High-fidelity radiance recovery plays a crucial role in scene information reconstruction and understanding. Conventional cameras suffer from limited sensitivity in dynamic range, bit depth, and spectral response, etc. In this paper, we propose to use event cameras with bio-inspired silicon sensors, which are sensitive to radiance changes, to recover precise radiance values. We reveal that, under active lighting conditions, the transient frequency of event signals triggering linearly reflects the radiance value. We propose an innovative method to convert the high temporal resolution of event signals into precise radiance values. The precise radiance values yields several capabilities in image analysis. We demonstrate the feasibility of recovering radiance values solely from the transient event frequency (TEF) through multiple experiments.

Neural radiance fields (NeRF) excel at synthesizing new views given multi-view, calibrated images of a static scene. When scenes include distractors, which are not persistent during image capture (moving objects, lighting variations, shadows), artifacts appear as view-dependent effects or ‘floaters’. To cope with distractors, we advocate a form of robust estimation for NeRF training, modeling distractors in training data as outliers of an optimization problem. Our method successfully removes outliers from a scene and improves upon our baselines, on synthetic and real-world scenes. Our technique is simple to incorporate in modern NeRF frameworks, with few hyper-parameters. It does not assume a priori knowledge of the types of distractors, and is instead focused on the optimization problem rather than pre-processing or modeling transient objects. More results on our page https://robustnerf.github.io/public.

2D-to-3D reconstruction is an ill-posed problem, yet humans are good at solving this problem due to their prior knowledge of the 3D world developed over years. Driven by this observation, we propose NeRDi, a single-view NeRF synthesis framework with general image priors from 2D diffusion models. Formulating single-view reconstruction as an image-conditioned 3D generation problem, we optimize the NeRF representations by minimizing a diffusion loss on its arbitrary view renderings with a pretrained image diffusion model under the input-view constraint. We leverage off-the-shelf vision-language models and introduce a two-section language guidance as conditioning inputs to the diffusion model. This is essentially helpful for improving multiview content coherence as it narrows down the general image prior conditioned on the semantic and visual features of the single-view input image. Additionally, we introduce a geometric loss based on estimated depth maps to regularize the underlying 3D geometry of the NeRF. Experimental results on the DTU MVS dataset show that our method can synthesize novel views with higher quality even compared to existing methods trained on this dataset. We also demonstrate our generalizability in zero-shot NeRF synthesis for in-the-wild images.

In this work, we focus on synthesizing high-fidelity novel view images for arbitrary human performers, given a set of sparse multi-view images. It is a challenging task due to the large variation among articulated body poses and heavy self-occlusions. To alleviate this, we introduce an effective generalizable framework Generalizable Model-based Neural Radiance Fields (GM-NeRF) to synthesize free-viewpoint images. Specifically, we propose a geometry-guided attention mechanism to register the appearance code from multi-view 2D images to a geometry proxy which can alleviate the misalignment between inaccurate geometry prior and pixel space. On top of that, we further conduct neural rendering and partial gradient backpropagation for efficient perceptual supervision and improvement of the perceptual quality of synthesis. To evaluate our method, we conduct experiments on synthesized datasets THuman2.0 and Multi-garment, and real-world datasets Genebody and ZJUMocap. The results demonstrate that our approach outperforms state-of-the-art methods in terms of novel view synthesis and geometric reconstruction.

Neural Radiance Field (NeRF) has broken new ground in the novel view synthesis due to its simple concept and state-of-the-art quality. However, it suffers from severe performance degradation unless trained with a dense set of images with different camera poses, which hinders its practical applications. Although previous methods addressing this problem achieved promising results, they relied heavily on the additional training resources, which goes against the philosophy of sparse-input novel-view synthesis pursuing the training efficiency. In this work, we propose MixNeRF, an effective training strategy for novel view synthesis from sparse inputs by modeling a ray with a mixture density model. Our MixNeRF estimates the joint distribution of RGB colors along the ray samples by modeling it with mixture of distributions. We also propose a new task of ray depth estimation as a useful training objective, which is highly correlated with 3D scene geometry. Moreover, we remodel the colors with regenerated blending weights based on the estimated ray depth and further improves the robustness for colors and viewpoints. Our MixNeRF outperforms other state-of-the-art methods in various standard benchmarks with superior efficiency of training and inference.

Neural Radiance Fields (NeRFs) have emerged as a popular approach for novel view synthesis. While NeRFs are quickly being adapted for a wider set of applications, intuitively editing NeRF scenes is still an open challenge. One important editing task is the removal of unwanted objects from a 3D scene, such that the replaced region is visually plausible and consistent with its context. We refer to this task as 3D inpainting. In 3D, solutions must be both consistent across multiple views and geometrically valid. In this paper, we propose a novel 3D inpainting method that addresses these challenges. Given a small set of posed images and sparse annotations in a single input image, our framework first rapidly obtains a 3D segmentation mask for a target object. Using the mask, a perceptual optimization-based approach is then introduced that leverages learned 2D image inpainters, distilling their information into 3D space, while ensuring view consistency. We also address the lack of a diverse benchmark for evaluating 3D scene inpainting methods by introducing a dataset comprised of challenging real-world scenes. In particular, our dataset contains views of the same scene with and without a target object, enabling more principled benchmarking of the 3D inpainting task. We first demonstrate the superiority of our approach on multiview segmentation, comparing to NeRF-based methods and 2D segmentation approaches. We then evaluate on the task of 3D inpainting, establishing state-of-the-art performance against other NeRF manipulation algorithms, as well as a strong 2D image inpainter baseline.

Neural radiance fields (NeRF) have demonstrated the potential of coordinate-based neural representation (neural fields or implicit neural representation) in neural rendering. However, using a multi-layer perceptron (MLP) to represent a 3D scene or object requires enormous computational resources and time. There have been recent studies on how to reduce these computational inefficiencies by using additional data structures, such as grids or trees. Despite the promising performance, the explicit data structure necessitates a substantial amount of memory. In this work, we present a method to reduce the size without compromising the advantages of having additional data structures. In detail, we propose using the wavelet transform on grid-based neural fields. Grid-based neural fields are for fast convergence, and the wavelet transform, whose efficiency has been demonstrated in high-performance standard codecs, is to improve the parameter efficiency of grids. Furthermore, in order to achieve a higher sparsity of grid coefficients while maintaining reconstruction quality, we present a novel trainable masking approach. Experimental results demonstrate that non-spatial grid coefficients, such as wavelet coefficients, are capable of attaining a higher level of sparsity than spatial grid coefficients, resulting in a more compact representation. With our proposed mask and compression pipeline, we achieved state-of-the-art performance within a memory budget of 2 MB. Our code is available at https://github.com/daniel03c1/maskedwaveletnerf.

Recent advances in neural radiance fields have enabled the high-fidelity 3D reconstruction of complex scenes for novel view synthesis. However, it remains underexplored how the appearance of such representations can be efficiently edited while maintaining photorealism. In this work, we present PaletteNeRF, a novel method for photorealistic appearance editing of neural radiance fields (NeRF) based on 3D color decomposition. Our method decomposes the appearance of each 3D point into a linear combination of palette-based bases (i.e., 3D segmentations defined by a group of NeRF-type functions) that are shared across the scene. While our palette-based bases are view-independent, we also predict a view-dependent function to capture the color residual (e.g., specular shading). During training, we jointly optimize the basis functions and the color palettes, and we also introduce novel regularizers to encourage the spatial coherence of the decomposition. Our method allows users to efficiently edit the appearance of the 3D scene by modifying the color palettes. We also extend our framework with compressed semantic features for semantic-aware appearance editing. We demonstrate that our technique is superior to baseline methods both quantitatively and qualitatively for appearance editing of complex real-world scenes.

Neural Radiance Fields (NeRF) have demonstrated superior novel view synthesis performance but are slow at rendering. To speed up the volume rendering process, many acceleration methods have been proposed at the cost of large memory consumption. To push the frontier of the efficiency-memory trade-off, we explore a new perspective to accelerate NeRF rendering, leveraging a key fact that the viewpoint change is usually smooth and continuous in interactive viewpoint control. This allows us to leverage the information of preceding viewpoints to reduce the number of rendered pixels as well as the number of sampled points along the ray of the remaining pixels. In our pipeline, a low-resolution feature map is rendered first by volume rendering, then a lightweight 2D neural renderer is applied to generate the output image at target resolution leveraging the features of preceding and current frames. We show that the proposed method can achieve competitive rendering quality while reducing the rendering time with little memory overhead, enabling 30FPS at 1080P image resolution with a low memory footprint.

Recent advances in 3D scene representation and novel view synthesis have witnessed the rise of Neural Radiance Fields (NeRFs). Nevertheless, it is not trivial to exploit NeRF for the photorealistic 3D scene stylization task, which aims to generate visually consistent and photorealistic stylized scenes from novel views. Simply coupling NeRF with photorealistic style transfer (PST) will result in cross-view inconsistency and degradation of stylized view syntheses. Through a thorough analysis, we demonstrate that this non-trivial task can be simplified in a new light: When transforming the appearance representation of a pre-trained NeRF with Lipschitz mapping, the consistency and photorealism across source views will be seamlessly encoded into the syntheses. That motivates us to build a concise and flexible learning framework namely LipRF, which upgrades arbitrary 2D PST methods with Lipschitz mapping tailored for the 3D scene. Technically, LipRF first pre-trains a radiance field to reconstruct the 3D scene, and then emulates the style on each view by 2D PST as the prior to learn a Lipschitz network to stylize the pre-trained appearance. In view of that Lipschitz condition highly impacts the expressivity of the neural network, we devise an adaptive regularization to balance the reconstruction and stylization. A gradual gradient aggregation strategy is further introduced to optimize LipRF in a cost-efficient manner. We conduct extensive experiments to show the high quality and robust performance of LipRF on both photorealistic 3D stylization and object appearance editing.

Our everyday lives are filled with occlusions that we strive to see through. By aggregating desired background information from different viewpoints, we can easily eliminate such occlusions without any external occlusion-free supervision. Though several occlusion removal methods have been proposed to empower machine vision systems with such ability, their performances are still unsatisfactory due to reliance on external supervision. We propose a novel method for occlusion removal by directly building a mapping between position and viewing angles and the corresponding occlusion-free scene details leveraging Neural Radiance Fields (NeRF). We also develop an effective scheme to jointly optimize camera parameters and scene reconstruction when occlusions are present. An additional depth constraint is applied to supervise the entire optimization without labeled external data for training. The experimental results on existing and newly collected datasets validate the effectiveness of our method.

Existing learning-based methods for point cloud rendering adopt various 3D representations and feature querying mechanisms to alleviate the sparsity problem of point clouds. However, artifacts still appear in the rendered images, due to the challenges in extracting continuous and discriminative 3D features from point clouds. In this paper, we present a dense while lightweight 3D representation, named TriVol, that can be combined with NeRF to render photo-realistic images from point clouds. Our TriVol consists of triple slim volumes, each of which is encoded from the input point cloud. Our representation has two advantages. First, it fuses the respective fields at different scales and thus extracts local and non-local features for discriminative representation. Second, since the volume size is greatly reduced, our 3D decoder can be efficiently inferred, allowing us to increase the resolution of the 3D space to render more point details. Extensive experiments on different benchmarks with varying kinds of scenes/objects demonstrate our framework’s effectiveness compared with current approaches. Moreover, our framework has excellent generalization ability to render a category of scenes or objects without fine-tuning.

Current Dynamic Texture Synthesis (DyTS) models can synthesize realistic videos. However, they require a slow iterative optimization process to synthesize a single fixed-size short video, and they do not offer any post-training control over the synthesis process. We propose Dynamic Neural Cellular Automata (DyNCA), a framework for real-time and controllable dynamic texture synthesis. Our method is built upon the recently introduced NCA models and can synthesize infinitely long and arbitrary-size realistic video textures in real-time. We quantitatively and qualitatively evaluate our model and show that our synthesized videos appear more realistic than the existing results. We improve the SOTA DyTS performance by 2~4 orders of magnitude. Moreover, our model offers several real-time video controls including motion speed, motion direction, and an editing brush tool. We exhibit our trained models in an online interactive demo that runs on local hardware and is accessible on personal computers and smartphones.

In this work, we aim to reconstruct a time-varying 3D model, capable of rendering photo-realistic renderings with independent control of viewpoint, illumination, and time, from Internet photos of large-scale landmarks. The core challenges are twofold. First, different types of temporal changes, such as illumination and changes to the underlying scene itself (such as replacing one graffiti artwork with another) are entangled together in the imagery. Second, scene-level temporal changes are often discrete and sporadic over time, rather than continuous. To tackle these problems, we propose a new scene representation equipped with a novel temporal step function encoding method that can model discrete scene-level content changes as piece-wise constant functions over time. Specifically, we represent the scene as a space-time radiance field with a per-image illumination embedding, where temporally-varying scene changes are encoded using a set of learned step functions. To facilitate our task of chronology reconstruction from Internet imagery, we also collect a new dataset of four scenes that exhibit various changes over time. We demonstrate that our method exhibits state-of-the-art view synthesis results on this dataset, while achieving independent control of viewpoint, time, and illumination. Code and data are available at https://zju3dv.github.io/NeuSC/.

In this paper, we focus on the problem of rendering novel views from a Neural Radiance Field (NeRF) under unobserved light conditions. To this end, we introduce a novel dataset, dubbed ReNe (Relighting NeRF), framing real world objects under one-light-at-time (OLAT) conditions, annotated with accurate ground-truth camera and light poses. Our acquisition pipeline leverages two robotic arms holding, respectively, a camera and an omni-directional point-wise light source. We release a total of 20 scenes depicting a variety of objects with complex geometry and challenging materials. Each scene includes 2000 images, acquired from 50 different points of views under 40 different OLAT conditions. By leveraging the dataset, we perform an ablation study on the relighting capability of variants of the vanilla NeRF architecture and identify a lightweight architecture that can render novel views of an object under novel light conditions, which we use to establish a non-trivial baseline for the dataset. Dataset and benchmark are available at https://eyecan-ai.github.io/rene.

Reflections on glossy objects contain valuable and hidden information about the surrounding environment. By converting these objects into cameras, we can unlock exciting applications, including imaging beyond the camera’s field-of-view and from seemingly impossible vantage points, e.g. from reflections on the human eye. However, this task is challenging because reflections depend jointly on object geometry, material properties, the 3D environment, and the observer’s viewing direction. Our approach converts glossy objects with unknown geometry into radiance-field cameras to image the world from the object’s perspective. Our key insight is to convert the object surface into a virtual sensor that captures cast reflections as a 2D projection of the 5D environment radiance field visible to and surrounding the object. We show that recovering the environment radiance fields enables depth and radiance estimation from the object to its surroundings in addition to beyond field-of-view novel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene, but not the observer. Moreover, using the radiance field we can image around occluders caused by close-by objects in the scene. Our method is trained end-to-end on multi-view images of the object and jointly estimates object geometry, diffuse radiance, and the 5D environment radiance field.

