensive experiments verify the effectiveness of the proposed method by outperforming state-of-the-art post-hoc OOD detection methods on CIFAR-10, CIFAR-100, and ImageNet datasets.

Multi-view clustering can partition data samples into their categories by learning a consensus representation in unsupervised way and has received more and more attention in recent years. However, most existing deep clustering methods learn consensus representation or view-specific representations from multiple views via view-wise aggregation way, where they ignore structure relationship of all samples. In this paper, we propose a novel multi-view clustering network to address these problems, called Global and Cross-view Feature Aggregation for Multi-View Clustering (GCFAggMVC). Specifically, the consensus data presentation from multiple views is obtained via cross-sample and cross-view feature aggregation, which fully explores the complementary of similar samples. Moreover, we align the consensus representation and the view-specific representation by the structure-guided contrastive learning module, which makes the view-specific representations from different samples with high structure relationship similar. The proposed module is a flexible multi-view data representation module, which can be also embedded to the incomplete multi-view data clustering task via plugging our module into other frameworks. Extensive experiments show that the proposed method achieves excellent performance in both complete multi-view data clustering tasks and incomplete multi-view data clustering tasks.

Classifying incomplete multi-view data is inevitable since arbitrary view missing widely exists in real-world applications. Although great progress has been achieved, existing incomplete multi-view methods are still difficult to obtain a trustworthy prediction due to the relatively high uncertainty nature of missing views. First, the missing view is of high uncertainty, and thus it is not reasonable to provide a single deterministic imputation. Second, the quality of the imputed data itself is of high uncertainty. To explore and exploit the uncertainty, we propose an Uncertainty-induced Incomplete Multi-View Data Classification (UIMC) model to classify the incomplete multi-view data under a stable and reliable framework. We construct a distribution and sample multiple times to characterize the uncertainty of missing views, and adaptively utilize them according to the sampling quality. Accordingly, the proposed method realizes more perceivable imputation and controllable fusion. Specifically, we model each missing data with a distribution conditioning on the available views and thus introducing uncertainty. Then an evidence-based fusion strategy is employed to guarantee the trustworthy integration of the imputed views. Extensive experiments are conducted on multiple benchmark data sets and our method establishes a state-of-the-art performance in terms of both performance and trustworthiness.

As one of the most fundamental techniques in multimodal learning, cross-modal matching aims to project various sensory modalities into a shared feature space. To achieve this, massive and correctly aligned data pairs are required for model training. However, unlike unimodal datasets, multimodal datasets are extremely harder to collect and annotate precisely. As an alternative, the co-occurred data pairs (e.g., image-text pairs) collected from the Internet have been widely exploited in the area. Unfortunately, the cheaply collected dataset unavoidably contains many mismatched data pairs, which have been proven to be harmful to the model’s performance. To address this, we propose a general framework called BiCro (Bidirectional Cross-modal similarity consistency), which can be easily integrated into existing cross-modal matching models and improve their robustness against noisy data. Specifically, BiCro aims to estimate soft labels for noisy data pairs to reflect their true correspondence degree. The basic idea of BiCro is motivated by that -- taking image-text matching as an example -- similar images should have similar textual descriptions and vice versa. Then the consistency of these two similarities can be recast as the estimated soft labels to train the matching model. The experiments on three popular cross-modal matching datasets demonstrate that our method significantly improves the noise-robustness of various matching models, and surpass the state-of-the-art by a clear margin.

It is well-known that zero-shot learning (ZSL) can suffer severely from the problem of domain shift, where the true and learned data distributions for the unseen classes do not match. Although transductive ZSL (TZSL) attempts to improve this by allowing the use of unlabelled examples from the unseen classes, there is still a high level of distribution shift. We propose a novel TZSL model (named as Bi-VAEGAN), which largely improves the shift by a strengthened distribution alignment between the visual and auxiliary spaces. The key proposal of the model design includes (1) a bi-directional distribution alignment, (2) a simple but effective L_2-norm based feature normalization approach, and (3) a more sophisticated unseen class prior estimation approach. In benchmark evaluation using four datasets, Bi-VAEGAN achieves the new state of the arts under both the standard and generalized TZSL settings. Code could be found at https://github.com/Zhicaiwww/Bi-VAEGAN.

Supervision for metric learning has long been given in the form of equivalence between human-labeled classes. Although this type of supervision has been a basis of metric learning for decades, we argue that it hinders further advances in the field. In this regard, we propose a new regularization method, dubbed HIER, to discover the latent semantic hierarchy of training data, and to deploy the hierarchy to provide richer and more fine-grained supervision than inter-class separability induced by common metric learning losses. HIER achieves this goal with no annotation for the semantic hierarchy but by learning hierarchical proxies in hyperbolic spaces. The hierarchical proxies are learnable parameters, and each of them is trained to serve as an ancestor of a group of data or other proxies to approximate the semantic hierarchy among them. HIER deals with the proxies along with data in hyperbolic space since the geometric properties of the space are well-suited to represent their hierarchical structure. The efficacy of HIER is evaluated on four standard benchmarks, where it consistently improved the performance of conventional methods when integrated with them, and consequently achieved the best records, surpassing even the existing hyperbolic metric learning technique, in almost all settings.

Deep learning has achieved great success in recent years with the aid of advanced neural network structures and large-scale human-annotated datasets. However, it is often costly and difficult to accurately and efficiently annotate large-scale datasets, especially for some specialized domains where fine-grained labels are required. In this setting, coarse labels are much easier to acquire as they do not require expert knowledge. In this work, we propose a contrastive learning method, called masked contrastive learning (MaskCon) to address the under-explored problem setting, where we learn with a coarse-labelled dataset in order to address a finer labelling problem. More specifically, within the contrastive learning framework, for each sample our method generates soft-labels with the aid of coarse labels against other samples and another augmented view of the sample in question. By contrast to self-supervised contrastive learning where only the sample’s augmentations are considered hard positives, and in supervised contrastive learning where only samples with the same coarse labels are considered hard positives, we propose soft labels based on sample distances, that are masked by the coarse labels. This allows us to utilize both inter-sample relations and coarse labels. We demonstrate that our method can obtain as special cases many existing state-of-the-art works and that it provides tighter bounds on the generalization error. Experimentally, our method achieves significant improvement over the current state-of-the-art in various datasets, including CIFAR10, CIFAR100, ImageNet-1K, Standford Online Products and Stanford Cars196 datasets. Code and annotations are available at https://github.com/MrChenFeng/MaskCon_CVPR2023.

The recent growth in the consumption of online media by children during early childhood necessitates data-driven tools enabling educators to filter out appropriate educational content for young learners. This paper presents an approach for detecting educational content in online videos. We focus on two widely used educational content classes: literacy and math. For each class, we choose prominent codes (sub-classes) based on the Common Core Standards. For example, literacy codes include ‘letter names’, ‘letter sounds’, and math codes include ‘counting’, ‘sorting’. We pose this as a fine-grained multilabel classification problem as videos can contain multiple types of educational content and the content classes can get visually similar (e.g., ‘letter names’ vs ‘letter sounds’). We propose a novel class prototypes based supervised contrastive learning approach that can handle fine-grained samples associated with multiple labels. We learn a class prototype for each class and a loss function is employed to minimize the distances between a class prototype and the samples from the class. Similarly, distances between a class prototype and the samples from other classes are maximized. As the alignment between visual and audio cues are crucial for effective comprehension, we consider a multimodal transformer network to capture the interaction between visual and audio cues in videos while learning the embedding for videos. For evaluation, we present a dataset, APPROVE, employing educational videos from YouTube labeled with fine-grained education classes by education researchers. APPROVE consists of 193 hours of expert-annotated videos with 19 classes. The proposed approach outperforms strong baselines on APPROVE and other benchmarks such as Youtube-8M, and COIN. The dataset is available at https://nusci.csl.sri.com/project/APPROVE.

Training deep neural networks (DNN) with noisy labels is challenging since DNN can easily memorize inaccurate labels, leading to poor generalization ability. Recently, the meta-learning based label correction strategy is widely adopted to tackle this problem via identifying and correcting potential noisy labels with the help of a small set of clean validation data. Although training with purified labels can effectively improve performance, solving the meta-learning problem inevitably involves a nested loop of bi-level optimization between model weights and hyperparameters (i.e., label distribution). As compromise, previous methods resort toa coupled learning process with alternating update. In this paper, we empirically find such simultaneous optimization over both model weights and label distribution can not achieve an optimal routine, consequently limiting the representation ability of backbone and accuracy of corrected labels. From this observation, a novel multi-stage label purifier named DMLP is proposed. DMLP decouples the label correction process into label-free representation learning and a simple meta label purifier, In this way, DMLP can focus on extracting discriminative feature and label correction in two distinctive stages. DMLP is a plug-and-play label purifier, the purified labels can be directly reused in naive end-to-end network retraining or other robust learning methods, where state-of-the-art results are obtained on several synthetic and real-world noisy datasets, especially under high noise levels.

Modern image classifiers perform well on populated classes while degrading considerably on tail classes with only a few instances. Humans, by contrast, effortlessly handle the long-tailed recognition challenge, since they can learn the tail representation based on different levels of semantic abstraction, making the learned tail features more discriminative. This phenomenon motivated us to propose SuperDisco, an algorithm that discovers super-class representations for long-tailed recognition using a graph model. We learn to construct the super-class graph to guide the representation learning to deal with long-tailed distributions. Through message passing on the super-class graph, image representations are rectified and refined by attending to the most relevant entities based on the semantic similarity among their super-classes. Moreover, we propose to meta-learn the super-class graph under the supervision of a prototype graph constructed from a small amount of imbalanced data. By doing so, we obtain a more robust super-class graph that further improves the long-tailed recognition performance. The consistent state-of-the-art experiments on the long-tailed CIFAR-100, ImageNet, Places, and iNaturalist demonstrate the benefit of the discovered super-class graph for dealing with long-tailed distributions.

International benchmarking competitions have become fundamental for the comparative performance assessment of image analysis methods. However, little attention has been given to investigating what can be learnt from these competitions. Do they really generate scientific progress? What are common and successful participation strategies? What makes a solution superior to a competing method? To address this gap in the literature, we performed a multi-center study with all 80 competitions that were conducted in the scope of IEEE ISBI 2021 and MICCAI 2021. Statistical analyses performed based on comprehensive descriptions of the submitted algorithms linked to their rank as well as the underlying participation strategies revealed common characteristics of winning solutions. These typically include the use of multi-task learning (63%) and/or multi-stage pipelines (61%), and a focus on augmentation (100%), image preprocessing (97%), data curation (79%), and postprocessing (66%). The “typical” lead of a winning team is a computer scientist with a doctoral degree, five years of experience in biomedical image analysis, and four years of experience in deep learning. Two core general development strategies stood out for highly-ranked teams: the reflection of the metrics in the method design and the focus on analyzing and handling failure cases. According to the organizers, 43% of the winning algorithms exceeded the state of the art but only 11% completely solved the respective domain problem. The insights of our study could help researchers (1) improve algorithm development strategies when approaching new problems, and (2) focus on open research questions revealed by this work.

Many real-world recognition problems are characterized by long-tailed label distributions. These distributions make representation learning highly challenging due to limited generalization over the tail classes. If the test distribution differs from the training distribution, e.g. uniform versus long-tailed, the problem of the distribution shift needs to be addressed. A recent line of work proposes learning multiple diverse experts to tackle this issue. Ensemble diversity is encouraged by various techniques, e.g. by specializing different experts in the head and the tail classes. In this work, we take an analytical approach and extend the notion of logit adjustment to ensembles to form a Balanced Product of Experts (BalPoE). BalPoE combines a family of experts with different test-time target distributions, generalizing several previous approaches. We show how to properly define these distributions and combine the experts in order to achieve unbiased predictions, by proving that the ensemble is Fisher-consistent for minimizing the balanced error. Our theoretical analysis shows that our balanced ensemble requires calibrated experts, which we achieve in practice using mixup. We conduct extensive experiments and our method obtains new state-of-the-art results on three long-tailed datasets: CIFAR-100-LT, ImageNet-LT, and iNaturalist-2018. Our code is available at https://github.com/emasa/BalPoE-CalibratedLT.

How to estimate the uncertainty of a given model is a crucial problem. Current calibration techniques treat different classes equally and thus implicitly assume that the distribution of training data is balanced, but ignore the fact that real-world data often follows a long-tailed distribution. In this paper, we explore the problem of calibrating the model trained from a long-tailed distribution. Due to the difference between the imbalanced training distribution and balanced test distribution, existing calibration methods such as temperature scaling can not generalize well to this problem. Specific calibration methods for domain adaptation are also not applicable because they rely on unlabeled target domain instances which are not available. Models trained from a long-tailed distribution tend to be more overconfident to head classes. To this end, we propose a novel knowledge-transferring-based calibration method by estimating the importance weights for samples of tail classes to realize long-tailed calibration. Our method models the distribution of each class as a Gaussian distribution and views the source statistics of head classes as a prior to calibrate the target distributions of tail classes. We adaptively transfer knowledge from head classes to get the target probability density of tail classes. The importance weight is estimated by the ratio of the target probability density over the source probability density. Extensive experiments on CIFAR-10-LT, MNIST-LT, CIFAR-100-LT, and ImageNet-LT datasets demonstrate the effectiveness of our method.

Although Domain Adaptation in Semantic Scene Segmentation has shown impressive improvement in recent years, the fairness concerns in the domain adaptation have yet to be well defined and addressed. In addition, fairness is one of the most critical aspects when deploying the segmentation models into human-related real-world applications, e.g., autonomous driving, as any unfair predictions could influence human safety. In this paper, we propose a novel Fairness Domain Adaptation (FREDOM) approach to semantic scene segmentation. In particular, from the proposed formulated fairness objective, a new adaptation framework will be introduced based on the fair treatment of class distributions. Moreover, to generally model the context of structural dependency, a new conditional structural constraint is introduced to impose the consistency of predicted segmentation. Thanks to the proposed Conditional Structure Network, the self-attention mechanism has sufficiently modeled the structural information of segmentation. Through the ablation studies, the proposed method has shown the performance improvement of the segmentation models and promoted fairness in the model predictions. The experimental results on the two standard benchmarks, i.e., SYNTHIA -> Cityscapes and GTA5 -> Cityscapes, have shown that our method achieved State-of-the-Art (SOTA) performance.

Unsupervised domain adaptation (UDA) aims to transfer the knowledge from a labeled source domain to an unlabeled target domain. Typically, to guarantee desirable knowledge transfer, aligning the distribution between source and target domain from a global perspective is widely adopted in UDA. Recent researchers further point out the importance of local-level alignment and propose to construct instance-pair alignment by leveraging on Optimal Transport (OT) theory. However, existing OT-based UDA approaches are limited to handling class imbalance challenges and introduce a heavy computation overhead when considering a large-scale training situation. To cope with two aforementioned issues, we propose a Clustering-based Optimal Transport (COT) algorithm, which formulates the alignment procedure as an Optimal Transport problem and constructs a mapping between clustering centers in the source and target domain via an end-to-end manner. With this alignment on clustering centers, our COT eliminates the negative effect caused by class imbalance and reduces the computation cost simultaneously. Empirically, our COT achieves state-of-the-art performance on several authoritative benchmark datasets.

Source free domain adaptation (SFDA) aims to transfer a trained source model to the unlabeled target domain without accessing the source data. However, the SFDA setting faces a performance bottleneck due to the absence of source data and target supervised information, as evidenced by the limited performance gains of the newest SFDA methods. Active source free domain adaptation (ASFDA) can break through the problem by exploring and exploiting a small set of informative samples via active learning. In this paper, we first find that those satisfying the properties of neighbor-chaotic, individual-different, and source-dissimilar are the best points to select. We define them as the minimum happy (MH) points challenging to explore with existing methods. We propose minimum happy points learning (MHPL) to explore and exploit MH points actively. We design three unique strategies: neighbor environment uncertainty, neighbor diversity relaxation, and one-shot querying, to explore the MH points. Further, to fully exploit MH points in the learning process, we design a neighbor focal loss that assigns the weighted neighbor purity to the cross entropy loss of MH points to make the model focus more on them. Extensive experiments verify that MHPL remarkably exceeds the various types of baselines and achieves significant performance gains at a small cost of labeling.

Deep neural networks (DNNs) often perform poorly in the presence of domain shift and category shift. How to upcycle DNNs and adapt them to the target task remains an important open problem. Unsupervised Domain Adaptation (UDA), especially recently proposed Source-free Domain Adaptation (SFDA), has become a promising technology to address this issue. Nevertheless, most existing SFDA methods require that the source domain and target domain share the same label space, consequently being only applicable to the vanilla closed-set setting. In this paper, we take one step further and explore the Source-free Universal Domain Adaptation (SF-UniDA). The goal is to identify “known” data samples under both domain and category shift, and reject those “unknown” data samples (not present in source classes), with only the knowledge from standard pre-trained source model. To this end, we introduce an innovative global and local clustering learning technique (GLC). Specifically, we design a novel, adaptive one-vs-all global clustering algorithm to achieve the distinction across different target classes and introduce a local k-NN clustering strategy to alleviate negative transfer. We examine the superiority of our GLC on multiple benchmarks with different category shift scenarios, including partial-set, open-set, and open-partial-set DA. More remarkably, in the most challenging open-partial-set DA scenario, GLC outperforms UMAD by 14.8% on the VisDA benchmark.

Multimodal learning (MML) aims to jointly exploit the common priors of different modalities to compensate for their inherent limitations. However, existing MML methods often optimize a uniform objective for different modalities, leading to the notorious “modality imbalance” problem and counterproductive MML performance. To address the problem, some existing methods modulate the learning pace based on the fused modality, which is dominated by the better modality and eventually results in a limited improvement on the worse modal. To better exploit the features of multimodal, we propose Prototypical Modality Rebalance (PMR) to perform stimulation on the particular slow-learning modality without interference from other modalities. Specifically, we introduce the prototypes that represent general features for each class, to build the non-parametric classifiers for uni-modal performance evaluation. Then, we try to accelerate the slow-learning modality by enhancing its clustering toward prototypes. Furthermore, to alleviate the suppression from the dominant modality, we introduce a prototype-based entropy regularization term during the early training stage to prevent premature convergence. Besides, our method only relies on the representations of each modality and without restrictions from model structures and fusion methods, making it with great application potential for various scenarios. The source code is available here.