Reflective flare is a phenomenon that occurs when light reflects inside lenses, causing bright spots or a “ghosting effect” in photos, which can impact their quality. Eliminating reflective flare is highly desirable but challenging. Many existing methods rely on manually designed features to detect these bright spots, but they often fail to identify reflective flares created by various types of light and may even mistakenly remove the light sources in scenarios with multiple light sources. To address these challenges, we propose an optical center symmetry prior, which suggests that the reflective flare and light source are always symmetrical around the lens’s optical center. This prior helps to locate the reflective flare’s proposal region more accurately and can be applied to most smartphone cameras. Building on this prior, we create the first reflective flare removal dataset called BracketFlare, which contains diverse and realistic reflective flare patterns. We use continuous bracketing to capture the reflective flare pattern in the underexposed image and combine it with a normally exposed image to synthesize a pair of flare-corrupted and flare-free images. With the dataset, neural networks can be trained to remove the reflective flares effectively. Extensive experiments demonstrate the effectiveness of our method on both synthetic and real-world datasets.

A light stage uses a series of calibrated cameras and lights to capture a subject’s facial appearance under varying illumination and viewpoint. This captured information is crucial for facial reconstruction and relighting. Unfortunately, light stages are often inaccessible: they are expensive and require significant technical expertise for construction and operation. In this paper, we present SunStage: a lightweight alternative to a light stage that captures comparable data using only a smartphone camera and the sun. Our method only requires the user to capture a selfie video outdoors, rotating in place, and uses the varying angles between the sun and the face as guidance in joint reconstruction of facial geometry, reflectance, camera pose, and lighting parameters. Despite the in-the-wild un-calibrated setting, our approach is able to reconstruct detailed facial appearance and geometry, enabling compelling effects such as relighting, novel view synthesis, and reflectance editing.

Most camera lens systems are designed in isolation, separately from downstream computer vision methods. Recently, joint optimization approaches that design lenses alongside other components of the image acquisition and processing pipeline--notably, downstream neural networks--have achieved improved imaging quality or better performance on vision tasks. However, these existing methods optimize only a subset of lens parameters and cannot optimize glass materials given their categorical nature. In this work, we develop a differentiable spherical lens simulation model that accurately captures geometrical aberrations. We propose an optimization strategy to address the challenges of lens design--notorious for non-convex loss function landscapes and many manufacturing constraints--that are exacerbated in joint optimization tasks. Specifically, we introduce quantized continuous glass variables to facilitate the optimization and selection of glass materials in an end-to-end design context, and couple this with carefully designed constraints to support manufacturability. In automotive object detection, we report improved detection performance over existing designs even when simplifying designs to two- or three-element lenses, despite significantly degrading the image quality.

This paper proposes a practical method of microscale 3D shape capturing by a teleidoscopic imaging system. The main challenge in microscale 3D shape reconstruction is to capture the target from multiple viewpoints with a large enough depth-of-field. Our idea is to employ a teleidoscopic measurement system consisting of three planar mirrors and monocentric lens. The planar mirrors virtually define multiple viewpoints by multiple reflections, and the monocentric lens realizes a high magnification with less blurry and surround view even in closeup imaging. Our contributions include, a structured ray-pixel camera model which handles refractive and reflective projection rays efficiently, analytical evaluations of depth of field of our teleidoscopic imaging system, and a practical calibration algorithm of the teleidoscppic imaging system. Evaluations with real images prove the concept of our measurement system.

Neural implicit methods have achieved high-quality 3D object surfaces under slight specular highlights. However, high specular reflections (HSR) often appear in front of target objects when we capture them through glasses. The complex ambiguity in these scenes violates the multi-view consistency, then makes it challenging for recent methods to reconstruct target objects correctly. To remedy this issue, we present a novel surface reconstruction framework, NeuS-HSR, based on implicit neural rendering. In NeuS-HSR, the object surface is parameterized as an implicit signed distance function (SDF). To reduce the interference of HSR, we propose decomposing the rendered image into two appearances: the target object and the auxiliary plane. We design a novel auxiliary plane module by combining physical assumptions and neural networks to generate the auxiliary plane appearance. Extensive experiments on synthetic and real-world datasets demonstrate that NeuS-HSR outperforms state-of-the-art approaches for accurate and robust target surface reconstruction against HSR.

We present a novel method, called NeuralUDF, for reconstructing surfaces with arbitrary topologies from 2D images via volume rendering. Recent advances in neural rendering based reconstruction have achieved compelling results. However, these methods are limited to objects with closed surfaces since they adopt Signed Distance Function (SDF) as surface representation which requires the target shape to be divided into inside and outside. In this paper, we propose to represent surfaces as the Unsigned Distance Function (UDF) and develop a new volume rendering scheme to learn the neural UDF representation. Specifically, a new density function that correlates the property of UDF with the volume rendering scheme is introduced for robust optimization of the UDF fields. Experiments on the DTU and DeepFashion3D datasets show that our method not only enables high-quality reconstruction of non-closed shapes with complex typologies, but also achieves comparable performance to the SDF based methods on the reconstruction of closed surfaces. Visit our project page at https://www.xxlong.site/NeuralUDF/.

In recent years, neural distance functions trained via volumetric ray marching have been widely adopted for multi-view 3D reconstruction. These methods, however, apply the ray marching procedure for the entire scene volume, leading to reduced sampling efficiency and, as a result, lower reconstruction quality in the areas of high-frequency details. In this work, we address this problem via joint training of the implicit function and our new coarse sphere-based surface reconstruction. We use the coarse representation to efficiently exclude the empty volume of the scene from the volumetric ray marching procedure without additional forward passes of the neural surface network, which leads to an increased fidelity of the reconstructions compared to the base systems. We evaluate our approach by incorporating it into the training procedures of several implicit surface modeling methods and observe uniform improvements across both synthetic and real-world datasets. Our codebase can be accessed via the project page.

Reconstructing 3D shapes from planar cross-sections is a challenge inspired by downstream applications like medical imaging and geographic informatics. The input is an in/out indicator function fully defined on a sparse collection of planes in space, and the output is an interpolation of the indicator function to the entire volume. Previous works addressing this sparse and ill-posed problem either produce low quality results, or rely on additional priors such as target topology, appearance information, or input normal directions. In this paper, we present OReX, a method for 3D shape reconstruction from slices alone, featuring a Neural Field as the interpolation prior. A modest neural network is trained on the input planes to return an inside/outside estimate for a given 3D coordinate, yielding a powerful prior that induces smoothness and self-similarities. The main challenge for this approach is high-frequency details, as the neural prior is overly smoothing. To alleviate this, we offer an iterative estimation architecture and a hierarchical input sampling scheme that encourage coarse-to-fine training, allowing the training process to focus on high frequencies at later stages. In addition, we identify and analyze a ripple-like effect stemming from the mesh extraction step. We mitigate it by regularizing the spatial gradients of the indicator function around input in/out boundaries during network training, tackling the problem at the root. Through extensive qualitative and quantitative experimentation, we demonstrate our method is robust, accurate, and scales well with the size of the input. We report state-of-the-art results compared to previous approaches and recent potential solutions, and demonstrate the benefit of our individual contributions through analysis and ablation studies.

Despite increasingly realistic image quality, recent 3D image generative models often operate on 3D volumes of fixed extent with limited camera motions. We investigate the task of unconditionally synthesizing unbounded nature scenes, enabling arbitrarily large camera motion while maintaining a persistent 3D world model. Our scene representation consists of an extendable, planar scene layout grid, which can be rendered from arbitrary camera poses via a 3D decoder and volume rendering, and a panoramic skydome. Based on this representation, we learn a generative world model solely from single-view internet photos. Our method enables simulating long flights through 3D landscapes, while maintaining global scene consistency---for instance, returning to the starting point yields the same view of the scene. Our approach enables scene extrapolation beyond the fixed bounds of current 3D generative models, while also supporting a persistent, camera-independent world representation that stands in contrast to auto-regressive 3D prediction models. Our project page: https://chail.github.io/persistent-nature/.

Diffusion models have emerged as the state-of-the-art for image generation, among other tasks. Here, we present an efficient diffusion-based model for 3D-aware generation of neural fields. Our approach pre-processes training data, such as ShapeNet meshes, by converting them to continuous occupancy fields and factoring them into a set of axis-aligned triplane feature representations. Thus, our 3D training scenes are all represented by 2D feature planes, and we can directly train existing 2D diffusion models on these representations to generate 3D neural fields with high quality and diversity, outperforming alternative approaches to 3D-aware generation. Our approach requires essential modifications to existing triplane factorization pipelines to make the resulting features easy to learn for the diffusion model. We demonstrate state-of-the-art results on 3D generation on several object classes from ShapeNet.

We propose a 3D shape generation framework (SDF-Diffusion in short) that uses denoising diffusion models with continuous 3D representation via signed distance fields (SDF). Unlike most existing methods that depend on discontinuous forms, such as point clouds, SDF-Diffusion generates high-resolution 3D shapes while alleviating memory issues by separating the generative process into two-stage: generation and super-resolution. In the first stage, a diffusion-based generative model generates a low-resolution SDF of 3D shapes. Using the estimated low-resolution SDF as a condition, the second stage diffusion model performs super-resolution to generate high-resolution SDF. Our framework can generate a high-fidelity 3D shape despite the extreme spatial complexity. On the ShapeNet dataset, our model shows competitive performance to the state-of-the-art methods and shows applicability on the shape completion task without modification.

While current multi-frame restoration methods combine information from multiple input images using 2D alignment techniques, recent advances in novel view synthesis are paving the way for a new paradigm relying on volumetric scene representations. In this work, we introduce the first 3D-based multi-frame denoising method that significantly outperforms its 2D-based counterparts with lower computational requirements. Our method extends the multiplane image (MPI) framework for novel view synthesis by introducing a learnable encoder-renderer pair manipulating multiplane representations in feature space. The encoder fuses information across views and operates in a depth-wise manner while the renderer fuses information across depths and operates in a view-wise manner. The two modules are trained end-to-end and learn to separate depths in an unsupervised way, giving rise to Multiplane Feature (MPF) representations. Experiments on the Spaces and Real Forward-Facing datasets as well as on raw burst data validate our approach for view synthesis, multi-frame denoising, and view synthesis under noisy conditions.

Recent CLIP-guided 3D optimization methods, such as DreamFields and PureCLIPNeRF, have achieved impressive results in zero-shot text-to-3D synthesis. However, due to scratch training and random initialization without prior knowledge, these methods often fail to generate accurate and faithful 3D structures that conform to the input text. In this paper, we make the first attempt to introduce explicit 3D shape priors into the CLIP-guided 3D optimization process. Specifically, we first generate a high-quality 3D shape from the input text in the text-to-shape stage as a 3D shape prior. We then use it as the initialization of a neural radiance field and optimize it with the full prompt. To address the challenging text-to-shape generation task, we present a simple yet effective approach that directly bridges the text and image modalities with a powerful text-to-image diffusion model. To narrow the style domain gap between the images synthesized by the text-to-image diffusion model and shape renderings used to train the image-to-shape generator, we further propose to jointly optimize a learnable text prompt and fine-tune the text-to-image diffusion model for rendering-style image generation. Our method, Dream3D, is capable of generating imaginative 3D content with superior visual quality and shape accuracy compared to state-of-the-art methods. Our project page is at https://bluestyle97.github.io/dream3d/.

Despite the great success in 2D editing using user-friendly tools, such as Photoshop, semantic strokes, or even text prompts, similar capabilities in 3D areas are still limited, either relying on 3D modeling skills or allowing editing within only a few categories. In this paper, we present a novel semantic-driven NeRF editing approach, which enables users to edit a neural radiance field with a single image, and faithfully delivers edited novel views with high fidelity and multi-view consistency. To achieve this goal, we propose a prior-guided editing field to encode fine-grained geometric and texture editing in 3D space, and develop a series of techniques to aid the editing process, including cyclic constraints with a proxy mesh to facilitate geometric supervision, a color compositing mechanism to stabilize semantic-driven texture editing, and a feature-cluster-based regularization to preserve the irrelevant content unchanged. Extensive experiments and editing examples on both real-world and synthetic data demonstrate that our method achieves photo-realistic 3D editing using only a single edited image, pushing the bound of semantic-driven editing in 3D real-world scenes.

We present 3D Highlighter, a technique for localizing semantic regions on a mesh using text as input. A key feature of our system is the ability to interpret “out-of-domain” localizations. Our system demonstrates the ability to reason about where to place non-obviously related concepts on an input 3D shape, such as adding clothing to a bare 3D animal model. Our method contextualizes the text description using a neural field and colors the corresponding region of the shape using a probability-weighted blend. Our neural optimization is guided by a pre-trained CLIP encoder, which bypasses the need for any 3D datasets or 3D annotations. Thus, 3D Highlighter is highly flexible, general, and capable of producing localizations on a myriad of input shapes.

StyleGAN has achieved great progress in 2D face reconstruction and semantic editing via image inversion and latent editing. While studies over extending 2D StyleGAN to 3D faces have emerged, a corresponding generic 3D GAN inversion framework is still missing, limiting the applications of 3D face reconstruction and semantic editing. In this paper, we study the challenging problem of 3D GAN inversion where a latent code is predicted given a single face image to faithfully recover its 3D shapes and detailed textures. The problem is ill-posed: innumerable compositions of shape and texture could be rendered to the current image. Furthermore, with the limited capacity of a global latent code, 2D inversion methods cannot preserve faithful shape and texture at the same time when applied to 3D models. To solve this problem, we devise an effective self-training scheme to constrain the learning of inversion. The learning is done efficiently without any real-world 2D-3D training pairs but proxy samples generated from a 3D GAN. In addition, apart from a global latent code that captures the coarse shape and texture information, we augment the generation network with a local branch, where pixel-aligned features are added to faithfully reconstruct face details. We further consider a new pipeline to perform 3D view-consistent editing. Extensive experiments show that our method outperforms state-of-the-art inversion methods in both shape and texture reconstruction quality.

Synthesis and reconstruction of 3D human head has gained increasing interests in computer vision and computer graphics recently. Existing state-of-the-art 3D generative adversarial networks (GANs) for 3D human head synthesis are either limited to near-frontal views or hard to preserve 3D consistency in large view angles. We propose PanoHead, the first 3D-aware generative model that enables high-quality view-consistent image synthesis of full heads in 360° with diverse appearance and detailed geometry using only in-the-wild unstructured images for training. At its core, we lift up the representation power of recent 3D GANs and bridge the data alignment gap when training from in-the-wild images with widely distributed views. Specifically, we propose a novel two-stage self-adaptive image alignment for robust 3D GAN training. We further introduce a tri-grid neural volume representation that effectively addresses front-face and back-head feature entanglement rooted in the widely-adopted tri-plane formulation. Our method instills prior knowledge of 2D image segmentation in adversarial learning of 3D neural scene structures, enabling compositable head synthesis in diverse backgrounds. Benefiting from these designs, our method significantly outperforms previous 3D GANs, generating high-quality 3D heads with accurate geometry and diverse appearances, even with long wavy and afro hairstyles, renderable from arbitrary poses. Furthermore, we show that our system can reconstruct full 3D heads from single input images for personalized realistic 3D avatars.