Multimodal learning has shown great potentials in numerous scenes and attracts increasing interest recently. However, it often encounters the problem of missing modality data and thus suffers severe performance degradation in practice. To this end, we propose a general framework called MMANet to assist incomplete multimodal learning. It consists of three components: the deployment network used for inference, the teacher network transferring comprehensive multimodal information to the deployment network, and the regularization network guiding the deployment network to balance weak modality combinations. Specifically, we propose a novel margin-aware distillation (MAD) to assist the information transfer by weighing the sample contribution with the classification uncertainty. This encourages the deployment network to focus on the samples near decision boundaries and acquire the refined inter-class margin. Besides, we design a modality-aware regularization (MAR) algorithm to mine the weak modality combinations and guide the regularization network to calculate prediction loss for them. This forces the deployment network to improve its representation ability for the weak modality combinations adaptively. Finally, extensive experiments on multimodal classification and segmentation tasks demonstrate that our MMANet outperforms the state-of-the-art significantly.

Test time adaptation (TTA) aims to adapt deep neural networks when receiving out of distribution test domain samples. In this setting, the model can only access online unlabeled test samples and pre-trained models on the training domains. We first address TTA as a feature revision problem due to the domain gap between source domains and target domains. After that, we follow the two measurements alignment and uniformity to discuss the test time feature revision. For test time feature uniformity, we propose a test time self-distillation strategy to guarantee the consistency of uniformity between representations of the current batch and all the previous batches. For test time feature alignment, we propose a memorized spatial local clustering strategy to align the representations among the neighborhood samples for the upcoming batch. To deal with the common noisy label problem, we propound the entropy and consistency filters to select and drop the possible noisy labels. To prove the scalability and efficacy of our method, we conduct experiments on four domain generalization benchmarks and four medical image segmentation tasks with various backbones. Experiment results show that our method not only improves baseline stably but also outperforms existing state-of-the-art test time adaptation methods.

Prototypical Network is a popular few-shot solver that aims at establishing a feature metric generalizable to novel few-shot classification (FSC) tasks using deep neural networks. However, its performance drops dramatically when generalizing to the FSC tasks in new domains. In this study, we revisit this problem and argue that the devil lies in the simplicity bias pitfall in neural networks. In specific, the network tends to focus on some biased shortcut features (e.g., color, shape, etc.) that are exclusively sufficient to distinguish very few classes in the meta-training tasks within a pre-defined domain, but fail to generalize across domains as some desirable semantic features. To mitigate this problem, we propose a Local-global Distillation Prototypical Network (LDP-net). Different from the standard Prototypical Network, we establish a two-branch network to classify the query image and its random local crops, respectively. Then, knowledge distillation is conducted among these two branches to enforce their class affiliation consistency. The rationale behind is that since such global-local semantic relationship is expected to hold regardless of data domains, the local-global distillation is beneficial to exploit some cross-domain transferable semantic features for feature metric establishment. Moreover, such local-global semantic consistency is further enforced among different images of the same class to reduce the intra-class semantic variation of the resultant feature. In addition, we propose to update the local branch as Exponential Moving Average (EMA) over training episodes, which makes it possible to better distill cross-episode knowledge and further enhance the generalization performance. Experiments on eight cross-domain FSC benchmarks empirically clarify our argument and show the state-of-the-art results of LDP-net. Code is available in https://github.com/NWPUZhoufei/LDP-Net

Machine learning models have been found to learn shortcuts---unintended decision rules that are unable to generalize---undermining models’ reliability. Previous works address this problem under the tenuous assumption that only a single shortcut exists in the training data. Real-world images are rife with multiple visual cues from background to texture. Key to advancing the reliability of vision systems is understanding whether existing methods can overcome multiple shortcuts or struggle in a Whac-A-Mole game, i.e., where mitigating one shortcut amplifies reliance on others. To address this shortcoming, we propose two benchmarks: 1) UrbanCars, a dataset with precisely controlled spurious cues, and 2) ImageNet-W, an evaluation set based on ImageNet for watermark, a shortcut we discovered affects nearly every modern vision model. Along with texture and background, ImageNet-W allows us to study multiple shortcuts emerging from training on natural images. We find computer vision models, including large foundation models---regardless of training set, architecture, and supervision---struggle when multiple shortcuts are present. Even methods explicitly designed to combat shortcuts struggle in a Whac-A-Mole dilemma. To tackle this challenge, we propose Last Layer Ensemble, a simple-yet-effective method to mitigate multiple shortcuts without Whac-A-Mole behavior. Our results surface multi-shortcut mitigation as an overlooked challenge critical to advancing the reliability of vision systems. The datasets and code are released: https://github.com/facebookresearch/Whac-A-Mole.

In a multi-task learning (MTL) setting, a single model is trained to tackle a diverse set of tasks jointly. Despite rapid progress in the field, MTL remains challenging due to optimization issues such as conflicting and dominating gradients. In this work, we propose using a condition number of a linear system of gradients as a stability criterion of an MTL optimization. We theoretically demonstrate that a condition number reflects the aforementioned optimization issues. Accordingly, we present Aligned-MTL, a novel MTL optimization approach based on the proposed criterion, that eliminates instability in the training process by aligning the orthogonal components of the linear system of gradients. While many recent MTL approaches guarantee convergence to a minimum, task trade-offs cannot be specified in advance. In contrast, Aligned-MTL provably converges to an optimal point with pre-defined task-specific weights, which provides more control over the optimization result. Through experiments, we show that the proposed approach consistently improves performance on a diverse set of MTL benchmarks, including semantic and instance segmentation, depth estimation, surface normal estimation, and reinforcement learning.

In this work, we introduce MDL-NAS, a unified framework that integrates multiple vision tasks into a manageable supernet and optimizes these tasks collectively under diverse dataset domains. MDL-NAS is storage-efficient since multiple models with a majority of shared parameters can be deposited into a single one. Technically, MDL-NAS constructs a coarse-to-fine search space, where the coarse search space offers various optimal architectures for different tasks while the fine search space provides fine-grained parameter sharing to tackle the inherent obstacles of multi-domain learning. In the fine search space, we suggest two parameter sharing policies, i.e., sequential sharing policy and mask sharing policy. Compared with previous works, such two sharing policies allow for the partial sharing and non-sharing of parameters at each layer of the network, hence attaining real fine-grained parameter sharing. Finally, we present a joint-subnet search algorithm that finds the optimal architecture and sharing parameters for each task within total resource constraints, challenging the traditional practice that downstream vision tasks are typically equipped with backbone networks designed for image classification. Experimentally, we demonstrate that MDL-NAS families fitted with non-hierarchical or hierarchical transformers deliver competitive performance for all tasks compared with state-of-the-art methods while maintaining efficient storage deployment and computation. We also demonstrate that MDL-NAS allows incremental learning and evades catastrophic forgetting when generalizing to a new task.

Foundation models have shown outstanding performance and generalization capabilities across domains. Since most studies on foundation models mainly focus on the pretraining phase, a naive strategy to minimize a single task-specific loss is adopted for fine-tuning. However, such fine-tuning methods do not fully leverage other losses that are potentially beneficial for the target task. Therefore, we propose MEta Loss TRansformer (MELTR), a plug-in module that automatically and non-linearly combines various loss functions to aid learning the target task via auxiliary learning. We formulate the auxiliary learning as a bi-level optimization problem and present an efficient optimization algorithm based on Approximate Implicit Differentiation (AID). For evaluation, we apply our framework to various video foundation models (UniVL, Violet and All-in-one), and show significant performance gain on all four downstream tasks: text-to-video retrieval, video question answering, video captioning, and multi-modal sentiment analysis. Our qualitative analyses demonstrate that MELTR adequately ‘transforms’ individual loss functions and ‘melts’ them into an effective unified loss. Code is available at https://github.com/mlvlab/MELTR.

Fine-tuning large-scale pre-trained vision models to downstream tasks is a standard technique for achieving state-of-the-art performance on computer vision benchmarks. However, fine-tuning the whole model with millions of parameters is inefficient as it requires storing a same-sized new model copy for each task. In this work, we propose LoRand, a method for fine-tuning large-scale vision models with a better trade-off between task performance and the number of trainable parameters. LoRand generates tiny adapter structures with low-rank synthesis while keeping the original backbone parameters fixed, resulting in high parameter sharing. To demonstrate LoRand’s effectiveness, we implement extensive experiments on object detection, semantic segmentation, and instance segmentation tasks. By only training a small percentage (1% to 3%) of the pre-trained backbone parameters, LoRand achieves comparable performance to standard fine-tuning on COCO and ADE20K and outperforms fine-tuning in low-resource PASCAL VOC dataset.

Batch Normalization (BN) and its variants has been extensively studied for neural nets in various computer vision tasks, but relatively little work has been dedicated to studying the effect of BN in continual learning. To that end, we develop a new update patch for BN, particularly tailored for the exemplar-based class-incremental learning (CIL). The main issue of BN in CIL is the imbalance of training data between current and past tasks in a mini-batch, which makes the empirical mean and variance as well as the learnable affine transformation parameters of BN heavily biased toward the current task --- contributing to the forgetting of past tasks. While one of the recent BN variants has been developed for “online” CIL, in which the training is done with a single epoch, we show that their method does not necessarily bring gains for “offline” CIL, in which a model is trained with multiple epochs on the imbalanced training data. The main reason for the ineffectiveness of their method lies in not fully addressing the data imbalance issue, especially in computing the gradients for learning the affine transformation parameters of BN. Accordingly, our new hyperparameter-free variant, dubbed as Task-Balanced BN (TBBN), is proposed to more correctly resolve the imbalance issue by making a horizontally-concatenated task-balanced batch using both reshape and repeat operations during training. Based on our experiments on class incremental learning of CIFAR-100, ImageNet-100, and five dissimilar task datasets, we demonstrate that our TBBN, which works exactly the same as the vanilla BN in the inference time, is easily applicable to most existing exemplar-based offline CIL algorithms and consistently outperforms other BN variants.

In this paper, we study a novel task that enables partial knowledge transfer from pre-trained models, which we term as Partial Network Cloning (PNC). Unlike prior methods that update all or at least part of the parameters in the target network throughout the knowledge transfer process, PNC conducts partial parametric “cloning” from a source network and then injects the cloned module to the target, without modifying its parameters. Thanks to the transferred module, the target network is expected to gain additional functionality, such as inference on new classes; whenever needed, the cloned module can be readily removed from the target, with its original parameters and competence kept intact. Specifically, we introduce an innovative learning scheme that allows us to identify simultaneously the component to be cloned from the source and the position to be inserted within the target network, so as to ensure the optimal performance. Experimental results on several datasets demonstrate that, our method yields a significant improvement of 5% in accuracy and 50% in locality when compared with parameter-tuning based methods.

Machine unlearning can fortify the privacy and security of machine learning applications. Unfortunately, the exact unlearning approaches are inefficient, and the approximate unlearning approaches are unsuitable for complicated CNNs. Moreover, the approximate approaches have serious security flaws because even unlearning completely different data points can produce the same contribution estimation as unlearning the target data points. To address the above problems, we try to define machine unlearning from the knowledge perspective, and we propose a knowledge-level machine unlearning method, namely ERM-KTP. Specifically, we propose an entanglement-reduced mask (ERM) structure to reduce the knowledge entanglement among classes during the training phase. When receiving the unlearning requests, we transfer the knowledge of the non-target data points from the original model to the unlearned model and meanwhile prohibit the knowledge of the target data points via our proposed knowledge transfer and prohibition (KTP) method. Finally, we will get the unlearned model as the result and delete the original model to accomplish the unlearning process. Especially, our proposed ERM-KTP is an interpretable unlearning method because the ERM structure and the crafted masks in KTP can explicitly explain the operation and the effect of unlearning data points. Extensive experiments demonstrate the effectiveness, efficiency, high fidelity, and scalability of the ERM-KTP unlearning method.

With the continual expansion of face datasets, feature-based distillation prevails for large-scale face recognition. In this work, we attempt to remove identity supervision in student training, to spare the GPU memory from saving massive class centers. However, this naive removal leads to inferior distillation result. We carefully inspect the performance degradation from the perspective of intrinsic dimension, and argue that the gap in intrinsic dimension, namely the intrinsic gap, is intimately connected to the infamous capacity gap problem. By constraining the teacher’s search space with reverse distillation, we narrow the intrinsic gap and unleash the potential of feature-only distillation. Remarkably, the proposed reverse distillation creates universally student-friendly teacher that demonstrates outstanding student improvement. We further enhance its effectiveness by designing a student proxy to better bridge the intrinsic gap. As a result, the proposed method surpasses state-of-the-art distillation techniques with identity supervision on various face recognition benchmarks, and the improvements are consistent across different teacher-student pairs.

Continual learning aims to learn on non-stationary data streams without catastrophically forgetting previous knowledge. Prevalent replay-based methods address this challenge by rehearsing on a small buffer holding the seen data, for which a delicate sample selection strategy is required. However, existing selection schemes typically seek only to maximize the utility of the ongoing selection, overlooking the interference between successive rounds of selection. Motivated by this, we dissect the interaction of sequential selection steps within a framework built on influence functions. We manage to identify a new class of second-order influences that will gradually amplify incidental bias in the replay buffer and compromise the selection process. To regularize the second-order effects, a novel selection objective is proposed, which also has clear connections to two widely adopted criteria. Furthermore, we present an efficient implementation for optimizing the proposed criterion. Experiments on multiple continual learning benchmarks demonstrate the advantage of our approach over state-of-the-art methods. Code is available at https://github.com/feifeiobama/InfluenceCL.

Most existing online knowledge distillation(OKD) techniques typically require sophisticated modules to produce diverse knowledge for improving students’ generalization ability. In this paper, we strive to fully utilize multi-model settings instead of well-designed modules to achieve a distillation effect with excellent generalization performance. Generally, model generalization can be reflected in the flatness of the loss landscape. Since averaging parameters of multiple models can find flatter minima, we are inspired to extend the process to the sampled convex combinations of multi-student models in OKD. Specifically, by linearly weighting students’ parameters in each training batch, we construct a Hybrid-Weight Model(HWM) to represent the parameters surrounding involved students. The supervision loss of HWM can estimate the landscape’s curvature of the whole region around students to measure the generalization explicitly. Hence we integrate HWM’s loss into students’ training and propose a novel OKD framework via parameter hybridization(OKDPH) to promote flatter minima and obtain robust solutions. Considering the redundancy of parameters could lead to the collapse of HWM, we further introduce a fusion operation to keep the high similarity of students. Compared to the state-of-the-art(SOTA) OKD methods and SOTA methods of seeking flat minima, our OKDPH achieves higher performance with fewer parameters, benefiting OKD with lightweight and robust characteristics. Our code is publicly available at https://github.com/tianlizhang/OKDPH.

The dilemma between plasticity and stability arises as a common challenge for incremental learning. In contrast, the human memory system is able to remedy this dilemma owing to its multi-level memory structure, which motivates us to propose a Bilevel Memory system with Knowledge Projection (BMKP) for incremental learning. BMKP decouples the functions of learning and knowledge remembering via a bilevel-memory design: a working memory responsible for adaptively model learning, to ensure plasticity; a long-term memory in charge of enduringly storing the knowledge incorporated within the learned model, to guarantee stability. However, an emerging issue is how to extract the learned knowledge from the working memory and assimilate it into the long-term memory. To approach this issue, we reveal that the model learned by the working memory are actually residing in a redundant high-dimensional space, and the knowledge incorporated in the model can have a quite compact representation under a group of pattern basis shared by all incremental learning tasks. Therefore, we propose a knowledge projection process to adapatively maintain the shared basis, with which the loosely organized model knowledge of working memory is projected into the compact representation to be remembered in the long-term memory. We evaluate BMKP on CIFAR-10, CIFAR-100, and Tiny-ImageNet. The experimental results show that BMKP achieves state-of-the-art performance with lower memory usage.

A primary goal of class-incremental learning is to strike a balance between stability and plasticity, where models should be both stable enough to retain knowledge learned from previously seen classes, and plastic enough to learn concepts from new classes. While previous works demonstrate strong performance on class-incremental benchmarks, it is not clear whether their success comes from the models being stable, plastic, or a mixture of both. This paper aims to shed light on how effectively recent class-incremental learning algorithms address the stability-plasticity trade-off. We establish analytical tools that measure the stability and plasticity of feature representations, and employ such tools to investigate models trained with various algorithms on large-scale class-incremental benchmarks. Surprisingly, we find that the majority of class-incremental learning algorithms heavily favor stability over plasticity, to the extent that the feature extractor of a model trained on the initial set of classes is no less effective than that of the final incremental model. Our observations not only inspire two simple algorithms that highlight the importance of feature representation analysis, but also suggest that class-incremental learning approaches, in general, should strive for better feature representation learning.

Recently, a number of iterative learning methods have been introduced to improve generalization. These typically rely on training for longer periods of time in exchange for improved generalization. LLF (later-layer-forgetting) is a state-of-the-art method in this category. It strengthens learning in early layers by periodically re-initializing the last few layers of the network. Our principal innovation in this work is to use Simulated annealing in EArly Layers (SEAL) of the network in place of re-initialization of later layers. Essentially, later layers go through the normal gradient descent process, while the early layers go through short stints of gradient ascent followed by gradient descent. Extensive experiments on the popular Tiny-ImageNet dataset benchmark and a series of transfer learning and few-shot learning tasks show that we outperform LLF by a significant margin. We further show that, compared to normal training, LLF features, although improving on the target task, degrade the transfer learning performance across all datasets we explored. In comparison, our method outperforms LLF across the same target datasets by a large margin. We also show that the prediction depth of our method is significantly lower than that of LLF and normal training, indicating on average better prediction performance.

Deep reinforcement learning (DRL) gives the promise that an agent learns good policy from high-dimensional information, whereas representation learning removes irrelevant and redundant information and retains pertinent information. In this work, we demonstrate that the learned representation of the Q-network and its target Q-network should, in theory, satisfy a favorable distinguishable representation property. Specifically, there exists an upper bound on the representation similarity of the value functions of two adjacent time steps in a typical DRL setting. However, through illustrative experiments, we show that the learned DRL agent may violate this property and lead to a sub-optimal policy. Therefore, we propose a simple yet effective regularizer called Policy Evaluation with Easy Regularization on Representation (PEER), which aims to maintain the distinguishable representation property via explicit regularization on internal representations. And we provide the convergence rate guarantee of PEER. Implementing PEER requires only one line of code. Our experiments demonstrate that incorporating PEER into DRL can significantly improve performance and sample efficiency. Comprehensive experiments show that PEER achieves state-of-the-art performance on all 4 environments on PyBullet, 9 out of 12 tasks on DMControl, and 19 out of 26 games on Atari. To the best of our knowledge, PEER is the first work to study the inherent representation property of Q-network and its target. Our code is available at https://sites.google.com/view/peer-cvpr2023/.