High-fidelity kinship face synthesis has many potential applications, such as kinship verification, missing child identification, and social media analysis. However, it is challenging to synthesize high-quality descendant faces with genetic relations due to the lack of large-scale, high-quality annotated kinship data. This paper proposes RFG (Region-level Facial Gene) extraction framework to address this issue. We propose to use IGE (Image-based Gene Encoder), LGE (Latent-based Gene Encoder) and Gene Decoder to learn the RFGs of a given face image, and the relationships between RFGs and the latent space of StyleGAN2. As cycle-like losses are designed to measure the L_2 distances between the output of Gene Decoder and image encoder, and that between the output of LGE and IGE, only face images are required to train our framework, i.e. no paired kinship face data is required. Based upon the proposed RFGs, a crossover and mutation module is further designed to inherit the facial parts of parents. A Gene Pool has also been used to introduce the variations into the mutation of RFGs. The diversity of the faces of descendants can thus be significantly increased. Qualitative, quantitative, and subjective experiments on FIW, TSKinFace, and FF-Databases clearly show that the quality and diversity of kinship faces generated by our approach are much better than the existing state-of-the-art methods.

Face parsing is defined as the per-pixel labeling of images containing human faces. The labels are defined to identify key facial regions like eyes, lips, nose, hair, etc. In this work, we make use of the structural consistency of the human face to propose a lightweight face-parsing method using a Local Implicit Function network, FP-LIIF. We propose a simple architecture having a convolutional encoder and a pixel MLP decoder that uses 1/26th number of parameters compared to the state-of-the-art models and yet matches or outperforms state-of-the-art models on multiple datasets, like CelebAMask-HQ and LaPa. We do not use any pretraining, and compared to other works, our network can also generate segmentation at different resolutions without any changes in the input resolution. This work enables the use of facial segmentation on low-compute or low-bandwidth devices because of its higher FPS and smaller model size.

The function of constructing the hierarchy of objects is important to the visual process of the human brain. Previous studies have successfully adopted capsule networks to decompose the digits and faces into parts in an unsupervised manner to investigate the similar perception mechanism of neural networks. However, their descriptions are restricted to the 2D space, limiting their capacities to imitate the intrinsic 3D perception ability of humans. In this paper, we propose an Inverse Graphics Capsule Network (IGC-Net) to learn the hierarchical 3D face representations from large-scale unlabeled images. The core of IGC-Net is a new type of capsule, named graphics capsule, which represents 3D primitives with interpretable parameters in computer graphics (CG), including depth, albedo, and 3D pose. Specifically, IGC-Net first decomposes the objects into a set of semantic-consistent part-level descriptions and then assembles them into object-level descriptions to build the hierarchy. The learned graphics capsules reveal how the neural networks, oriented at visual perception, understand faces as a hierarchy of 3D models. Besides, the discovered parts can be deployed to the unsupervised face segmentation task to evaluate the semantic consistency of our method. Moreover, the part-level descriptions with explicit physical meanings provide insight into the face analysis that originally runs in a black box, such as the importance of shape and texture for face recognition. Experiments on CelebA, BP4D, and Multi-PIE demonstrate the characteristics of our IGC-Net.

3D-aware generative adversarial networks (GANs) synthesize high-fidelity and multi-view-consistent facial images using only collections of single-view 2D imagery. Towards fine-grained control over facial attributes, recent efforts incorporate 3D Morphable Face Model (3DMM) to describe deformation in generative radiance fields either explicitly or implicitly. Explicit methods provide fine-grained expression control but cannot handle topological changes caused by hair and accessories, while implicit ones can model varied topologies but have limited generalization caused by the unconstrained deformation fields. We propose a novel 3D GAN framework for unsupervised learning of generative, high-quality and 3D-consistent facial avatars from unstructured 2D images. To achieve both deformation accuracy and topological flexibility, we propose a 3D representation called Generative Texture-Rasterized Tri-planes. The proposed representation learns Generative Neural Textures on top of parametric mesh templates and then projects them into three orthogonal-viewed feature planes through rasterization, forming a tri-plane feature representation for volume rendering. In this way, we combine both fine-grained expression control of mesh-guided explicit deformation and the flexibility of implicit volumetric representation. We further propose specific modules for modeling mouth interior which is not taken into account by 3DMM. Our method demonstrates state-of-the-art 3Daware synthesis quality and animation ability through extensive experiments. Furthermore, serving as 3D prior, our animatable 3D representation boosts multiple applications including one-shot facial avatars and 3D-aware stylization.

We propose a novel 3D morphable model for complete human heads based on hybrid neural fields. At the core of our model lies a neural parametric representation that disentangles identity and expressions in disjoint latent spaces. To this end, we capture a person’s identity in a canonical space as a signed distance field (SDF), and model facial expressions with a neural deformation field. In addition, our representation achieves high-fidelity local detail by introducing an ensemble of local fields centered around facial anchor points. To facilitate generalization, we train our model on a newly-captured dataset of over 3700 head scans from 203 different identities using a custom high-end 3D scanning setup. Our dataset significantly exceeds comparable existing datasets, both with respect to quality and completeness of geometry, averaging around 3.5M mesh faces per scan. Finally, we demonstrate that our approach outperforms state-of-the-art methods in terms of fitting error and reconstruction quality.

Recent popular Role-Playing Games (RPGs) saw the great success of character auto-creation systems. The bone-driven face model controlled by continuous parameters (like the position of bones) and discrete parameters (like the hairstyles) makes it possible for users to personalize and customize in-game characters. Previous in-game character auto-creation systems are mostly image-driven, where facial parameters are optimized so that the rendered character looks similar to the reference face photo. This paper proposes a novel text-to-parameter translation method (T2P) to achieve zero-shot text-driven game character auto-creation. With our method, users can create a vivid in-game character with arbitrary text description without using any reference photo or editing hundreds of parameters manually. In our method, taking the power of large-scale pre-trained multi-modal CLIP and neural rendering, T2P searches both continuous facial parameters and discrete facial parameters in a unified framework. Due to the discontinuous parameter representation, previous methods have difficulty in effectively learning discrete facial parameters. T2P, to our best knowledge, is the first method that can handle the optimization of both discrete and continuous parameters. Experimental results show that T2P can generate high-quality and vivid game characters with given text prompts. T2P outperforms other SOTA text-to-3D generation methods on both objective evaluations and subjective evaluations.

In this paper, we propose a novel hybrid representation and end-to-end trainable network architecture to model fully editable and customizable neural avatars. At the core of our work lies a representation that combines the modeling power of neural fields with the ease of use and inherent 3D consistency of skinned meshes. To this end, we construct a trainable feature codebook to store local geometry and texture features on the vertices of a deformable body model, thus exploiting its consistent topology under articulation. This representation is then employed in a generative auto-decoder architecture that admits fitting to unseen scans and sampling of realistic avatars with varied appearances and geometries. Furthermore, our representation allows local editing by swapping local features between 3D assets. To verify our method for avatar creation and editing, we contribute a new high-quality dataset, dubbed CustomHumans, for training and evaluation. Our experiments quantitatively and qualitatively show that our method generates diverse detailed avatars and achieves better model fitting performance compared to state-of-the-art methods. Our code and dataset are available at https://ait.ethz.ch/custom-humans.

Real-time and robust photorealistic avatars for telepresence in AR/VR have been highly desired for enabling immersive photorealistic telepresence. However, there still exists one key bottleneck: the considerable computational expense needed to accurately infer facial expressions captured from headset-mounted cameras with a quality level that can match the realism of the avatar’s human appearance. To this end, we propose a framework called Auto-CARD, which for the first time enables real-time and robust driving of Codec Avatars when exclusively using merely on-device computing resources. This is achieved by minimizing two sources of redundancy. First, we develop a dedicated neural architecture search technique called AVE-NAS for avatar encoding in AR/VR, which explicitly boosts both the searched architectures’ robustness in the presence of extreme facial expressions and hardware friendliness on fast evolving AR/VR headsets. Second, we leverage the temporal redundancy in consecutively captured images during continuous rendering and develop a mechanism dubbed LATEX to skip the computation of redundant frames. Specifically, we first identify an opportunity from the linearity of the latent space derived by the avatar decoder and then propose to perform adaptive latent extrapolation for redundant frames. For evaluation, we demonstrate the efficacy of our Auto-CARD framework in real-time Codec Avatar driving settings, where we achieve a 5.05x speed-up on Meta Quest 2 while maintaining a comparable or even better animation quality than state-of-the-art avatar encoder designs.

Translating spoken languages into Sign languages is necessary for open communication between the hearing and hearing-impaired communities. To achieve this goal, we propose the first method for animating a text written in HamNoSys, a lexical Sign language notation, into signed pose sequences. As HamNoSys is universal by design, our proposed method offers a generic solution invariant to the target Sign language. Our method gradually generates pose predictions using transformer encoders that create meaningful representations of the text and poses while considering their spatial and temporal information. We use weak supervision for the training process and show that our method succeeds in learning from partial and inaccurate data. Additionally, we offer a new distance measurement that considers missing keypoints, to measure the distance between pose sequences using DTW-MJE. We validate its correctness using AUTSL, a large-scale Sign language dataset, show that it measures the distance between pose sequences more accurately than existing measurements, and use it to assess the quality of our generated pose sequences. Code for the data pre-processing, the model, and the distance measurement is publicly released for future research.

The ability to create realistic animatable and relightable head avatars from casual video sequences would open up wide ranging applications in communication and entertainment. Current methods either build on explicit 3D morphable meshes (3DMM) or exploit neural implicit representations. The former are limited by fixed topology, while the latter are non-trivial to deform and inefficient to render. Furthermore, existing approaches entangle lighting and albedo, limiting the ability to re-render the avatar in new environments. In contrast, we propose PointAvatar, a deformable point-based representation that disentangles the source color into intrinsic albedo and normal-dependent shading. We demonstrate that PointAvatar bridges the gap between existing mesh- and implicit representations, combining high-quality geometry and appearance with topological flexibility, ease of deformation and rendering efficiency. We show that our method is able to generate animatable 3D avatars using monocular videos from multiple sources including hand-held smartphones, laptop webcams and internet videos, achieving state-of-the-art quality in challenging cases where previous methods fail, e.g., thin hair strands, while being significantly more efficient in training than competing methods.

We propose PAniC-3D, a system to reconstruct stylized 3D character heads directly from illustrated (p)ortraits of (ani)me (c)haracters. Our anime-style domain poses unique challenges to single-view reconstruction; compared to natural images of human heads, character portrait illustrations have hair and accessories with more complex and diverse geometry, and are shaded with non-photorealistic contour lines. In addition, there is a lack of both 3D model and portrait illustration data suitable to train and evaluate this ambiguous stylized reconstruction task. Facing these challenges, our proposed PAniC-3D architecture crosses the illustration-to-3D domain gap with a line-filling model, and represents sophisticated geometries with a volumetric radiance field. We train our system with two large new datasets (11.2k Vroid 3D models, 1k Vtuber portrait illustrations), and evaluate on a novel AnimeRecon benchmark of illustration-to-3D pairs. PAniC-3D significantly outperforms baseline methods, and provides data to establish the task of stylized reconstruction from portrait illustrations.

We propose a novel framework to reconstruct accurate appearance and geometry with neural radiance fields (NeRF) for interacting hands, enabling the rendering of photo-realistic images and videos for gesture animation from arbitrary views. Given multi-view images of a single hand or interacting hands, an off-the-shelf skeleton estimator is first employed to parameterize the hand poses. Then we design a pose-driven deformation field to establish correspondence from those different poses to a shared canonical space, where a pose-disentangled NeRF for one hand is optimized. Such unified modeling efficiently complements the geometry and texture cues in rarely-observed areas for both hands. Meanwhile, we further leverage the pose priors to generate pseudo depth maps as guidance for occlusion-aware density learning. Moreover, a neural feature distillation method is proposed to achieve cross-domain alignment for color optimization. We conduct extensive experiments to verify the merits of our proposed HandNeRF and report a series of state-of-the-art results both qualitatively and quantitatively on the large-scale InterHand2.6M dataset.

The task of human reposing involves generating a realistic image of a model standing in an arbitrary conceivable pose. There are multiple difficulties in generating perceptually accurate images and existing methods suffers from limitations in preserving texture, maintaining pattern coherence, respecting cloth boundaries, handling occlusions, manipulating skin generation etc. These difficulties are further exacerbated by the fact that the possible space of pose orientation for humans is large and variable, the nature of clothing items are highly non-rigid and the diversity in body shape differ largely among the population. To alleviate these difficulties and synthesize perceptually accurate images, we propose VGFlow, a model which uses a visibility guided flow module to disentangle the flow into visible and invisible parts of the target for simultaneous texture preservation and style manipulation. Furthermore, to tackle distinct body shapes and avoid network artifacts, we also incorporate an a self-supervised patch-wise ”realness” loss to further improve the output. VGFlow achieves state-of-the-art results as observed qualitatively and quantitatively on different image quality metrics(SSIM, LPIPS, FID).

This paper addresses the challenge of capturing performance for the clothed humans from sparse-view or monocular videos. Previous methods capture the performance of full humans with a personalized template or recover the garments from a single frame with static human poses. However, it is inconvenient to extract cloth semantics and capture clothing motion with one-piece template, while single frame-based methods may suffer from instable tracking across videos. To address these problems, we propose a novel method for human performance capture by tracking clothing and human body motion separately with a double-layer neural radiance fields (NeRFs). Specifically, we propose a double-layer NeRFs for the body and garments, and track the densely deforming template of the clothing and body by jointly optimizing the deformation fields and the canonical double-layer NeRFs. In the optimization, we introduce a physics-aware cloth simulation network which can help generate physically plausible cloth dynamics and body-cloth interactions. Compared with existing methods, our method is fully differentiable and can capture both the body and clothing motion robustly from dynamic videos. Also, our method represents the clothing with an independent NeRFs, allowing us to model implicit fields of general clothes feasibly. The experimental evaluations validate its effectiveness on real multi-view or monocular videos.

Enable neural networks to capture 3D geometrical-aware features is essential in multi-view based vision tasks. Previous methods usually encode the 3D information of multi-view stereo into the 2D features. In contrast, we present a novel method, named POEM, that directly operates on the 3D POints Embedded in the Multi-view stereo for reconstructing hand mesh in it. Point is a natural form of 3D information and an ideal medium for fusing features across views, as it has different projections on different views. Our method is thus in light of a simple yet effective idea, that a complex 3D hand mesh can be represented by a set of 3D points that 1) are embedded in the multi-view stereo, 2) carry features from the multi-view images, and 3) encircle the hand. To leverage the power of points, we design two operations: point-based feature fusion and cross-set point attention mechanism. Evaluation on three challenging multi-view datasets shows that POEM outperforms the state-of-the-art in hand mesh reconstruction. Code and models are available for research at github.com/lixiny/POEM

We present FlexNeRF, a method for photorealistic free-viewpoint rendering of humans in motion from monocular videos. Our approach works well with sparse views, which is a challenging scenario when the subject is exhibiting fast/complex motions. We propose a novel approach which jointly optimizes a canonical time and pose configuration, with a pose-dependent motion field and pose-independent temporal deformations complementing each other. Thanks to our novel temporal and cyclic consistency constraints along with additional losses on intermediate representation such as segmentation, our approach provides high quality outputs as the observed views become sparser. We empirically demonstrate that our method significantly outperforms the state-of-the-art on public benchmark datasets as well as a self-captured fashion dataset. The project page is available at: https://flex-nerf.github.io/.