Behavior of neural networks is irremediably determined by the specific loss and data used during training. However it is often desirable to tune the model at inference time based on external factors such as preferences of the user or dynamic characteristics of the data. This is especially important to balance the perception-distortion trade-off of ill-posed image-to-image translation tasks. In this work, we propose to optimize a parametric tunable convolutional layer, which includes a number of different kernels, using a parametric multi-loss, which includes an equal number of objectives. Our key insight is to use a shared set of parameters to dynamically interpolate both the objectives and the kernels. During training, these parameters are sampled at random to explicitly optimize all possible combinations of objectives and consequently disentangle their effect into the corresponding kernels. During inference, these parameters become interactive inputs of the model hence enabling reliable and consistent control over the model behavior. Extensive experimental results demonstrate that our tunable convolutions effectively work as a drop-in replacement for traditional convolutions in existing neural networks at virtually no extra computational cost, outperforming state-of-the-art control strategies in a wide range of applications; including image denoising, deblurring, super-resolution, and style transfer.

The recent emergence of new algorithms for permuting models into functionally equivalent regions of the solution space has shed some light on the complexity of error surfaces and some promising properties like mode connectivity. However, finding the permutation that minimizes some objectives is challenging, and current optimization techniques are not differentiable, which makes it difficult to integrate into a gradient-based optimization, and often leads to sub-optimal solutions. In this paper, we propose a Sinkhorn re-basin network with the ability to obtain the transportation plan that better suits a given objective. Unlike the current state-of-art, our method is differentiable and, therefore, easy to adapt to any task within the deep learning domain. Furthermore, we show the advantage of our re-basin method by proposing a new cost function that allows performing incremental learning by exploiting the linear mode connectivity property. The benefit of our method is compared against similar approaches from the literature under several conditions for both optimal transport and linear mode connectivity. The effectiveness of our continual learning method based on re-basin is also shown for several common benchmark datasets, providing experimental results that are competitive with the state-of-art. The source code is provided at https://github.com/fagp/sinkhorn-rebasin.

Recently, flat minima are proven to be effective for improving generalization and sharpness-aware minimization (SAM) achieves state-of-the-art performance. Yet the current definition of flatness discussed in SAM and its follow-ups are limited to the zeroth-order flatness (i.e., the worst-case loss within a perturbation radius). We show that the zeroth-order flatness can be insufficient to discriminate minima with low generalization error from those with high generalization error both when there is a single minimum or multiple minima within the given perturbation radius. Thus we present first-order flatness, a stronger measure of flatness focusing on the maximal gradient norm within a perturbation radius which bounds both the maximal eigenvalue of Hessian at local minima and the regularization function of SAM. We also present a novel training procedure named Gradient norm Aware Minimization (GAM) to seek minima with uniformly small curvature across all directions. Experimental results show that GAM improves the generalization of models trained with current optimizers such as SGD and AdamW on various datasets and networks. Furthermore, we show that GAM can help SAM find flatter minima and achieve better generalization.

Network structure learning aims to optimize network architectures and make them more efficient without compromising performance. In this paper, we first study the astrocytes, a new mechanism to regulate connections in the classic M-P neuron. Then, with the astrocytes, we propose an AstroNet that can adaptively optimize neuron connections and therefore achieves structure learning to achieve higher accuracy and efficiency. AstroNet is based on our built Astrocyte-Neuron model, with a temporal regulation mechanism and a global connection mechanism, which is inspired by the bidirectional communication property of astrocytes. With the model, the proposed AstroNet uses a neural network (NN) for performing tasks, and an astrocyte network (AN) to continuously optimize the connections of NN, i.e., assigning weight to the neuron units in the NN adaptively. Experiments on the classification task demonstrate that our AstroNet can efficiently optimize the network structure while achieving state-of-the-art (SOTA) accuracy.

Recently, the sizes of deep neural networks and training datasets both increase drastically to pursue better performance in a practical sense. With the prevalence of transformer-based models in vision tasks, even more pressure is laid on the GPU platforms to train these heavy models, which consumes a large amount of time and computing resources as well. Therefore, it’s crucial to accelerate the training process of deep neural networks. In this paper, we propose a general network expansion method to reduce the practical time cost of the model training process. Specifically, we utilize both width- and depth-level sparsity of dense models to accelerate the training of deep neural networks. Firstly, we pick a sparse sub-network from the original dense model by reducing the number of parameters as the starting point of training. Then the sparse architecture will gradually expand during the training procedure and finally grow into a dense one. We design different expanding strategies to grow CNNs and ViTs respectively, due to the great heterogeneity in between the two architectures. Our method can be easily integrated into popular deep learning frameworks, which saves considerable training time and hardware resources. Extensive experiments show that our acceleration method can significantly speed up the training process of modern vision models on general GPU devices with negligible performance drop (e.g. 1.42x faster for ResNet-101 and 1.34x faster for DeiT-base on ImageNet-1k). The code is available at https://github.com/huawei-noah/Efficient-Computing/tree/master/TrainingAcceleration/NetworkExpansion and https://gitee.com/mindspore/hub/blob/master/mshubres/assets/noah-cvlab/gpu/1.8/networkexpansionv1.0_imagenet2012.md.

This paper aims to illustrate the concept-emerging phenomenon in a trained DNN. Specifically, we find that the inference score of a DNN can be disentangled into the effects of a few interactive concepts. These concepts can be understood as inference patterns in a sparse, symbolic graphical model, which explains the DNN. The faithfulness of using such a graphical model to explain the DNN is theoretically guaranteed, because we prove that the graphical model can well mimic the DNN’s outputs on an exponential number of different masked samples. Besides, such a graphical model can be further simplified and re-written as an And-Or graph (AOG), without losing much explanation accuracy. The code is released at https://github.com/sjtu-xai-lab/aog.

In this paper, we study the second order properties of the loss of trained deep neural networks with respect to the training data points to understand the curvature of the loss surface in the vicinity of these points. We find that there is an unexpected concentration of samples with very low curvature. We note that these low curvature samples are largely consistent across completely different architectures, and identifiable in the early epochs of training. We show that the curvature relates to the ‘cleanliness’ of the data points, with low curvatures samples corresponding to clean, higher clarity samples, representative of their category. Alternatively, high curvature samples are often occluded, have conflicting features and visually atypical of their category. Armed with this insight, we introduce SLo-Curves, a novel coreset identification and training algorithm. SLo-curves identifies the samples with low curvatures as being more data-efficient and trains on them with an additional regularizer that penalizes high curvature of the loss surface in their vicinity. We demonstrate the efficacy of SLo-Curves on CIFAR-10 and CIFAR-100 datasets, where it outperforms state of the art coreset selection methods at small coreset sizes by up to 9%. The identified coresets generalize across architectures, and hence can be pre-computed to generate condensed versions of datasets for use in downstream tasks.

Deep learning models are challenged by the distribution shift between the training data and test data. Recently, the large models pre-trained on diverse data have demonstrated unprecedented robustness to various distribution shifts. However, fine-tuning these models can lead to a trade-off between in-distribution (ID) performance and out-of-distribution (OOD) robustness. Existing methods for tackling this trade-off do not explicitly address the OOD robustness problem. In this paper, based on causal analysis of the aforementioned problems, we propose a novel fine-tuning method, which uses masked images as counterfactual samples that help improve the robustness of the fine-tuning model. Specifically, we mask either the semantics-related or semantics-unrelated patches of the images based on class activation map to break the spurious correlation, and refill the masked patches with patches from other images. The resulting counterfactual samples are used in feature-based distillation with the pre-trained model. Extensive experiments verify that regularizing the fine-tuning with the proposed masked images can achieve a better trade-off between ID and OOD performance, surpassing previous methods on the OOD performance. Our code is available at https://github.com/Coxy7/robust-finetuning.

Prior work has shown that Visual Recognition datasets frequently underrepresent bias groups B (e.g. Female) within class labels Y (e.g. Programmers). This dataset bias can lead to models that learn spurious correlations between class labels and bias groups such as age, gender, or race. Most recent methods that address this problem require significant architectural changes or additional loss functions requiring more hyper-parameter tuning. Alternatively, data sampling baselines from the class imbalance literature (eg Undersampling, Upweighting), which can often be implemented in a single line of code and often have no hyperparameters, offer a cheaper and more efficient solution. However, these methods suffer from significant shortcomings. For example, Undersampling drops a significant part of the input distribution per epoch while Oversampling repeats samples, causing overfitting. To address these shortcomings, we introduce a new class-conditioned sampling method: Bias Mimicking. The method is based on the observation that if a class c bias distribution, i.e., P_D(B|Y=c) is mimicked across every c’ != c, then Y and B are statistically independent. Using this notion, BM, through a novel training procedure, ensures that the model is exposed to the entire distribution per epoch without repeating samples. Consequently, Bias Mimicking improves underrepresented groups’ accuracy of sampling methods by 3% over four benchmarks while maintaining and sometimes improving performance over nonsampling methods. Code: https://github.com/mqraitem/Bias-Mimicking

The complicated architecture and high training cost of vision transformers urge the exploration of post-training quantization. However, the heavy-tailed distribution of vision transformer activations hinders the effectiveness of previous post-training quantization methods, even with advanced quantizer designs. Instead of tuning the quantizer to better fit the complicated activation distribution, this paper proposes NoisyQuant, a quantizer-agnostic enhancement for the post-training activation quantization performance of vision transformers. We make a surprising theoretical discovery that for a given quantizer, adding a fixed Uniform noisy bias to the values being quantized can significantly reduce the quantization error under provable conditions. Building on the theoretical insight, NoisyQuant achieves the first success on actively altering the heavy-tailed activation distribution with additive noisy bias to fit a given quantizer. Extensive experiments show NoisyQuant largely improves the post-training quantization performance of vision transformer with minimal computation overhead. For instance, on linear uniform 6-bit activation quantization, NoisyQuant improves SOTA top-1 accuracy on ImageNet by up to 1.7%, 1.1% and 0.5% for ViT, DeiT, and Swin Transformer respectively, achieving on-par or even higher performance than previous nonlinear, mixed-precision quantization.