In this paper we present a new method for deformable NeRF that can directly use optical flow as supervision. We overcome the major challenge with respect to the computationally inefficiency of enforcing the flow constraints to the backward deformation field, used by deformable NeRFs. Specifically, we show that inverting the backward deformation function is actually not needed for computing scene flows between frames. This insight dramatically simplifies the problem, as one is no longer constrained to deformation functions that can be analytically inverted. Instead, thanks to the weak assumptions required by our derivation based on the inverse function theorem, our approach can be extended to a broad class of commonly used backward deformation field. We present results on monocular novel view synthesis with rapid object motion, and demonstrate significant improvements over baselines without flow supervision.

We build rearticulable models for arbitrary everyday man-made objects containing an arbitrary number of parts that are connected together in arbitrary ways via 1-degree-of-freedom joints. Given point cloud videos of such everyday objects, our method identifies the distinct object parts, what parts are connected to what other parts, and the properties of the joints connecting each part pair. We do this by jointly optimizing the part segmentation, transformation, and kinematics using a novel energy minimization framework. Our inferred animatable models, enables retargeting to novel poses with sparse point correspondences guidance. We test our method on a new articulating robot dataset and the Sapiens dataset with common daily objects. Experiments show that our method outperforms two leading prior works on various metrics.

From an image of a person, we can easily infer the natural 3D pose and shape of the person even if ambiguity exists. This is because we have a mental model that allows us to imagine a person’s appearance at different viewing directions from a given image and utilize the consistency between them for inference. However, existing human mesh recovery methods only consider the direction in which the image was taken due to their structural limitations. Hence, we propose “Implicit 3D Human Mesh Recovery (ImpHMR)” that can implicitly imagine a person in 3D space at the feature-level via Neural Feature Fields. In ImpHMR, feature fields are generated by CNN-based image encoder for a given image. Then, the 2D feature map is volume-rendered from the feature field for a given viewing direction, and the pose and shape parameters are regressed from the feature. To utilize consistency with pose and shape from unseen-view, if there are 3D labels, the model predicts results including the silhouette from an arbitrary direction and makes it equal to the rotated ground-truth. In the case of only 2D labels, we perform self-supervised learning through the constraint that the pose and shape parameters inferred from different directions should be the same. Extensive evaluations show the efficacy of the proposed method.

Whole-body mesh recovery aims to estimate the 3D human body, face, and hands parameters from a single image. It is challenging to perform this task with a single network due to resolution issues, i.e., the face and hands are usually located in extremely small regions. Existing works usually detect hands and faces, enlarge their resolution to feed in a specific network to predict the parameter, and finally fuse the results. While this copy-paste pipeline can capture the fine-grained details of the face and hands, the connections between different parts cannot be easily recovered in late fusion, leading to implausible 3D rotation and unnatural pose. In this work, we propose a one-stage pipeline for expressive whole-body mesh recovery, named OSX, without separate networks for each part. Specifically, we design a Component Aware Transformer (CAT) composed of a global body encoder and a local face/hand decoder. The encoder predicts the body parameters and provides a high-quality feature map for the decoder, which performs a feature-level upsample-crop scheme to extract high-resolution part-specific features and adopt keypoint-guided deformable attention to estimate hand and face precisely. The whole pipeline is simple yet effective without any manual post-processing and naturally avoids implausible prediction. Comprehensive experiments demonstrate the effectiveness of OSX. Lastly, we build a large-scale Upper-Body dataset (UBody) with high-quality 2D and 3D whole-body annotations. It contains persons with partially visible bodies in diverse real-life scenarios to bridge the gap between the basic task and downstream applications.

We present Implicit Two Hands (Im2Hands), the first neural implicit representation of two interacting hands. Unlike existing methods on two-hand reconstruction that rely on a parametric hand model and/or low-resolution meshes, Im2Hands can produce fine-grained geometry of two hands with high hand-to-hand and hand-to-image coherency. To handle the shape complexity and interaction context between two hands, Im2Hands models the occupancy volume of two hands -- conditioned on an RGB image and coarse 3D keypoints -- by two novel attention-based modules responsible for (1) initial occupancy estimation and (2) context-aware occupancy refinement, respectively. Im2Hands first learns per-hand neural articulated occupancy in the canonical space designed for each hand using query-image attention. It then refines the initial two-hand occupancy in the posed space to enhance the coherency between the two hand shapes using query-anchor attention. In addition, we introduce an optional keypoint refinement module to enable robust two-hand shape estimation from predicted hand keypoints in a single-image reconstruction scenario. We experimentally demonstrate the effectiveness of Im2Hands on two-hand reconstruction in comparison to related methods, where ours achieves state-of-the-art results. Our code is publicly available at https://github.com/jyunlee/Im2Hands.

Synthesizing 3D human avatars interacting realistically with a scene is an important problem with applications in AR/VR, video games, and robotics. Towards this goal, we address the task of generating a virtual human -- hands and full body -- grasping everyday objects. Existing methods approach this problem by collecting a 3D dataset of humans interacting with objects and training on this data. However, 1) these methods do not generalize to different object positions and orientations or to the presence of furniture in the scene, and 2) the diversity of their generated full-body poses is very limited. In this work, we address all the above challenges to generate realistic, diverse full-body grasps in everyday scenes without requiring any 3D full-body grasping data. Our key insight is to leverage the existence of both full-body pose and hand-grasping priors, composing them using 3D geometrical constraints to obtain full-body grasps. We empirically validate that these constraints can generate a variety of feasible human grasps that are superior to baselines both quantitatively and qualitatively.

To enable general-purpose robots, we will require the robot to operate daily articulated objects as humans do. Current robot manipulation has heavily relied on using a parallel gripper, which restricts the robot to a limited set of objects. On the other hand, operating with a multi-finger robot hand will allow better approximation to human behavior and enable the robot to operate on diverse articulated objects. To this end, we propose a new benchmark called DexArt, which involves Dexterous manipulation with Articulated objects in a physical simulator. In our benchmark, we define multiple complex manipulation tasks, and the robot hand will need to manipulate diverse articulated objects within each task. Our main focus is to evaluate the generalizability of the learned policy on unseen articulated objects. This is very challenging given the high degrees of freedom of both hands and objects. We use Reinforcement Learning with 3D representation learning to achieve generalization. Through extensive studies, we provide new insights into how 3D representation learning affects decision making in RL with 3D point cloud inputs. More details can be found at https://www.chenbao.tech/dexart/.

We present CARTO, a novel approach for reconstructing multiple articulated objects from a single stereo RGB observation. We use implicit object-centric representations and learn a single geometry and articulation decoder for multiple object categories. Despite training on multiple categories, our decoder achieves a comparable reconstruction accuracy to methods that train bespoke decoders separately for each category. Combined with our stereo image encoder we infer the 3D shape, 6D pose, size, joint type, and the joint state of multiple unknown objects in a single forward pass. Our method achieves a 20.4% absolute improvement in mAP 3D IOU50 for novel instances when compared to a two-stage pipeline. Inference time is fast and can run on a NVIDIA TITAN XP GPU at 1 HZ for eight or less objects present. While only trained on simulated data, CARTO transfers to real-world object instances. Code and evaluation data is available at: http://carto.cs.uni-freiburg.de

Synthesizing 3D human motion in a contextual, ecological environment is important for simulating realistic activities people perform in the real world. However, conventional optics-based motion capture systems are not suited for simultaneously capturing human movements and complex scenes. The lack of rich contextual 3D human motion datasets presents a roadblock to creating high-quality generative human motion models. We propose a novel motion acquisition system in which the actor perceives and operates in a highly contextual virtual world while being motion captured in the real world. Our system enables rapid collection of high-quality human motion in highly diverse scenes, without the concern of occlusion or the need for physical scene construction in the real world. We present CIRCLE, a dataset containing 10 hours of full-body reaching motion from 5 subjects across nine scenes, paired with ego-centric information of the environment represented in various forms, such as RGBD videos. We use this dataset to train a model that generates human motion conditioned on scene information. Leveraging our dataset, the model learns to use ego-centric scene information to achieve nontrivial reaching tasks in the context of complex 3D scenes. To download the data please visit our website (https://stanford-tml.github.io/circle_dataset/).

We propose a method to reconstruct global human trajectories from videos in the wild. Our optimization method decouples the camera and human motion, which allows us to place people in the same world coordinate frame. Most existing methods do not model the camera motion; methods that rely on the background pixels to infer 3D human motion usually require a full scene reconstruction, which is often not possible for in-the-wild videos. However, even when existing SLAM systems cannot recover accurate scene reconstructions, the background pixel motion still provides enough signal to constrain the camera motion. We show that relative camera estimates along with data-driven human motion priors can resolve the scene scale ambiguity and recover global human trajectories. Our method robustly recovers the global 3D trajectories of people in challenging in-the-wild videos, such as PoseTrack. We quantify our improvement over existing methods on 3D human dataset Egobody. We further demonstrate that our recovered camera scale allows us to reason about motion of multiple people in a shared coordinate frame, which improves performance of downstream tracking in PoseTrack. Code and additional results can be found at https://vye16.github.io/slahmr/.

Garments are important to humans. A visual system that can estimate and track the complete garment pose can be useful for many downstream tasks and real-world applications. In this work, we present a complete package to address the category-level garment pose tracking task: (1) A recording system VR-Garment, with which users can manipulate virtual garment models in simulation through a VR interface. (2) A large-scale dataset VR-Folding, with complex garment pose configurations in manipulation like flattening and folding. (3) An end-to-end online tracking framework GarmentTracking, which predicts complete garment pose both in canonical space and task space given a point cloud sequence. Extensive experiments demonstrate that the proposed GarmentTracking achieves great performance even when the garment has large non-rigid deformation. It outperforms the baseline approach on both speed and accuracy. We hope our proposed solution can serve as a platform for future research. Codes and datasets are available in https://garment-tracking.robotflow.ai.

Understanding dynamic hand motions and actions from egocentric RGB videos is a fundamental yet challenging task due to self-occlusion and ambiguity. To address occlusion and ambiguity, we develop a transformer-based framework to exploit temporal information for robust estimation. Noticing the different temporal granularity of and the semantic correlation between hand pose estimation and action recognition, we build a network hierarchy with two cascaded transformer encoders, where the first one exploits the short-term temporal cue for hand pose estimation, and the latter aggregates per-frame pose and object information over a longer time span to recognize the action. Our approach achieves competitive results on two first-person hand action benchmarks, namely FPHA and H2O. Extensive ablation studies verify our design choices.

Existing methods of multi-person video 3D human Pose and Shape Estimation (PSE) typically adopt a two-stage strategy, which first detects human instances in each frame and then performs single-person PSE with temporal model. However, the global spatio-temporal context among spatial instances can not be captured. In this paper, we propose a new end-to-end multi-person 3D Pose and Shape estimation framework with progressive Video Transformer, termed PSVT. In PSVT, a spatio-temporal encoder (STE) captures the global feature dependencies among spatial objects. Then, spatio-temporal pose decoder (STPD) and shape decoder (STSD) capture the global dependencies between pose queries and feature tokens, shape queries and feature tokens, respectively. To handle the variances of objects as time proceeds, a novel scheme of progressive decoding is used to update pose and shape queries at each frame. Besides, we propose a novel pose-guided attention (PGA) for shape decoder to better predict shape parameters. The two components strengthen the decoder of PSVT to improve performance. Extensive experiments on the four datasets show that PSVT achieves stage-of-the-art results.

Normalizing flows (NFs) provide a powerful tool to construct an expressive distribution by a sequence of trackable transformations of a base distribution and form a probabilistic model of underlying data.Rotation, as an important quantity in computer vision, graphics, and robotics, can exhibit many ambiguities when occlusion and symmetry occur and thus demands such probabilistic models. Though much progress has been made for NFs in Euclidean space, there are no effective normalizing flows without discontinuity or many-to-one mapping tailored for SO(3) manifold. Given the unique non-Euclidean properties of the rotation manifold, adapting the existing NFs to SO(3) manifold is non-trivial. In this paper, we propose a novel normalizing flow on SO(3) by combining a Mobius transformation-based coupling layer and a quaternion affine transformation. With our proposed rotation normalizing flows, one can not only effectively express arbitrary distributions on SO(3), but also conditionally build the target distribution given input observations. Extensive experiments show that our rotation normalizing flows significantly outperform the baselines on both unconditional and conditional tasks.

Recent advances in machine learning and computer vision are revolutionizing the field of animal behavior by enabling researchers to track the poses and locations of freely moving animals without any marker attachment. However, large datasets of annotated images of animals for markerless pose tracking, especially high-resolution images taken from multiple angles with accurate 3D annotations, are still scant. Here, we propose a method that uses a motion capture (mo-cap) system to obtain a large amount of annotated data on animal movement and posture (2D and 3D) in a semi-automatic manner. Our method is novel in that it extracts the 3D positions of morphological keypoints (e.g eyes, beak, tail) in reference to the positions of markers attached to the animals. Using this method, we obtained, and offer here, a new dataset - 3D-POP with approximately 300k annotated frames (4 million instances) in the form of videos having groups of one to ten freely moving birds from 4 different camera views in a 3.6m x 4.2m area. 3D-POP is the first dataset of flocking birds with accurate keypoint annotations in 2D and 3D along with bounding box and individual identities and will facilitate the development of solutions for problems of 2D to 3D markerless pose, trajectory tracking, and identification in birds.

Test-time adaptation methods have been gaining attention recently as a practical solution for addressing source-to-target domain gaps by gradually updating the model without requiring labels on the target data. In this paper, we propose a method of test-time adaptation for category-level object pose estimation called TTA-COPE. We design a pose ensemble approach with a self-training loss using pose-aware confidence. Unlike previous unsupervised domain adaptation methods for category-level object pose estimation, our approach processes the test data in a sequential, online manner, and it does not require access to the source domain at runtime. Extensive experimental results demonstrate that the proposed pose ensemble and the self-training loss improve category-level object pose performance during test time under both semi-supervised and unsupervised settings.

Solving the camera-to-robot pose is a fundamental requirement for vision-based robot control, and is a process that takes considerable effort and cares to make accurate. Traditional approaches require modification of the robot via markers, and subsequent deep learning approaches enabled markerless feature extraction. Mainstream deep learning methods only use synthetic data and rely on Domain Randomization to fill the sim-to-real gap, because acquiring the 3D annotation is labor-intensive. In this work, we go beyond the limitation of 3D annotations for real-world data. We propose an end-to-end pose estimation framework that is capable of online camera-to-robot calibration and a self-supervised training method to scale the training to unlabeled real-world data. Our framework combines deep learning and geometric vision for solving the robot pose, and the pipeline is fully differentiable. To train the Camera-to-Robot Pose Estimation Network (CtRNet), we leverage foreground segmentation and differentiable rendering for image-level self-supervision. The pose prediction is visualized through a renderer and the image loss with the input image is back-propagated to train the neural network. Our experimental results on two public real datasets confirm the effectiveness of our approach over existing works. We also integrate our framework into a visual servoing system to demonstrate the promise of real-time precise robot pose estimation for automation tasks.

Recently, self-supervised 6D object pose estimation, where synthetic images with object poses (sometimes jointly with un-annotated real images) are used for training, has attracted much attention in computer vision. Some typical works in literature employ a time-consuming differentiable renderer for object pose prediction at the training stage, so that (i) their performances on real images are generally limited due to the gap between their rendered images and real images and (ii) their training process is computationally expensive. To address the two problems, we propose a novel Network for Self-supervised Monocular Object pose estimation by utilizing the predicted Camera poses from un-annotated real images, called SMOC-Net. The proposed network is explored under a knowledge distillation framework, consisting of a teacher model and a student model. The teacher model contains a backbone estimation module for initial object pose estimation, and an object pose refiner for refining the initial object poses using a geometric constraint (called relative-pose constraint) derived from relative camera poses. The student model gains knowledge for object pose estimation from the teacher model by imposing the relative-pose constraint. Thanks to the relative-pose constraint, SMOC-Net could not only narrow the domain gap between synthetic and real data but also reduce the training cost. Experimental results on two public datasets demonstrate that SMOC-Net outperforms several state-of-the-art methods by a large margin while requiring much less training time than the differentiable-renderer-based methods.

Previous methods solve feature matching and pose estimation using a two-stage process by first finding matches and then estimating the pose. As they ignore the geometric relationships between the two tasks, they focus on either improving the quality of matches or filtering potential outliers, leading to limited efficiency or accuracy. In contrast, we propose an iterative matching and pose estimation framework (IMP) leveraging the geometric connections between the two tasks: a few good matches are enough for a roughly accurate pose estimation; a roughly accurate pose can be used to guide the matching by providing geometric constraints. To this end, we implement a geometry-aware recurrent module with transformers which jointly outputs sparse matches and camera poses. Specifically, for each iteration, we first implicitly embed geometric information into the module via a pose-consistency loss, allowing it to predict geometry-aware matches progressively. Second, we introduce an efficient IMP (EIMP) to dynamically discard keypoints without potential matches, avoiding redundant updating and significantly reducing the quadratic time complexity of attention computation in transformers. Experiments on YFCC100m, Scannet, and Aachen Day-Night datasets demonstrate that the proposed method outperforms previous approaches in terms of accuracy and efficiency.

Virtually every object in the modern world was created, modified, analyzed and optimized using computer aided design (CAD) tools. An active CAD research area is the use of data-driven machine learning methods to learn from the massive repositories of geometric and program representations. However, the lack of labeled data in CAD’s native format, i.e., the parametric boundary representation (B-Rep), poses an obstacle at present difficult to overcome. Several datasets of mechanical parts in B-Rep format have recently been released for machine learning research. However, large-scale databases are mostly unlabeled, and labeled datasets are small. Additionally, task-specific label sets are rare and costly to annotate. This work proposes to leverage unlabeled CAD geometry on supervised learning tasks. We learn a novel, hybrid implicit/explicit surface representation for B-Rep geometry. Further, we show that this pre-training both significantly improves few-shot learning performance and achieves state-of-the-art performance on several current B-Rep benchmarks.

Supervised keypoint localization methods rely on large manually labeled image datasets, where objects can deform, articulate, or occlude. However, creating such large keypoint labels is time-consuming and costly, and is often error-prone due to inconsistent labeling. Thus, we desire an approach that can learn keypoint localization with fewer yet consistently annotated images. To this end, we present a novel formulation that learns to localize semantically consistent keypoint definitions, even for occluded regions, for varying object categories. We use a few user-labeled 2D images as input examples, which are extended via self-supervision using a larger unlabeled dataset. Unlike unsupervised methods, the few-shot images act as semantic shape constraints for object localization. Furthermore, we introduce 3D geometry-aware constraints to uplift keypoints, achieving more accurate 2D localization. Our general-purpose formulation paves the way for semantically conditioned generative modeling and attains competitive or state-of-the-art accuracy on several datasets, including human faces, eyes, animals, cars, and never-before-seen mouth interior (teeth) localization tasks, not attempted by the previous few-shot methods. Project page: https://xingzhehe.github.io/FewShot3DKP/

Camera pose estimation is a key step in standard 3D reconstruction pipelines that operates on a dense set of images of a single object or scene. However, methods for pose estimation often fail when there are only a few images available because they rely on the ability to robustly identify and match visual features between pairs of images. While these methods can work robustly with dense camera views, capturing a large set of images can be time consuming or impractical. Here, we propose Sparse-View Camera Pose Regression and Refinement (SparsePose) for recovering accurate camera poses given a sparse set of wide-baseline images (fewer than 10). The method learns to regress initial camera poses and then iteratively refine them after training on a large-scale dataset of objects (Co3D: Common Objects in 3D). SparsePose significantly outperforms conventional and learning-based baselines in recovering accurate camera rotations and translations. We also demonstrate our pipeline for high-fidelity 3D reconstruction using only 5-9 images of an object.

We present a large-scale dataset of Planes in 3D, Pi3D, of roughly 1000 planes observed in 10 000 images from the 1DSfM dataset, and HEB, a large-scale homography estimation benchmark leveraging Pi3D. The applications of the Pi3D dataset are diverse, e.g. training or evaluating monocular depth, surface normal estimation and image matching algorithms. The HEB dataset consists of 226 260 homographies and includes roughly 4M correspondences. The homographies link images that often undergo significant viewpoint and illumination changes. As applications of HEB, we perform a rigorous evaluation of a wide range of robust estimators and deep learning-based correspondence filtering methods, establishing the current state-of-the-art in robust homography estimation. We also evaluate the uncertainty of the SIFT orientations and scales w.r.t. the ground truth coming from the underlying homographies and provide codes for comparing uncertainty of custom detectors.

Cross-modal training using 2D-3D paired datasets, such as those containing multi-view images and 3D scene scans, presents an effective way to enhance 2D scene understanding by introducing geometric and view-invariance priors into 2D features. However, the need for large-scale scene datasets can impede scalability and further improvements. This paper explores an alternative learning method by leveraging a lightweight and publicly available type of 3D data in the form of CAD models. We construct a 3D space with geometric-aware alignment where the similarity in this space reflects the geometric similarity of CAD models based on the Chamfer distance. The acquired geometric-aware properties are then induced into 2D features, which boost performance on downstream tasks more effectively than existing RGB-CAD approaches. Our technique is not limited to paired RGB-CAD datasets. By training exclusively on pseudo pairs generated from CAD-based reconstruction methods, we enhance the performance of SOTA 2D pre-trained models that use ResNet-50 or ViT-B backbones on various 2D understanding tasks. We also achieve comparable results to SOTA methods trained on scene scans on four tasks in NYUv2, SUNRGB-D, indoor ADE20k, and indoor/outdoor COCO, despite using lightweight CAD models or pseudo data.

A fully automated object reconstruction pipeline is crucial for digital content creation. While the area of 3D reconstruction has witnessed profound developments, the removal of background to obtain a clean object model still relies on different forms of manual labor, such as bounding box labeling, mask annotations, and mesh manipulations. In this paper, we propose a novel framework named AutoRecon for the automated discovery and reconstruction of an object from multi-view images. We demonstrate that foreground objects can be robustly located and segmented from SfM point clouds by leveraging self-supervised 2D vision transformer features. Then, we reconstruct decomposed neural scene representations with dense supervision provided by the decomposed point clouds, resulting in accurate object reconstruction and segmentation. Experiments on the DTU, BlendedMVS and CO3D-V2 datasets demonstrate the effectiveness and robustness of AutoRecon. The code and supplementary material are available on the project page: https://zju3dv.github.io/autorecon/.

We present a new multi-sensor dataset for multi-view 3D surface reconstruction. It includes registered RGB and depth data from sensors of different resolutions and modalities: smartphones, Intel RealSense, Microsoft Kinect, industrial cameras, and structured-light scanner. The scenes are selected to emphasize a diverse set of material properties challenging for existing algorithms. We provide around 1.4 million images of 107 different scenes acquired from 100 viewing directions under 14 lighting conditions. We expect our dataset will be useful for evaluation and training of 3D reconstruction algorithms and for related tasks. The dataset is available at skoltech3d.appliedai.tech.

Monocular 3D object localization in driving scenes is a crucial task, but challenging due to its ill-posed nature. Estimating 3D coordinates for each pixel on the object surface holds great potential as it provides dense 2D-3D geometric constraints for the underlying PnP problem. However, high-quality ground truth supervision is not available in driving scenes due to sparsity and various artifacts of Lidar data, as well as the practical infeasibility of collecting per-instance CAD models. In this work, we present NeurOCS, a framework that uses instance masks and 3D boxes as input to learn 3D object shapes by means of differentiable rendering, which further serves as supervision for learning dense object coordinates. Our approach rests on insights in learning a category-level shape prior directly from real driving scenes, while properly handling single-view ambiguities. Furthermore, we study and make critical design choices to learn object coordinates more effectively from an object-centric view. Altogether, our framework leads to new state-of-the-art in monocular 3D localization that ranks 1st on the KITTI-Object benchmark among published monocular methods.

Neural implicit surface representation methods show impressive reconstruction results but struggle to handle texture-less planar regions that widely exist in indoor scenes. Existing approaches addressing this leverage image prior that requires assistive networks trained with large-scale annotated datasets. In this work, we introduce a self-supervised super-plane constraint by exploring the free geometry cues from the predicted surface, which can further regularize the reconstruction of plane regions without any other ground truth annotations. Specifically, we introduce an iterative training scheme, where (i) grouping of pixels to formulate a super-plane (analogous to super-pixels), and (ii) optimizing of the scene reconstruction network via a super-plane constraint, are progressively conducted. We demonstrate that the model trained with super-planes surprisingly outperforms the one using conventional annotated planes, as individual super-plane statistically occupies a larger area and leads to more stable training. Extensive experiments show that our self-supervised super-plane constraint significantly improves 3D reconstruction quality even better than using ground truth plane segmentation. Additionally, the plane reconstruction results from our model can be used for auto-labeling for other vision tasks. The code and models are available at https: //github.com/botaoye/S3PRecon.

Multiple near frontal-parallel planes based depth representation demonstrated impressive results in self-supervised monocular depth estimation (MDE). Whereas, such a representation would cause the discontinuity of the ground as it is perpendicular to the frontal-parallel planes, which is detrimental to the identification of drivable space in autonomous driving. In this paper, we propose the PlaneDepth, a novel orthogonal planes based presentation, including vertical planes and ground planes. PlaneDepth estimates the depth distribution using a Laplacian Mixture Model based on orthogonal planes for an input image. These planes are used to synthesize a reference view to provide the self-supervision signal. Further, we find that the widely used resizing and cropping data augmentation breaks the orthogonality assumptions, leading to inferior plane predictions. We address this problem by explicitly constructing the resizing cropping transformation to rectify the predefined planes and predicted camera pose. Moreover, we propose an augmented self-distillation loss supervised with a bilateral occlusion mask to boost the robustness of orthogonal planes representation for occlusions. Thanks to our orthogonal planes representation, we can extract the ground plane in an unsupervised manner, which is important for autonomous driving. Extensive experiments on the KITTI dataset demonstrate the effectiveness and efficiency of our method. The code is available at https://github.com/svip-lab/PlaneDepth.

In this paper, we propose a single image scale estimation method based on a novel scale field representation. A scale field defines the local pixel-to-metric conversion ratio along the gravity direction on all the ground pixels. This representation resolves the ambiguity in camera parameters, allowing us to use a simple yet effective way to collect scale annotations on arbitrary images from human annotators. By training our model on calibrated panoramic image data and the in-the-wild human annotated data, our single image scene scale estimation network generates robust scale field on a variety of image, which can be utilized in various 3D understanding and scale-aware image editing applications.

In contrast to sparse keypoints, a handful of line segments can concisely encode the high-level scene layout, as they often delineate the main structural elements. In addition to offering strong geometric cues, they are also omnipresent in urban landscapes and indoor scenes. Despite their apparent advantages, current line-based reconstruction methods are far behind their point-based counterparts. In this paper we aim to close the gap by introducing LIMAP, a library for 3D line mapping that robustly and efficiently creates 3D line maps from multi-view imagery. This is achieved through revisiting the degeneracy problem of line triangulation, carefully crafted scoring and track building, and exploiting structural priors such as line coincidence, parallelism, and orthogonality. Our code integrates seamlessly with existing point-based Structure-from-Motion methods and can leverage their 3D points to further improve the line reconstruction. Furthermore, as a byproduct, the method is able to recover 3D association graphs between lines and points / vanishing points (VPs). In thorough experiments, we show that LIMAP significantly outperforms existing approaches for 3D line mapping. Our robust 3D line maps also open up new research directions. We show two example applications: visual localization and bundle adjustment, where integrating lines alongside points yields the best results. Code is available at https://github.com/cvg/limap.

Representing visual signals with implicit coordinate-based neural networks, as an effective replacement of the traditional discrete signal representation, has gained considerable popularity in computer vision and graphics. In contrast to existing implicit neural representations which focus on modelling the scene only, this paper proposes a novel implicit camera model which represents the physical imaging process of a camera as a deep neural network. We demonstrate the power of this new implicit camera model on two inverse imaging tasks: i) generating all-in-focus photos, and ii) HDR imaging. Specifically, we devise an implicit blur generator and an implicit tone mapper to model the aperture and exposure of the camera’s imaging process, respectively. Our implicit camera model is jointly learned together with implicit scene models under multi-focus stack and multi-exposure bracket supervision. We have demonstrated the effectiveness of our new model on large number of test images and videos, producing accurate and visually appealing all-in-focus and high dynamic range images. In principle, our new implicit neural camera model has the potential to benefit a wide array of other inverse imaging tasks.

Estimating a dense depth map from a single view is geometrically ill-posed, and state-of-the-art methods rely on learning depth’s relation with visual appearance using deep neural networks. On the other hand, Structure from Motion (SfM) leverages multi-view constraints to produce very accurate but sparse maps, as matching across images is typically limited by locally discriminative texture. In this work, we combine the strengths of both approaches by proposing a novel test-time refinement (TTR) method, denoted as SfM-TTR, that boosts the performance of single-view depth networks at test time using SfM multi-view cues. Specifically, and differently from the state of the art, we use sparse SfM point clouds as test-time self-supervisory signal, fine-tuning the network encoder to learn a better representation of the test scene. Our results show how the addition of SfM-TTR to several state-of-the-art self-supervised and supervised networks improves significantly their performance, outperforming previous TTR baselines mainly based on photometric multi-view consistency. The code is available at https://github.com/serizba/SfM-TTR.

Monocular depth estimation is fundamental for 3D scene understanding and downstream applications. However, even under the supervised setup, it is still challenging and ill-posed due to the lack of geometric constraints. We observe that although a scene can consist of millions of pixels, there are much fewer high-level patterns. We propose iDisc to learn those patterns with internal discretized representations. The method implicitly partitions the scene into a set of high-level concepts. In particular, our new module, Internal Discretization (ID), implements a continuous-discrete-continuous bottleneck to learn those concepts without supervision. In contrast to state-of-the-art methods, the proposed model does not enforce any explicit constraints or priors on the depth output. The whole network with the ID module can be trained in an end-to-end fashion thanks to the bottleneck module based on attention. Our method sets the new state of the art with significant improvements on NYU-Depth v2 and KITTI, outperforming all published methods on the official KITTI benchmark. iDisc can also achieve state-of-the-art results on surface normal estimation. Further, we explore the model generalization capability via zero-shot testing. From there, we observe the compelling need to promote diversification in the outdoor scenario and we introduce splits of two autonomous driving datasets, DDAD and Argoverse. Code is available at http://vis.xyz/pub/idisc/.