Due to data privacy issues, accelerating networks with tiny training sets has become a critical need in practice. Previous methods mainly adopt filter-level pruning to accelerate networks with scarce training samples. In this paper, we reveal that dropping blocks is a fundamentally superior approach in this scenario. It enjoys a higher acceleration ratio and results in a better latency-accuracy performance under the few-shot setting. To choose which blocks to drop, we propose a new concept namely recoverability to measure the difficulty of recovering the compressed network. Our recoverability is efficient and effective for choosing which blocks to drop. Finally, we propose an algorithm named PRACTISE to accelerate networks using only tiny sets of training images. PRACTISE outperforms previous methods by a significant margin. For 22% latency reduction, PRACTISE surpasses previous methods by on average 7% on ImageNet-1k. It also enjoys high generalization ability, working well under data-free or out-of-domain data settings, too. Our code is at https://github.com/DoctorKey/Practise.

Most recent test-time adaptation methods focus on only classification tasks, use specialized network architectures, destroy model calibration or rely on lightweight information from the source domain. To tackle these issues, this paper proposes a novel Test-time Self-Learning method with automatic Adversarial augmentation dubbed TeSLA for adapting a pre-trained source model to the unlabeled streaming test data. In contrast to conventional self-learning methods based on cross-entropy, we introduce a new test-time loss function through an implicitly tight connection with the mutual information and online knowledge distillation. Furthermore, we propose a learnable efficient adversarial augmentation module that further enhances online knowledge distillation by simulating high entropy augmented images. Our method achieves state-of-the-art classification and segmentation results on several benchmarks and types of domain shifts, particularly on challenging measurement shifts of medical images. TeSLA also benefits from several desirable properties compared to competing methods in terms of calibration, uncertainty metrics, insensitivity to model architectures, and source training strategies, all supported by extensive ablations. Our code and models are available at https://github.com/devavratTomar/TeSLA.

Despite excellent performance in image generation, Generative Adversarial Networks (GANs) are notorious for its requirements of enormous storage and intensive computation. As an awesome “performance maker”, knowledge distillation is demonstrated to be particularly efficacious in exploring low-priced GANs. In this paper, we investigate the irreplaceability of teacher discriminator and present an inventive discriminator-cooperated distillation, abbreviated as DCD, towards refining better feature maps from the generator. In contrast to conventional pixel-to-pixel match methods in feature map distillation, our DCD utilizes teacher discriminator as a transformation to drive intermediate results of the student generator to be perceptually close to corresponding outputs of the teacher generator. Furthermore, in order to mitigate mode collapse in GAN compression, we construct a collaborative adversarial training paradigm where the teacher discriminator is from scratch established to co-train with student generator in company with our DCD. Our DCD shows superior results compared with existing GAN compression methods. For instance, after reducing over 40× MACs and 80× parameters of CycleGAN, we well decrease FID metric from 61.53 to 48.24 while the current SoTA method merely has 51.92. This work’s source code has been made accessible at https://github.com/poopit/DCD-official.

Privacy-preserving image generation has been important for segments such as medical domains that have sensitive and limited data. The benefits of guaranteed privacy come at the costs of generated images’ quality and utility due to the privacy budget constraints. The utility is currently measured by the gen2real accuracy (g2r%), i.e., the accuracy on real data of a downstream classifier trained using generated data. However, apart from this standard utility, we identify the “reversed utility” as another crucial aspect, which computes the accuracy on generated data of a classifier trained using real data, dubbed as real2gen accuracy (r2g%). Jointly considering these two views of utility, the standard and the reversed, could help the generation model better improve transferability between fake and real data. Therefore, we propose a novel private image generation method that incorporates a dual-purpose auxiliary classifier, which alternates between learning from real data and fake data, into the training of differentially private GANs. Additionally, our deliberate training strategies such as sequential training contributes to accelerating the generator’s convergence and further boosting the performance upon exhausting the privacy budget. Our results achieve new state-of-the-arts over all metrics on three benchmarks: MNIST, Fashion-MNIST, and CelebA.

Recent studies have shown that higher accuracy on ImageNet usually leads to better robustness against different corruptions. In this paper, instead of following the traditional research paradigm that investigates new out-of-distribution corruptions or perturbations deep models may encounter, we conduct model debugging in in-distribution data to explore which object attributes a model may be sensitive to. To achieve this goal, we create a toolkit for object editing with controls of backgrounds, sizes, positions, and directions, and create a rigorous benchmark named ImageNet-E(diting) for evaluating the image classifier robustness in terms of object attributes. With our ImageNet-E, we evaluate the performance of current deep learning models, including both convolutional neural networks and vision transformers. We find that most models are quite sensitive to attribute changes. An imperceptible change in the background can lead to an average of 9.23% drop on top-1 accuracy. We also evaluate some robust models including both adversarially trained models and other robust trained models and find that some models show worse robustness against attribute changes than vanilla models. Based on these findings, we discover ways to enhance attribute robustness with preprocessing, architecture designs, and training strategies. We hope this work can provide some insights to the community and open up a new avenue for research in robust computer vision. The code and dataset will be publicly available.

Position Embeddings (PEs), an arguably indispensable component in Vision Transformers (ViTs), have been shown to improve the performance of ViTs on many vision tasks. However, PEs have a potentially high risk of privacy leakage since the spatial information of the input patches is exposed. This caveat naturally raises a series of interesting questions about the impact of PEs on accuracy, privacy, prediction consistency, etc. To tackle these issues, we propose a Masked Jigsaw Puzzle (MJP) position embedding method. In particular, MJP first shuffles the selected patches via our block-wise random jigsaw puzzle shuffle algorithm, and their corresponding PEs are occluded. Meanwhile, for the non-occluded patches, the PEs remain the original ones but their spatial relation is strengthened via our dense absolute localization regressor. The experimental results reveal that 1) PEs explicitly encode the 2D spatial relationship and lead to severe privacy leakage problems under gradient inversion attack; 2) Training ViTs with the naively shuffled patches can alleviate the problem, but it harms the accuracy; 3) Under a certain shuffle ratio, the proposed MJP not only boosts the performance and robustness on large-scale datasets (i.e., ImageNet-1K and ImageNet-C, -A/O) but also improves the privacy preservation ability under typical gradient attacks by a large margin. The source code and trained models are available at https://github.com/yhlleo/MJP.

Semi-supervised video anomaly detection (VAD) is a critical task in the intelligent surveillance system. However, an essential type of anomaly in VAD named scene-dependent anomaly has not received the attention of researchers. Moreover, there is no research investigating anomaly anticipation, a more significant task for preventing the occurrence of anomalous events. To this end, we propose a new comprehensive dataset, NWPU Campus, containing 43 scenes, 28 classes of abnormal events, and 16 hours of videos. At present, it is the largest semi-supervised VAD dataset with the largest number of scenes and classes of anomalies, the longest duration, and the only one considering the scene-dependent anomaly. Meanwhile, it is also the first dataset proposed for video anomaly anticipation. We further propose a novel model capable of detecting and anticipating anomalous events simultaneously. Compared with 7 outstanding VAD algorithms in recent years, our method can cope with scene-dependent anomaly detection and anomaly anticipation both well, achieving state-of-the-art performance on ShanghaiTech, CUHK Avenue, IITB Corridor and the newly proposed NWPU Campus datasets consistently. Our dataset and code is available at: https://campusvad.github.io.

We propose a simple and application-friendly network (called SimpleNet) for detecting and localizing anomalies. SimpleNet consists of four components: (1) a pre-trained Feature Extractor that generates local features, (2) a shallow Feature Adapter that transfers local features towards target domain, (3) a simple Anomaly Feature Generator that counterfeits anomaly features by adding Gaussian noise to normal features, and (4) a binary Anomaly Discriminator that distinguishes anomaly features from normal features. During inference, the Anomaly Feature Generator would be discarded. Our approach is based on three intuitions. First, transforming pre-trained features to target-oriented features helps avoid domain bias. Second, generating synthetic anomalies in feature space is more effective, as defects may not have much commonality in the image space. Third, a simple discriminator is much efficient and practical. In spite of simplicity, SimpleNet outperforms previous methods quantitatively and qualitatively. On the MVTec AD benchmark, SimpleNet achieves an anomaly detection AUROC of 99.6%, reducing the error by 55.5% compared to the next best performing model. Furthermore, SimpleNet is faster than existing methods, with a high frame rate of 77 FPS on a 3080ti GPU. Additionally, SimpleNet demonstrates significant improvements in performance on the One-Class Novelty Detection task. Code: https://github.com/DonaldRR/SimpleNet.