Smartphone cameras today are increasingly approaching the versatility and quality of professional cameras through a combination of hardware and software advancements. However, fixed aperture remains a key limitation, preventing users from controlling the depth of field (DoF) of captured images. At the same time, many smartphones now have multiple cameras with different fixed apertures - specifically, an ultra-wide camera with wider field of view and deeper DoF and a higher resolution primary camera with shallower DoF. In this work, we propose DC^2, a system for defocus control for synthetically varying camera aperture, focus distance and arbitrary defocus effects by fusing information from such a dual-camera system. Our key insight is to leverage real-world smartphone camera dataset by using image refocus as a proxy task for learning to control defocus. Quantitative and qualitative evaluations on real-world data demonstrate our system’s efficacy where we outperform state-of-the-art on defocus deblurring, bokeh rendering, and image refocus. Finally, we demonstrate creative post-capture defocus control enabled by our method, including tilt-shift and content-based defocus effects.

We present the design of a productionized end-to-end stereo depth sensing system that does pre-processing, online stereo rectification, and stereo depth estimation with a fallback to monocular depth estimation when rectification is unreliable. The output of our depth sensing system is then used in a novel view generation pipeline to create 3D computational photography effects using point-of-view images captured by smart glasses. All these steps are executed on-device on the stringent compute budget of a mobile phone, and because we expect the users can use a wide range of smartphones, our design needs to be general and cannot be dependent on a particular hardware or ML accelerator such as a smartphone GPU. Although each of these steps is well studied, a description of a practical system is still lacking. For such a system, all these steps need to work in tandem with one another and fallback gracefully on failures within the system or less than ideal input data. We show how we handle unforeseen changes to calibration, e.g., due to heat, robustly support depth estimation in the wild, and still abide by the memory and latency constraints required for a smooth user experience. We show that our trained models are fast, and run in less than 1s on a six-year-old Samsung Galaxy S8 phone’s CPU. Our models generalize well to unseen data and achieve good results on Middlebury and in-the-wild images captured from the smart glasses.

Recent cascade Multi-View Stereo (MVS) methods can efficiently estimate high-resolution depth maps through narrowing hypothesis ranges. However, previous methods ignored the vital geometric information embedded in coarse stages, leading to vulnerable cost matching and sub-optimal reconstruction results. In this paper, we propose a geometry awareness model, termed GeoMVSNet, to explicitly integrate geometric clues implied in coarse stages for delicate depth estimation. In particular, we design a two-branch geometry fusion network to extract geometric priors from coarse estimations to enhance structural feature extraction at finer stages. Besides, we embed the coarse probability volumes, which encode valuable depth distribution attributes, into the lightweight regularization network to further strengthen depth-wise geometry intuition. Meanwhile, we apply the frequency domain filtering to mitigate the negative impact of the high-frequency regions and adopt the curriculum learning strategy to progressively boost the geometry integration of the model. To intensify the full-scene geometry perception of our model, we present the depth distribution similarity loss based on the Gaussian-Mixture Model assumption. Extensive experiments on DTU and Tanks and Temples (T&T) datasets demonstrate that our GeoMVSNet achieves state-of-the-art results and ranks first on the T&T-Advanced set. Code is available at https://github.com/doubleZ0108/GeoMVSNet.

With the rapid development of virtual reality, 360° images have gained increasing popularity. Their wide field of view necessitates high resolution to ensure image quality. This, however, makes it harder to acquire, store and even process such 360° images. To alleviate this issue, we propose the first attempt at 360° image rescaling, which refers to downscaling a 360° image to a visually valid low-resolution (LR) counterpart and then upscaling to a high-resolution (HR) 360° image given the LR variant. Specifically, we first analyze two 360° image datasets and observe several findings that characterize how 360° images typically change along their latitudes. Inspired by these findings, we propose a novel deformable invertible neural network (INN), named DINN360, for latitude-aware 360° image rescaling. In DINN360, a deformable INN is designed to downscale the LR image, and project the high-frequency (HF) component to the latent space by adaptively handling various deformations occurring at different latitude regions. Given the downscaled LR image, the high-quality HR image is then reconstructed in a conditional latitude-aware manner by recovering the structure-related HF component from the latent space. Extensive experiments over four public datasets show that our DINN360 method performs considerably better than other state-of-the-art methods for 2×, 4× and 8× 360° image rescaling.

Estimating depth from four large field of view (FoV) cameras has been a difficult and understudied problem. In this paper, we proposed a novel and simple system that can convert this difficult problem into easier binocular depth estimation. We name this system OmniVidar, as its results are similar to LiDAR, but rely only on vision. OmniVidar contains three components: (1) a new camera model to address the shortcomings of existing models, (2) a new multi-fisheye camera based epipolar rectification method for solving the image distortion and simplifying the depth estimation problem, (3) an improved binocular depth estimation network, which achieves a better balance between accuracy and efficiency. Unlike other omnidirectional stereo vision methods, OmniVidar does not contain any 3D convolution, so it can achieve higher resolution depth estimation at fast speed. Results demonstrate that OmniVidar outperforms all other methods in terms of accuracy and performance.

Multi-frame depth estimation generally achieves high accuracy relying on the multi-view geometric consistency. When applied in dynamic scenes, e.g., autonomous driving, this consistency is usually violated in the dynamic areas, leading to corrupted estimations. Many multi-frame methods handle dynamic areas by identifying them with explicit masks and compensating the multi-view cues with monocular cues represented as local monocular depth or features. The improvements are limited due to the uncontrolled quality of the masks and the underutilized benefits of the fusion of the two types of cues. In this paper, we propose a novel method to learn to fuse the multi-view and monocular cues encoded as volumes without needing the heuristically crafted masks. As unveiled in our analyses, the multi-view cues capture more accurate geometric information in static areas, and the monocular cues capture more useful contexts in dynamic areas. To let the geometric perception learned from multi-view cues in static areas propagate to the monocular representation in dynamic areas and let monocular cues enhance the representation of multi-view cost volume, we propose a cross-cue fusion (CCF) module, which includes the cross-cue attention (CCA) to encode the spatially non-local relative intra-relations from each source to enhance the representation of the other. Experiments on real-world datasets prove the significant effectiveness and generalization ability of the proposed method.

Effectively localizing an agent in a realistic, noisy setting is crucial for many embodied vision tasks. Visual Odometry (VO) is a practical substitute for unreliable GPS and compass sensors, especially in indoor environments. While SLAM-based methods show a solid performance without large data requirements, they are less flexible and robust w.r.t. to noise and changes in the sensor suite compared to learning-based approaches. Recent deep VO models, however, limit themselves to a fixed set of input modalities, e.g., RGB and depth, while training on millions of samples. When sensors fail, sensor suites change, or modalities are intentionally looped out due to available resources, e.g., power consumption, the models fail catastrophically. Furthermore, training these models from scratch is even more expensive without simulator access or suitable existing models that can be fine-tuned. While such scenarios get mostly ignored in simulation, they commonly hinder a model’s reusability in real-world applications. We propose a Transformer-based modality-invariant VO approach that can deal with diverse or changing sensor suites of navigation agents. Our model outperforms previous methods while training on only a fraction of the data. We hope this method opens the door to a broader range of real-world applications that can benefit from flexible and learned VO models.

The task of 3D semantic scene graph (3DSSG) prediction in the point cloud is challenging since (1) the 3D point cloud only captures geometric structures with limited semantics compared to 2D images, and (2) long-tailed relation distribution inherently hinders the learning of unbiased prediction. Since 2D images provide rich semantics and scene graphs are in nature coped with languages, in this study, we propose Visual-Linguistic Semantics Assisted Training (VL-SAT) scheme that can significantly empower 3DSSG prediction models with discrimination about long-tailed and ambiguous semantic relations. The key idea is to train a powerful multi-modal oracle model to assist the 3D model. This oracle learns reliable structural representations based on semantics from vision, language, and 3D geometry, and its benefits can be heterogeneously passed to the 3D model during the training stage. By effectively utilizing visual-linguistic semantics in training, our VL-SAT can significantly boost common 3DSSG prediction models, such as SGFN and SGGpoint, only with 3D inputs in the inference stage, especially when dealing with tail relation triplets. Comprehensive evaluations and ablation studies on the 3DSSG dataset have validated the effectiveness of the proposed scheme. Code is available at https://github.com/wz7in/CVPR2023-VLSAT.

In this paper, we address the problem of detecting 3D objects from multi-view images. Current query-based methods rely on global 3D position embeddings (PE) to learn the geometric correspondence between images and 3D space. We claim that directly interacting 2D image features with global 3D PE could increase the difficulty of learning view transformation due to the variation of camera extrinsics. Thus we propose a novel method based on CAmera view Position Embedding, called CAPE. We form the 3D position embeddings under the local camera-view coordinate system instead of the global coordinate system, such that 3D position embedding is free of encoding camera extrinsic parameters. Furthermore, we extend our CAPE to temporal modeling by exploiting the object queries of previous frames and encoding the ego motion for boosting 3D object detection. CAPE achieves the state-of-the-art performance (61.0% NDS and 52.5% mAP) among all LiDAR-free methods on standard nuScenes dataset. Codes and models are available.

Recent LSS-based multi-view 3D object detection has made tremendous progress, by processing the features in Brid-Eye-View (BEV) via the convolutional detector. However, the typical convolution ignores the radial symmetry of the BEV features and increases the difficulty of the detector optimization. To preserve the inherent property of the BEV features and ease the optimization, we propose an azimuth-equivariant convolution (AeConv) and an azimuth-equivariant anchor. The sampling grid of AeConv is always in the radial direction, thus it can learn azimuth-invariant BEV features. The proposed anchor enables the detection head to learn predicting azimuth-irrelevant targets. In addition, we introduce a camera-decoupled virtual depth to unify the depth prediction for the images with different camera intrinsic parameters. The resultant detector is dubbed Azimuth-equivariant Detector (AeDet). Extensive experiments are conducted on nuScenes, and AeDet achieves a 62.0% NDS, surpassing the recent multi-view 3D object detectors such as PETRv2 and BEVDepth by a large margin.

This paper proposes a novel method to improve the performance of a trained object detector on scenes with fixed camera perspectives based on self-supervised adaptation. Given a specific scene, the trained detector is adapted using pseudo-ground truth labels generated by the detector itself and an object tracker in a cross-teaching manner. When the camera perspective is fixed, our method can utilize the background equivariance by proposing artifact-free object mixup as a means of data augmentation, and utilize accurate background extraction as an additional input modality. We also introduce a large-scale and diverse dataset for the development and evaluation of scene-adaptive object detection. Experiments on this dataset show that our method can improve the average precision of the original detector, outperforming the previous state-of-the-art self-supervised domain adaptive object detection methods by a large margin. Our dataset and code are published at https://github.com/cvlab-stonybrook/scenes100.

3D object detection is an essential perception task in autonomous driving to understand the environments. The Bird’s-Eye-View (BEV) representations have significantly improved the performance of 3D detectors with camera inputs on popular benchmarks. However, there still lacks a systematic understanding of the robustness of these vision-dependent BEV models, which is closely related to the safety of autonomous driving systems. In this paper, we evaluate the natural and adversarial robustness of various representative models under extensive settings, to fully understand their behaviors influenced by explicit BEV features compared with those without BEV. In addition to the classic settings, we propose a 3D consistent patch attack by applying adversarial patches in the 3D space to guarantee the spatiotemporal consistency, which is more realistic for the scenario of autonomous driving. With substantial experiments, we draw several findings: 1) BEV models tend to be more stable than previous methods under different natural conditions and common corruptions due to the expressive spatial representations; 2) BEV models are more vulnerable to adversarial noises, mainly caused by the redundant BEV features; 3) Camera-LiDAR fusion models have superior performance under different settings with multi-modal inputs, but BEV fusion model is still vulnerable to adversarial noises of both point cloud and image. These findings alert the safety issue in the applications of BEV detectors and could facilitate the development of more robust models.

While most recent autonomous driving system focuses on developing perception methods on ego-vehicle sensors, people tend to overlook an alternative approach to leverage intelligent roadside cameras to extend the perception ability beyond the visual range. We discover that the state-of-the-art vision-centric bird’s eye view detection methods have inferior performances on roadside cameras. This is because these methods mainly focus on recovering the depth regarding the camera center, where the depth difference between the car and the ground quickly shrinks while the distance increases. In this paper, we propose a simple yet effective approach, dubbed BEVHeight, to address this issue. In essence, instead of predicting the pixel-wise depth, we regress the height to the ground to achieve a distance-agnostic formulation to ease the optimization process of camera-only perception methods. On popular 3D detection benchmarks of roadside cameras, our method surpasses all previous vision-centric methods by a significant margin. The code is available at https://github.com/ADLab-AutoDrive/BEVHeight.

This paper proposes a novel method for vision-based metric cross-view geolocalization (CVGL) that matches the camera images captured from a ground-based vehicle with an aerial image to determine the vehicle’s geo-pose. Since aerial images are globally available at low cost, they represent a potential compromise between two established paradigms of autonomous driving, i.e. using expensive high-definition prior maps or relying entirely on the sensor data captured at runtime. We present an end-to-end differentiable model that uses the ground and aerial images to predict a probability distribution over possible vehicle poses. We combine multiple vehicle datasets with aerial images from orthophoto providers on which we demonstrate the feasibility of our method. Since the ground truth poses are often inaccurate w.r.t. the aerial images, we implement a pseudo-label approach to produce more accurate ground truth poses and make them publicly available. While previous works require training data from the target region to achieve reasonable localization accuracy (i.e. same-area evaluation), our approach overcomes this limitation and outperforms previous results even in the strictly more challenging cross-area case. We improve the previous state-of-the-art by a large margin even without ground or aerial data from the test region, which highlights the model’s potential for global-scale application. We further integrate the uncertainty-aware predictions in a tracking framework to determine the vehicle’s trajectory over time resulting in a mean position error on KITTI-360 of 0.78m.

Humans can orient themselves in their 3D environments using simple 2D maps. Differently, algorithms for visual localization mostly rely on complex 3D point clouds that are expensive to build, store, and maintain over time. We bridge this gap by introducing OrienterNet, the first deep neural network that can localize an image with sub-meter accuracy using the same 2D semantic maps that humans use. OrienterNet estimates the location and orientation of a query image by matching a neural Bird’s-Eye View with open and globally available maps from OpenStreetMap, enabling anyone to localize anywhere such maps are available. OrienterNet is supervised only by camera poses but learns to perform semantic matching with a wide range of map elements in an end-to-end manner. To enable this, we introduce a large crowd-sourced dataset of images captured across 12 cities from the diverse viewpoints of cars, bikes, and pedestrians. OrienterNet generalizes to new datasets and pushes the state of the art in both robotics and AR scenarios. The code is available at https://github.com/facebookresearch/OrienterNet

Fusing LiDAR and camera information is essential for accurate and reliable 3D object detection in autonomous driving systems. This is challenging due to the difficulty of combining multi-granularity geometric and semantic features from two drastically different modalities. Recent approaches aim at exploring the semantic densities of camera features through lifting points in 2D camera images (referred to as “seeds”) into 3D space, and then incorporate 2D semantics via cross-modal interaction or fusion techniques. However, depth information is under-investigated in these approaches when lifting points into 3D space, thus 2D semantics can not be reliably fused with 3D points. Moreover, their multi-modal fusion strategy, which is implemented as concatenation or attention, either can not effectively fuse 2D and 3D information or is unable to perform fine-grained interactions in the voxel space. To this end, we propose a novel framework with better utilization of the depth information and fine-grained cross-modal interaction between LiDAR and camera, which consists of two important components. First, a Multi-Depth Unprojection (MDU) method is used to enhance the depth quality of the lifted points at each interaction level. Second, a Gated Modality-Aware Convolution (GMA-Conv) block is applied to modulate voxels involved with the camera modality in a fine-grained manner and then aggregate multi-modal features into a unified space. Together they provide the detection head with more comprehensive features from LiDAR and camera. On the nuScenes test benchmark, our proposed method, abbreviated as MSMDFusion, achieves state-of-the-art results on both 3D object detection and tracking tasks without using test-time-augmentation and ensemble techniques. The code is available at https://github.com/SxJyJay/MSMDFusion.

Recently, virtual/pseudo-point-based 3D object detection that seamlessly fuses RGB images and LiDAR data by depth completion has gained great attention. However, virtual points generated from an image are very dense, introducing a huge amount of redundant computation during detection. Meanwhile, noises brought by inaccurate depth completion significantly degrade detection precision. This paper proposes a fast yet effective backbone, termed VirConvNet, based on a new operator VirConv (Virtual Sparse Convolution), for virtual-point-based 3D object detection. The VirConv consists of two key designs: (1) StVD (Stochastic Voxel Discard) and (2) NRConv (Noise-Resistant Submanifold Convolution). The StVD alleviates the computation problem by discarding large amounts of nearby redundant voxels. The NRConv tackles the noise problem by encoding voxel features in both 2D image and 3D LiDAR space. By integrating our VirConv, we first develop an efficient pipeline VirConv-L based on an early fusion design. Then, we build a high-precision pipeline VirConv-T based on a transformed refinement scheme. Finally, we develop a semi-supervised pipeline VirConv-S based on a pseudo-label framework. On the KITTI car 3D detection test leaderboard, our VirConv-L achieves 85% AP with a fast running speed of 56ms. Our VirConv-T and VirConv-S attains a high-precision of 86.3% and 87.2% AP, and currently rank 2nd and 1st, respectively. The code is available at https://github.com/hailanyi/VirConv.

The accuracy of crowd counting in images has improved greatly in recent years due to the development of deep neural networks for predicting crowd density maps. However, most methods do not further explore the ability to localize people in the density map, with those few works adopting simple methods, like finding the local peaks in the density map. In this paper, we propose the optimal transport minimization (OT-M) algorithm for crowd localization with density maps. The objective of OT-M is to find a target point map that has the minimal Sinkhorn distance with the input density map, and we propose an iterative algorithm to compute the solution. We then apply OT-M to generate hard pseudo-labels (point maps) for semi-supervised counting, rather than the soft pseudo-labels (density maps) used in previous methods. Our hard pseudo-labels provide stronger supervision, and also enable the use of recent density-to-point loss functions for training. We also propose a confidence weighting strategy to give higher weight to the more reliable unlabeled data. Extensive experiments show that our methods achieve outstanding performance on both crowd localization and semi-supervised counting. Code is available at https://github.com/Elin24/OT-M.

3D object detectors usually rely on hand-crafted proxies, e.g., anchors or centers, and translate well-studied 2D frameworks to 3D. Thus, sparse voxel features need to be densified and processed by dense prediction heads, which inevitably costs extra computation. In this paper, we instead propose VoxelNext for fully sparse 3D object detection. Our core insight is to predict objects directly based on sparse voxel features, without relying on hand-crafted proxies. Our strong sparse convolutional network VoxelNeXt detects and tracks 3D objects through voxel features entirely. It is an elegant and efficient framework, with no need for sparse-to-dense conversion or NMS post-processing. Our method achieves a better speed-accuracy trade-off than other mainframe detectors on the nuScenes dataset. For the first time, we show that a fully sparse voxel-based representation works decently for LIDAR 3D object detection and tracking. Extensive experiments on nuScenes, Waymo, and Argoverse2 benchmarks validate the effectiveness of our approach. Without bells and whistles, our model outperforms all existing LIDAR methods on the nuScenes tracking test benchmark.

3D object detection within large 3D scenes is challenging not only due to the sparse and irregular 3D point clouds, but also due to both the extreme foreground-background scene imbalance and class imbalance. A common approach is to add ground-truth objects from other scenes. Differently, we propose to modify the scenes by removing elements (voxels), rather than adding ones. Our approach selects the “meaningful” voxels, in a manner that addresses both types of dataset imbalance. The approach is general and can be applied to any voxel-based detector, yet the meaningfulness of a voxel is network-dependent. Our voxel selection is shown to improve the performance of several prominent 3D detection methods.

LiDAR and camera are two modalities available for 3D semantic segmentation in autonomous driving. The popular LiDAR-only methods severely suffer from inferior segmentation on small and distant objects due to insufficient laser points, while the robust multi-modal solution is under-explored, where we investigate three crucial inherent difficulties: modality heterogeneity, limited sensor field of view intersection, and multi-modal data augmentation. We propose a multi-modal 3D semantic segmentation model (MSeg3D) with joint intra-modal feature extraction and inter-modal feature fusion to mitigate the modality heterogeneity. The multi-modal fusion in MSeg3D consists of geometry-based feature fusion GF-Phase, cross-modal feature completion, and semantic-based feature fusion SF-Phase on all visible points. The multi-modal data augmentation is reinvigorated by applying asymmetric transformations on LiDAR point cloud and multi-camera images individually, which benefits the model training with diversified augmentation transformations. MSeg3D achieves state-of-the-art results on nuScenes, Waymo, and SemanticKITTI datasets. Under the malfunctioning multi-camera input and the multi-frame point clouds input, MSeg3D still shows robustness and improves the LiDAR-only baseline. Our code is publicly available at https://github.com/jialeli1/lidarseg3d.

Densely annotating LiDAR point clouds is costly, which often restrains the scalability of fully-supervised learning methods. In this work, we study the underexplored semi-supervised learning (SSL) in LiDAR semantic segmentation. Our core idea is to leverage the strong spatial cues of LiDAR point clouds to better exploit unlabeled data. We propose LaserMix to mix laser beams from different LiDAR scans and then encourage the model to make consistent and confident predictions before and after mixing. Our framework has three appealing properties. 1) Generic: LaserMix is agnostic to LiDAR representations (e.g., range view and voxel), and hence our SSL framework can be universally applied. 2) Statistically grounded: We provide a detailed analysis to theoretically explain the applicability of the proposed framework. 3) Effective: Comprehensive experimental analysis on popular LiDAR segmentation datasets (nuScenes, SemanticKITTI, and ScribbleKITTI) demonstrates our effectiveness and superiority. Notably, we achieve competitive results over fully-supervised counterparts with 2x to 5x fewer labels and improve the supervised-only baseline significantly by relatively 10.8%. We hope this concise yet high-performing framework could facilitate future research in semi-supervised LiDAR segmentation. Code is publicly available.

Self-supervised pretraining on large unlabeled datasets has shown tremendous success on improving the task performance of many computer vision tasks. However, such techniques have not been widely used for outdoor LiDAR point cloud perception due to its scene complexity and wide range. This prevents impactful application from 2D pretraining frameworks. In this paper, we propose ISCC, a new self-supervised pretraining method, core of which are two pretext tasks newly designed for LiDAR point clouds. The first task focuses on learning semantic information by sorting local groups of points in the scene into a globally consistent set of semantically meaningful clusters using contrastive learning. This is augmented with a second task which reasons about precise surfaces of various parts of the scene through implicit surface reconstruction to learn geometric structures. We demonstrate their effectiveness on transfer learning performance on 3D object detection and semantic segmentation in real world LiDAR scenes. We further design an unsupervised semantic grouping task to showcase the highly semantically meaningful features learned by our approach.

Given a 3D object, kinematic motion prediction aims to identify the mobile parts as well as the corresponding motion parameters. Due to the large variations in both topological structure and geometric details of 3D objects, this remains a challenging task and the lack of large scale labeled data also constrain the performance of deep learning based approaches. In this paper, we tackle the task of object kinematic motion prediction problem in a semi-weakly supervised manner. Our key observations are two-fold. First, although 3D dataset with fully annotated motion labels is limited, there are existing datasets and methods for object part semantic segmentation at large scale. Second, semantic part segmentation and mobile part segmentation is not always consistent but it is possible to detect the mobile parts from the underlying 3D structure. Towards this end, we propose a graph neural network to learn the map between hierarchical part-level segmentation and mobile parts parameters, which are further refined based on geometric alignment. This network can be first trained on PartNet-Mobility dataset with fully labeled mobility information and then applied on PartNet dataset with fine-grained and hierarchical part-level segmentation. The network predictions yield a large scale of 3D objects with pseudo labeled mobility information and can further be used for weakly-supervised learning with pre-existing segmentation. Our experiments show there are significant performance boosts with the augmented data for previous method designed for kinematic motion prediction on 3D partial scans.

Generalizable 3D part segmentation is important but challenging in vision and robotics. Training deep models via conventional supervised methods requires large-scale 3D datasets with fine-grained part annotations, which are costly to collect. This paper explores an alternative way for low-shot part segmentation of 3D point clouds by leveraging a pretrained image-language model, GLIP, which achieves superior performance on open-vocabulary 2D detection. We transfer the rich knowledge from 2D to 3D through GLIP-based part detection on point cloud rendering and a novel 2D-to-3D label lifting algorithm. We also utilize multi-view 3D priors and few-shot prompt tuning to boost performance significantly. Extensive evaluation on PartNet and PartNet-Mobility datasets shows that our method enables excellent zero-shot 3D part segmentation. Our few-shot version not only outperforms existing few-shot approaches by a large margin but also achieves highly competitive results compared to the fully supervised counterpart. Furthermore, we demonstrate that our method can be directly applied to iPhone-scanned point clouds without significant domain gaps.

Image restoration under multiple adverse weather conditions aims to remove weather-related artifacts by using the single set of network parameters. In this paper, we find that distorted images under different weather conditions contain general characteristics as well as their specific characteristics. Inspired by this observation, we design an efficient unified framework with a two-stage training strategy to explore the weather-general and weather-specific features. The first training stage aims to learn the weather-general features by taking the images under various weather conditions as the inputs and outputting the coarsely restored results. The second training stage aims to learn to adaptively expand the specific parameters for each weather type in the deep model, where requisite positions for expansion of weather-specific parameters are learned automatically. Hence, we can obtain an efficient and unified model for image restoration under multiple adverse weather conditions. Moreover, we build the first real-world benchmark dataset with multiple weather conditions to better deal with real-world weather scenarios. Experimental results show that our method achieves superior performance on all the synthetic and real-world benchmark datasets.

Despite the significant recent progress made on 3D point cloud semantic segmentation, the current methods require training data for all classes at once, and are not suitable for real-life scenarios where new categories are being continuously discovered. Substantial memory storage and expensive re-training is required to update the model to sequentially arriving data for new concepts. In this paper, to continually learn new categories using previous knowledge, we introduce class-incremental semantic segmentation of 3D point cloud. Unlike 2D images, 3D point clouds are disordered and unstructured, making it difficult to store and transfer knowledge especially when the previous data is not available. We further face the challenge of semantic shift, where previous/future classes are indiscriminately collapsed and treated as the background in the current step, causing a dramatic performance drop on past classes. We exploit the structure of point cloud and propose two strategies to address these challenges. First, we design a geometry-aware distillation module that transfers point-wise feature associations in terms of their geometric characteristics. To counter forgetting caused by the semantic shift, we further develop an uncertainty-aware pseudo-labelling scheme that eliminates noise in uncertain pseudo-labels by label propagation within a local neighborhood. Our extensive experiments on S3DIS and ScanNet in a class-incremental setting show impressive results comparable to the joint training strategy (upper bound). Code is available at: https://github.com/leolyj/3DPC-CISS

Pre-training by numerous image data has become de-facto for robust 2D representations. In contrast, due to the expensive data processing, a paucity of 3D datasets severely hinders the learning for high-quality 3D features. In this paper, we propose an alternative to obtain superior 3D representations from 2D pre-trained models via Image-to-Point Masked Autoencoders, named as I2P-MAE. By self-supervised pre-training, we leverage the well learned 2D knowledge to guide 3D masked autoencoding, which reconstructs the masked point tokens with an encoder-decoder architecture. Specifically, we first utilize off-the-shelf 2D models to extract the multi-view visual features of the input point cloud, and then conduct two types of image-to-point learning schemes. For one, we introduce a 2D-guided masking strategy that maintains semantically important point tokens to be visible. Compared to random masking, the network can better concentrate on significant 3D structures with key spatial cues. For another, we enforce these visible tokens to reconstruct multi-view 2D features after the decoder. This enables the network to effectively inherit high-level 2D semantics for discriminative 3D modeling. Aided by our image-to-point pre-training, the frozen I2P-MAE, without any fine-tuning, achieves 93.4% accuracy for linear SVM on ModelNet40, competitive to existing fully trained methods. By further fine-tuning on on ScanObjectNN’s hardest split, I2P-MAE attains the state-of-the-art 90.11% accuracy, +3.68% to the second-best, demonstrating superior transferable capacity. Code is available at https://github.com/ZrrSkywalker/I2P-MAE.