Federated Distillation (FD) has recently attracted increasing attention for its efficiency in aggregating multiple diverse local models trained from statistically heterogeneous data of distributed clients. Existing FD methods generally treat these models equally by merely computing the average of their output soft predictions for some given input distillation sample, which does not take the diversity across all local models into account, thus leading to degraded performance of the aggregated model, especially when some local models learn little knowledge about the sample. In this paper, we propose a new perspective that treats the local data in each client as a specific domain and design a novel domain knowledge aware federated distillation method, dubbed DaFKD, that can discern the importance of each model to the distillation sample, and thus is able to optimize the ensemble of soft predictions from diverse models. Specifically, we employ a domain discriminator for each client, which is trained to identify the correlation factor between the sample and the corresponding domain. Then, to facilitate the training of the domain discriminator while saving communication costs, we propose sharing its partial parameters with the classification model. Extensive experiments on various datasets and settings show that the proposed method can improve the model accuracy by up to 6.02% compared to state-of-the-art baselines.

Federated learning can coordinate multiple users to participate in data training while ensuring data privacy. The collaboration of multiple agents allows for a natural connection between federated learning and collective intelligence. When there are large differences in data distribution among clients, it is crucial for federated learning to design a reliable client selection strategy and an interpretable client communication framework to better utilize group knowledge. Herein, a reliable personalized federated learning approach, termed RIPFL, is proposed and fully interpreted from the perspective of social learning. RIPFL reliably selects and divides the clients involved in training such that each client can use different amounts of social information and more effectively communicate with other clients. Simultaneously, the method effectively integrates personal information with the social information generated by the global model from the perspective of Bayesian decision rules and evidence theory, enabling individuals to grow better with the help of collective wisdom. An interpretable federated learning mind is well scalable, and the experimental results indicate that the proposed method has superior robustness and accuracy than other state-of-the-art federated learning algorithms.

Owing to the non-i.i.d. nature of client data, channel neurons in federated-learned models may specialize to distinct features for different clients. Yet, existing channel-sparse federated learning (FL) algorithms prescribe fixed sparsity strategies for client models, and may thus prevent clients from training channel neurons collaboratively. To minimize the impact of sparsity on FL convergence, we propose Flado to improve the alignment of client model update trajectories by tailoring the sparsities of individual neurons in each client. Empirical results show that while other sparse methods are surprisingly impactful to convergence, Flado can not only attain the highest task accuracies with unlimited budget across a range of datasets, but also significantly reduce the amount of FLOPs required for training more than by 10x under the same communications budget, and push the Pareto frontier of communication/computation trade-off notably further than competing FL algorithms.

Learning models trained on biased datasets tend to observe correlations between categorical and undesirable features, which result in degraded performances. Most existing debiased learning models are designed for centralized machine learning, which cannot be directly applied to distributed settings like federated learning (FL), which collects data at distinct clients with privacy preserved. To tackle the challenging task of debiased federated learning, we present a novel FL framework of Bias-Eliminating Augmentation Learning (FedBEAL), which learns to deploy Bias-Eliminating Augmenters (BEA) for producing client-specific bias-conflicting samples at each client. Since the bias types or attributes are not known in advance, a unique learning strategy is presented to jointly train BEA with the proposed FL framework. Extensive image classification experiments on datasets with various bias types confirm the effectiveness and applicability of our FedBEAL, which performs favorably against state-of-the-art debiasing and FL methods for debiased FL.

Face anti-spoofing (FAS) based on domain generalization (DG) has been recently studied to improve the generalization on unseen scenarios. Previous methods typically rely on domain labels to align the distribution of each domain for learning domain-invariant representations. However, artificial domain labels are coarse-grained and subjective, which cannot reflect real domain distributions accurately. Besides, such domain-aware methods focus on domain-level alignment, which is not fine-grained enough to ensure that learned representations are insensitive to domain styles. To address these issues, we propose a novel perspective for DG FAS that aligns features on the instance level without the need for domain labels. Specifically, Instance-Aware Domain Generalization framework is proposed to learn the generalizable feature by weakening the features’ sensitivity to instance-specific styles. Concretely, we propose Asymmetric Instance Adaptive Whitening to adaptively eliminate the style-sensitive feature correlation, boosting the generalization. Moreover, Dynamic Kernel Generator and Categorical Style Assembly are proposed to first extract the instance-specific features and then generate the style-diversified features with large style shifts, respectively, further facilitating the learning of style-insensitive features. Extensive experiments and analysis demonstrate the superiority of our method over state-of-the-art competitors. Code will be publicly available at this link: https://github.com/qianyuzqy/IADG.

This paper considers the synthetic-to-real adaptation of point cloud semantic segmentation, which aims to segment the real-world point clouds with only synthetic labels available. Contrary to synthetic data which is integral and clean, point clouds collected by real-world sensors typically contain unexpected and irregular noise because the sensors may be impacted by various environmental conditions. Consequently, the model trained on ideal synthetic data may fail to achieve satisfactory segmentation results on real data. Influenced by such noise, previous adversarial training methods, which are conventional for 2D adaptation tasks, become less effective. In this paper, we aim to mitigate the domain gap caused by target noise via learning to mask the source points during the adaptation procedure. To this end, we design a novel learnable masking module, which takes source features and 3D coordinates as inputs. We incorporate Gumbel-Softmax operation into the masking module so that it can generate binary masks and be trained end-to-end via gradient back-propagation. With the help of adversarial training, the masking module can learn to generate source masks to mimic the pattern of irregular target noise, thereby narrowing the domain gap. We name our method “Adversarial Masking” as adversarial training and learnable masking module depend on each other and cooperate with each other to mitigate the domain gap. Experiments on two synthetic-to-real adaptation benchmarks verify the effectiveness of the proposed method.

As the scientific and technological achievements produced by human intellectual labor and computation cost, model intellectual property (IP) protection, which refers to preventing the usage of the well-trained model on an unauthorized domain, deserves further attention, so as to effectively mobilize the enthusiasm of model owners and creators. To this end, we propose a novel compact un-transferable isolation domain (CUTI-domain), which acts as a model barrier to block illegal transferring from the authorized domain to the unauthorized domain. Specifically, CUTI-domain is investigated to block cross-domain transferring by highlighting private style features of the authorized domain and lead to the failure of recognition on unauthorized domains that contain irrelative private style features. Furthermore, depending on whether the unauthorized domain is known or not, two solutions of using CUTI-domain are provided: target-specified CUTI-domain and target-free CUTI-domain. Comprehensive experimental results on four digit datasets, CIFAR10 & STL10, and VisDA-2017 dataset, demonstrate that our CUTI-domain can be easily implemented with different backbones as a plug-and-play module and provides an efficient solution for model IP protection.

We develop a novel method to remove injected backdoors in deep learning models. It works by cloning the benign behaviors of a trojaned model to a new model of the same structure. It trains the clone model from scratch on a very small subset of samples and aims to minimize a cloning loss that denotes the differences between the activations of important neurons across the two models. The set of important neurons varies for each input, depending on their magnitude of activations and their impact on the classification result. We theoretically show our method can better recover benign functions of the backdoor model. Meanwhile, we prove our method can be more effective in removing backdoors compared with fine-tuning. Our experiments show that our technique can effectively remove nine different types of backdoors with minor benign accuracy degradation, outperforming the state-of-the-art backdoor removal techniques that are based on fine-tuning, knowledge distillation, and neuron pruning.

Deep neural networks (DNNs) are known to be vulnerable to both backdoor attacks as well as adversarial attacks. In the literature, these two types of attacks are commonly treated as distinct problems and solved separately, since they belong to training-time and inference-time attacks respectively. However, in this paper we find an intriguing connection between them: for a model planted with backdoors, we observe that its adversarial examples have similar behaviors as its triggered samples, i.e., both activate the same subset of DNN neurons. It indicates that planting a backdoor into a model will significantly affect the model’s adversarial examples. Based on this observations, a novel Progressive Backdoor Erasing (PBE) algorithm is proposed to progressively purify the infected model by leveraging untargeted adversarial attacks. Different from previous backdoor defense methods, one significant advantage of our approach is that it can erase backdoor even when the additional clean dataset is unavailable. We empirically show that, against 5 state-of-the-art backdoor attacks, our AFT can effectively erase the backdoor triggers without obvious performance degradation on clean samples and significantly outperforms existing defense methods.

Model inversion attacks are a type of privacy attack that reconstructs private data used to train a machine learning model, solely by accessing the model. Recently, white-box model inversion attacks leveraging Generative Adversarial Networks (GANs) to distill knowledge from public datasets have been receiving great attention because of their excellent attack performance. On the other hand, current black-box model inversion attacks that utilize GANs suffer from issues such as being unable to guarantee the completion of the attack process within a predetermined number of query accesses or achieve the same level of performance as white-box attacks. To overcome these limitations, we propose a reinforcement learning-based black-box model inversion attack. We formulate the latent space search as a Markov Decision Process (MDP) problem and solve it with reinforcement learning. Our method utilizes the confidence scores of the generated images to provide rewards to an agent. Finally, the private data can be reconstructed using the latent vectors found by the agent trained in the MDP. The experiment results on various datasets and models demonstrate that our attack successfully recovers the private information of the target model by achieving state-of-the-art attack performance. We emphasize the importance of studies on privacy-preserving machine learning by proposing a more advanced black-box model inversion attack.

Compared to query-based black-box attacks, transfer-based black-box attacks do not require any information of the attacked models, which ensures their secrecy. However, most existing transfer-based approaches rely on ensembling multiple models to boost the attack transferability, which is time- and resource-intensive, not to mention the difficulty of obtaining diverse models on the same task. To address this limitation, in this work, we focus on the single-model transfer-based black-box attack on object detection, utilizing only one model to achieve a high-transferability adversarial attack on multiple black-box detectors. Specifically, we first make observations on the patch optimization process of the existing method and propose an enhanced attack framework by slightly adjusting its training strategies. Then, we analogize patch optimization with regular model optimization, proposing a series of self-ensemble approaches on the input data, the attacked model, and the adversarial patch to efficiently make use of the limited information and prevent the patch from overfitting. The experimental results show that the proposed framework can be applied with multiple classical base attack methods (e.g., PGD and MIM) to greatly improve the black-box transferability of the well-optimized patch on multiple mainstream detectors, meanwhile boosting white-box performance.

Classification has been the focal point of research on adversarial attacks, but only a few works investigate methods suited to denser prediction tasks, such as semantic segmentation. The methods proposed in these works do not accurately solve the adversarial segmentation problem and, therefore, overestimate the size of the perturbations required to fool models. Here, we propose a white-box attack for these models based on a proximal splitting to produce adversarial perturbations with much smaller l_infinity norms. Our attack can handle large numbers of constraints within a nonconvex minimization framework via an Augmented Lagrangian approach, coupled with adaptive constraint scaling and masking strategies. We demonstrate that our attack significantly outperforms previously proposed ones, as well as classification attacks that we adapted for segmentation, providing a first comprehensive benchmark for this dense task.

Transferability of adversarial examples is critical for black-box deep learning model attacks. While most existing studies focus on enhancing the transferability of untargeted adversarial attacks, few of them studied how to generate transferable targeted adversarial examples that can mislead models into predicting a specific class. Moreover, existing transferable targeted adversarial attacks usually fail to sufficiently characterize the target class distribution, thus suffering from limited transferability. In this paper, we propose the Transferable Targeted Adversarial Attack (TTAA), which can capture the distribution information of the target class from both label-wise and feature-wise perspectives, to generate highly transferable targeted adversarial examples. To this end, we design a generative adversarial training framework consisting of a generator to produce targeted adversarial examples, and feature-label dual discriminators to distinguish the generated adversarial examples from the target class images. Specifically, we design the label discriminator to guide the adversarial examples to learn label-related distribution information about the target class. Meanwhile, we design a feature discriminator, which extracts the feature-wise information with strong cross-model consistency, to enable the adversarial examples to learn the transferable distribution information. Furthermore, we introduce the random perturbation dropping to further enhance the transferability by augmenting the diversity of adversarial examples used in the training process. Experiments demonstrate that our method achieves excellent performance on the transferability of targeted adversarial examples. The targeted fooling rate reaches 95.13% when transferred from VGG-19 to DenseNet-121, which significantly outperforms the state-of-the-art methods.

The deep neural networks (DNNs) trained by adversarial training (AT) usually suffered from significant robust generalization gap, i.e., DNNs achieve high training robustness but low test robustness. In this paper, we propose a generic method to boost the robust generalization of AT methods from the novel perspective of attribution span. To this end, compared with standard DNNs, we discover that the generalization gap of adversarially trained DNNs is caused by the smaller attribution span on the input image. In other words, adversarially trained DNNs tend to focus on specific visual concepts on training images, causing its limitation on test robustness. In this way, to enhance the robustness, we propose an effective method to enlarge the learned attribution span. Besides, we use hybrid feature statistics for feature fusion to enrich the diversity of features. Extensive experiments show that our method can effectively improves robustness of adversarially trained DNNs, outperforming previous SOTA methods. Furthermore, we provide a theoretical analysis of our method to prove its effectiveness.

Deep neural networks obtained by standard training have been constantly plagued by adversarial examples. Although adversarial training demonstrates its capability to defend against adversarial examples, unfortunately, it leads to an inevitable drop in the natural generalization. To address the issue, we decouple the natural generalization and the robust generalization from joint training and formulate different training strategies for each one. Specifically, instead of minimizing a global loss on the expectation over these two generalization errors, we propose a bi-expert framework called Generalist where we simultaneously train base learners with task-aware strategies so that they can specialize in their own fields. The parameters of base learners are collected and combined to form a global learner at intervals during the training process. The global learner is then distributed to the base learners as initialized parameters for continued training. Theoretically, we prove that the risks of Generalist will get lower once the base learners are well trained. Extensive experiments verify the applicability of Generalist to achieve high accuracy on natural examples while maintaining considerable robustness to adversarial ones. Code is available at https://github.com/PKU-ML/Generalist.

Despite incredible advances, deep learning has been shown to be susceptible to adversarial attacks. Numerous approaches were proposed to train robust networks both empirically and certifiably. However, most of them defend against only a single type of attack, while recent work steps forward at defending against multiple attacks. In this paper, to understand multi-target robustness, we view this problem as a bargaining game in which different players (adversaries) negotiate to reach an agreement on a joint direction of parameter updating. We identify a phenomenon named player domination in the bargaining game, and show that with this phenomenon, some of the existing max-based approaches such as MAX and MSD do not converge. Based on our theoretical results, we design a novel framework that adjusts the budgets of different adversaries to avoid player domination. Experiments on two benchmarks show that employing the proposed framework to the existing approaches significantly advances multi-target robustness.

Classical adversarial attacks for Face Recognition (FR) models typically generate discrete examples for target identity with a single state image. However, such paradigm of point-wise attack exhibits poor generalization against numerous unknown states of identity and can be easily defended. In this paper, by rethinking the inherent relationship between the face of target identity and its variants, we introduce a new pipeline of Generalized Manifold Adversarial Attack (GMAA) to achieve a better attack performance by expanding the attack range. Specifically, this expansion lies on two aspects -- GMAA not only expands the target to be attacked from one to many to encourage a good generalization ability for the generated adversarial examples, but it also expands the latter from discrete points to manifold by leveraging the domain knowledge that face expression change can be continuous, which enhances the attack effect as a data augmentation mechanism did. Moreover, we further design a dual supervision with local and global constraints as a minor contribution to improve the visual quality of the generated adversarial examples. We demonstrate the effectiveness of our method based on extensive experiments, and reveal that GMAA promises a semantic continuous adversarial space with a higher generalization ability and visual quality.

The field of text-to-image generation has made remarkable strides in creating high-fidelity and photorealistic images. As this technology gains popularity, there is a growing concern about its potential security risks. However, there has been limited exploration into the robustness of these models from an adversarial perspective. Existing research has primarily focused on untargeted settings, and lacks holistic consideration for reliability (attack success rate) and stealthiness (imperceptibility). In this paper, we propose RIATIG, a reliable and imperceptible adversarial attack against text-to-image models via inconspicuous examples. By formulating the example crafting as an optimization process and solving it using a genetic-based method, our proposed attack can generate imperceptible prompts for text-to-image generation models in a reliable way. Evaluation of six popular text-to-image generation models demonstrates the efficiency and stealthiness of our attack in both white-box and black-box settings. To allow the community to build on top of our findings, we’ve made the artifacts available.

The success of deep learning based face recognition systems has given rise to serious privacy concerns due to their ability to enable unauthorized tracking of users in the digital world. Existing methods for enhancing privacy fail to generate naturalistic’ images that can protect facial privacy without compromising user experience. We propose a novel two-step approach for facial privacy protection that relies on finding adversarial latent codes in the low-dimensional manifold of a pretrained generative model. The first step inverts the given face image into the latent space and finetunes the generative model to achieve an accurate reconstruction of the given image from its latent code. This step produces a good initialization, aiding the generation of high-quality faces that resemble the given identity. Subsequently, user defined makeup text prompts and identity-preserving regularization are used to guide the search for adversarial codes in the latent space. Extensive experiments demonstrate that faces generated by our approach have stronger black-box transferability with an absolute gain of 12.06% over the state-of-the-art facial privacy protection approach under the face verification task. Finally, we demonstrate the effectiveness of the proposed approach for commercial face recognition systems. Our code is available at https://github.com/fahadshamshad/Clip2Protect.

In this paper we present TruFor, a forensic framework that can be applied to a large variety of image manipulation methods, from classic cheapfakes to more recent manipulations based on deep learning. We rely on the extraction of both high-level and low-level traces through a transformer-based fusion architecture that combines the RGB image and a learned noise-sensitive fingerprint. The latter learns to embed the artifacts related to the camera internal and external processing by training only on real data in a self-supervised manner. Forgeries are detected as deviations from the expected regular pattern that characterizes each pristine image. Looking for anomalies makes the approach able to robustly detect a variety of local manipulations, ensuring generalization. In addition to a pixel-level localization map and a whole-image integrity score, our approach outputs a reliability map that highlights areas where localization predictions may be error-prone. This is particularly important in forensic applications in order to reduce false alarms and allow for a large scale analysis. Extensive experiments on several datasets show that our method is able to reliably detect and localize both cheapfakes and deepfakes manipulations outperforming state-of-the-art works. Code is publicly available at https://grip-unina.github.io/TruFor/.