Recent years have witnessed significant developments in point cloud processing, including classification and segmentation. However, supervised learning approaches need a lot of well-labeled data for training, and annotation is labor- and time-intensive. Self-supervised learning, on the other hand, uses unlabeled data, and pre-trains a backbone with a pretext task to extract latent representations to be used with the downstream tasks. Compared to 2D images, self-supervised learning of 3D point clouds is under-explored. Existing models, for self-supervised learning of 3D point clouds, rely on a large number of data samples, and require significant amount of computational resources and training time. To address this issue, we propose a novel contrastive learning approach, referred to as ToThePoint. Different from traditional contrastive learning methods, which maximize agreement between features obtained from a pair of point clouds formed only with different types of augmentation, ToThePoint also maximizes the agreement between the permutation invariant features and features discarded after max pooling. We first perform self-supervised learning on the ShapeNet dataset, and then evaluate the performance of the network on different downstream tasks. In the downstream task experiments, performed on the ModelNet40, ModelNet40C, ScanobjectNN and ShapeNet-Part datasets, our proposed ToThePoint achieves competitive, if not better results compared to the state-of-the-art baselines, and does so with significantly less training time (200 times faster than baselines)

The remarkable breakthroughs in point cloud representation learning have boosted their usage in real-world applications such as self-driving cars and virtual reality. However, these applications usually have an urgent requirement for not only accurate but also efficient 3D object detection. Recently, knowledge distillation has been proposed as an effective model compression technique, which transfers the knowledge from an over-parameterized teacher to a lightweight student and achieves consistent effectiveness in 2D vision. However, due to point clouds’ sparsity and irregularity, directly applying previous image-based knowledge distillation methods to point cloud detectors usually leads to unsatisfactory performance. To fill the gap, this paper proposes PointDistiller, a structured knowledge distillation framework for point clouds-based 3D detection. Concretely, PointDistiller includes local distillation which extracts and distills the local geometric structure of point clouds with dynamic graph convolution and reweighted learning strategy, which highlights student learning on the critical points or voxels to improve knowledge distillation efficiency. Extensive experiments on both voxels-based and raw points-based detectors have demonstrated the effectiveness of our method over seven previous knowledge distillation methods. For instance, our 4X compressed PointPillars student achieves 2.8 and 3.4 mAP improvements on BEV and 3D object detection, outperforming its teacher by 0.9 and 1.8 mAP, respectively. Codes are available in the supplementary material and will be released on Github.

We introduce PointConvFormer, a novel building block for point cloud based deep network architectures. Inspired by generalization theory, PointConvFormer combines ideas from point convolution, where filter weights are only based on relative position, and Transformers which utilize feature-based attention. In PointConvFormer, attention computed from feature difference between points in the neighborhood is used to modify the convolutional weights at each point. Hence, we preserved the invariances from point convolution, whereas attention helps to select relevant points in the neighborhood for convolution. We experiment on both semantic segmentation and scene flow estimation tasks on point clouds with multiple datasets including ScanNet, SemanticKitti, FlyingThings3D and KITTI. Our results show that PointConvFormer substantially outperforms classic convolutions, regular transformers, and voxelized sparse convolution approaches with much smaller and faster networks. Visualizations show that PointConvFormer performs similarly to convolution on flat areas, whereas the neighborhood selection effect is stronger on object boundaries, showing that it has got the best of both worlds. The code will be available with the final version.

Transformers have shown superior performance on various computer vision tasks with their capabilities to capture long-range dependencies. Despite the success, it is challenging to directly apply Transformers on point clouds due to their quadratic cost in the number of points. In this paper, we present a Self-Positioning point-based Transformer (SPoTr), which is designed to capture both local and global shape contexts with reduced complexity. Specifically, this architecture consists of local self- attention and self-positioning point-based global cross-attention. The self-positioning points, adaptively located based on the input shape, consider both spatial and semantic information with disentangled attention to improve expressive power. With the self-positioning points, we propose a novel global cross-attention mechanism for point clouds, which improves the scalability of global self-attention by allowing the attention module to compute attention weights with only a small set of self-positioning points. Experiments show the effectiveness of SPoTr on three point cloud tasks such as shape classification, part segmentation, and scene segmentation. In particular, our proposed model achieves an accuracy gain of 2.6% over the previous best models on shape classification with ScanObjectNN. We also provide qualitative analyses to demonstrate the interpretability of self-positioning points. The code of SPoTr is available at https://github.com/mlvlab/SPoTr.

Feature invariance under different data transformations, i.e., transformation invariance, can be regarded as a type of self-supervision for representation learning. In this paper, we present PointClustering, a new unsupervised representation learning scheme that leverages transformation invariance for point cloud pre-training. PointClustering formulates the pretext task as deep clustering and employs transformation invariance as an inductive bias, following the philosophy that common point cloud transformation will not change the geometric properties and semantics. Technically, PointClustering iteratively optimizes the feature clusters and backbone, and delves into the transformation invariance as learning regularization from two perspectives: point level and instance level. Point-level invariance learning maintains local geometric properties through gathering point features of one instance across transformations, while instance-level invariance learning further measures clusters over the entire dataset to explore semantics of instances. Our PointClustering is architecture-agnostic and readily applicable to MLP-based, CNN-based and Transformer-based backbones. We empirically demonstrate that the models pre-learnt on the ScanNet dataset by PointClustering provide superior performances on six benchmarks, across downstream tasks of classification and segmentation. More remarkably, PointClustering achieves an accuracy of 94.5% on ModelNet40 with Transformer backbone. Source code is available at https://github.com/FuchenUSTC/PointClustering.

As a primitive 3D data representation, point clouds are prevailing in 3D sensing, yet short of intrinsic structural information of the underlying objects. Such discrepancy poses great challenges in directly establishing correspondences between point clouds sampled from deformable shapes. In light of this, we propose Neural Intrinsic Embedding (NIE) to embed each vertex into a high-dimensional space in a way that respects the intrinsic structure. Based upon NIE, we further present a weakly-supervised learning framework for non-rigid point cloud registration. Unlike the prior works, we do not require expansive and sensitive off-line basis construction (e.g., eigen-decomposition of Laplacians), nor do we require ground-truth correspondence labels for supervision. We empirically show that our framework performs on par with or even better than the state-of-the-art baselines, which generally require more supervision and/or more structural geometric input.

Parsing an unstructured point set into constituent local geometry structures (e.g., edges or surfaces) would be helpful for understanding and representing point clouds. This motivates us to design a deep architecture to model the hierarchical geometry from points, edges, surfaces (triangles), to super-surfaces (adjacent surfaces) for the thorough analysis of point clouds. In this paper, we present a novel Hierarchical Geometry Network (HGNet) that integrates such hierarchical geometry structures from super-surfaces, surfaces, edges, to points in a top-down manner for learning point cloud representations. Technically, we first construct the edges between every two neighbor points. A point-level representation is learnt with edge-to-point aggregation, i.e., aggregating all connected edges into the anchor point. Next, as every two neighbor edges compose a surface, we obtain the edge-level representation of each anchor edge via surface-to-edge aggregation over all neighbor surfaces. Furthermore, the surface-level representation is achieved through super-surface-to-surface aggregation by transforming all super-surfaces into the anchor surface. A Transformer structure is finally devised to unify all the point-level, edge-level, and surface-level features into the holistic point cloud representations. Extensive experiments on four point cloud analysis datasets demonstrate the superiority of HGNet for 3D object classification and part/semantic segmentation tasks. More remarkably, HGNet achieves the overall accuracy of 89.2% on ScanObjectNN, improving PointNeXt-S by 1.5%.

Deep Implicit Function (DIF) has gained much popularity as an efficient 3D shape representation. To capture geometry details, current mainstream methods divide 3D shapes into local regions and then learn each one with a local latent code via a decoder, where the decoder shares the geometric similarities among different local regions. Although such local methods can capture more local details, a large diversity of different local regions increases the difficulty of learning an implicit function when treating all regions equally using only a single decoder. In addition, these local regions often exhibit imbalanced distributions, where certain regions have significantly fewer observations. This leads that fine geometry details could not be preserved well. To solve this problem, we propose a novel Local Pattern-specific Implicit Function, named LP-DIF, for representing a shape with some clusters of local regions and multiple decoders, where each decoder only focuses on one cluster of local regions which share a certain pattern. Specifically, we first extract local codes for all regions, and then cluster them into multiple groups in the latent space, where similar regions sharing a common pattern fall into one group. After that, we train multiple decoders for mining local patterns of different groups, which simplifies learning of fine geometric details by reducing the diversity of local regions seen by each decoder. To further alleviate the data-imbalance problem, we introduce a region re-weighting module to each pattern-specific decoder by kernel density estimator, which dynamically re-weights the regions during learning. Our LP-DIF can restore more geometry details, and thus improve the quality of 3D reconstruction. Experiments demonstrate that our method can achieve the state-of-the-art performance over previous methods. Code is available at https://github.com/gtyxyz/lpdif.

We consider the problem of finding a continuous and non-rigid matching between a 2D contour and a 3D mesh. While such problems can be solved to global optimality by finding a shortest path in the product graph between both shapes, existing solutions heavily rely on unrealistic prior assumptions to avoid degenerate solutions (e.g. knowledge to which region of the 3D shape each point of the 2D contour is matched). To address this, we propose a novel 2D-3D shape matching formalism based on the conjugate product graph between the 2D contour and the 3D shape. Doing so allows us for the first time to consider higher-order costs, i.e. defined for edge chains, as opposed to costs defined for single edges. This offers substantially more flexibility, which we utilise to incorporate a local rigidity prior. By doing so, we effectively circumvent degenerate solutions and thereby obtain smoother and more realistic matchings, even when using only a one-dimensional feature descriptor. Overall, our method finds globally optimal and continuous 2D-3D matchings, has the same asymptotic complexity as previous solutions, produces state-of-the-art results for shape matching and is even capable of matching partial shapes. Our code is publicly available (https://github.com/paul0noah/sm-2D3D).

Recently, Multiple Object Tracking has achieved great success, which consists of object detection, feature embedding, and identity association. Existing methods apply the three-step or two-step paradigm to generate robust trajectories, where identity association is independent of other components. However, the independent identity association results in the identity-aware knowledge contained in the tracklet not be used to boost the detection and embedding modules. To overcome the limitations of existing methods, we introduce a novel Unified Tracking Model (UTM) to bridge those three components for generating a positive feedback loop with mutual benefits. The key insight of UTM is the Identity-Aware Feature Enhancement (IAFE), which is applied to bridge and benefit these three components by utilizing the identity-aware knowledge to boost detection and embedding. Formally, IAFE contains the Identity-Aware Boosting Attention (IABA) and the Identity-Aware Erasing Attention (IAEA), where IABA enhances the consistent regions between the current frame feature and identity-aware knowledge, and IAEA suppresses the distracted regions in the current frame feature. With better detections and embeddings, higher-quality tracklets can also be generated. Extensive experiments of public and private detections on three benchmarks demonstrate the robustness of UTM.

Extracting discriminative local features that are invariant to imaging variations is an integral part of establishing correspondences between images. In this work, we introduce a self-supervised learning framework to extract discriminative rotation-invariant descriptors using group-equivariant CNNs. Thanks to employing group-equivariant CNNs, our method effectively learns to obtain rotation-equivariant features and their orientations explicitly, without having to perform sophisticated data augmentations. The resultant features and their orientations are further processed by group aligning, a novel invariant mapping technique that shifts the group-equivariant features by their orientations along the group dimension. Our group aligning technique achieves rotation-invariance without any collapse of the group dimension and thus eschews loss of discriminability. The proposed method is trained end-to-end in a self-supervised manner, where we use an orientation alignment loss for the orientation estimation and a contrastive descriptor loss for robust local descriptors to geometric/photometric variations. Our method demonstrates state-of-the-art matching accuracy among existing rotation-invariant descriptors under varying rotation and also shows competitive results when transferred to the task of keypoint matching and camera pose estimation.

Local feature matching aims at finding correspondences between a pair of images. Although current detector-free methods leverage Transformer architecture to obtain an impressive performance, few works consider maintaining local consistency. Meanwhile, most methods struggle with large scale variations. To deal with the above issues, we propose Adaptive Spot-Guided Transformer (ASTR) for local feature matching, which jointly models the local consistency and scale variations in a unified coarse-to-fine architecture. The proposed ASTR enjoys several merits. First, we design a spot-guided aggregation module to avoid interfering with irrelevant areas during feature aggregation. Second, we design an adaptive scaling module to adjust the size of grids according to the calculated depth information at fine stage. Extensive experimental results on five standard benchmarks demonstrate that our ASTR performs favorably against state-of-the-art methods.Our code will be released on https://astr2023.github.io.

Dense geometric matching determines the dense pixel-wise correspondence between a source and support image corresponding to the same 3D structure. Prior works employ an encoder of transformer blocks to correlate the two-frame features. However, existing monocular pretraining tasks, e.g., image classification, and masked image modeling (MIM), can not pretrain the cross-frame module, yielding less optimal performance. To resolve this, we reformulate the MIM from reconstructing a single masked image to reconstructing a pair of masked images, enabling the pretraining of transformer module. Additionally, we incorporate a decoder into pretraining for improved upsampling results. Further, to be robust to the textureless area, we propose a novel cross-frame global matching module (CFGM). Since the most textureless area is planar surfaces, we propose a homography loss to further regularize its learning. Combined together, we achieve the State-of-The-Art (SoTA) performance on geometric matching. Codes and models are available at https://github.com/ShngJZ/PMatch.

Recurrent All-Pairs Field Transforms (RAFT) has shown great potentials in matching tasks. However, all-pairs correlations lack non-local geometry knowledge and have difficulties tackling local ambiguities in ill-posed regions. In this paper, we propose Iterative Geometry Encoding Volume (IGEV-Stereo), a new deep network architecture for stereo matching. The proposed IGEV-Stereo builds a combined geometry encoding volume that encodes geometry and context information as well as local matching details, and iteratively indexes it to update the disparity map. To speed up the convergence, we exploit GEV to regress an accurate starting point for ConvGRUs iterations. Our IGEV-Stereo ranks first on KITTI 2015 and 2012 (Reflective) among all published methods and is the fastest among the top 10 methods. In addition, IGEV-Stereo has strong cross-dataset generalization as well as high inference efficiency. We also extend our IGEV to multi-view stereo (MVS), i.e. IGEV-MVS, which achieves competitive accuracy on DTU benchmark. Code is available at https://github.com/gangweiX/IGEV.

Geometric estimation problems in vision are often solved via minimization of statistical loss functions which account for the presence of outliers in the observations. The corresponding energy landscape often has many local minima. Many approaches attempt to avoid local minima by annealing the scale parameter of loss functions using methods such as graduated non-convexity (GNC). However, little attention has been paid to the annealing schedule, which is often carried out in a fixed manner, resulting in a poor speed-accuracy trade-off and unreliable convergence to the global minimum. In this paper, we propose a principled approach for adaptively annealing the scale for GNC by tracking the positive-definiteness (i.e. local convexity) of the Hessian of the cost function. We illustrate our approach using the classic problem of registering 3D correspondences in the presence of noise and outliers. We also develop approximations to the Hessian that significantly speeds up our method. The effectiveness of our approach is validated by comparing its performance with state-of-the-art 3D registration approaches on a number of synthetic and real datasets. Our approach is accurate and efficient and converges to the global solution more reliably than the state-of-the-art methods.

Optical flow estimation is a fundamental functionality in computer vision. An event-based camera, which asynchronously detects sparse intensity changes, is an ideal device for realizing low-latency estimation of the optical flow owing to its low-latency sensing mechanism. An existing method using local plane fitting of events could utilize the sparsity to realize incremental updates for low-latency estimation; however, its output is merely a normal component of the full optical flow. An alternative approach using a frame-based deep neural network could estimate the full flow; however, its intensive non-incremental dense operation prohibits the low-latency estimation. We propose tangentially elongated Gaussian (TEG) belief propagation (BP) that realizes incremental full-flow estimation. We model the probability of full flow as the joint distribution of TEGs from the normal flow measurements, such that the marginal of this distribution with correct prior equals the full flow. We formulate the marginalization using a message-passing based on the BP to realize efficient incremental updates using sparse measurements. In addition to the theoretical justification, we evaluate the effectiveness of the TEGBP in real-world datasets; it outperforms SOTA incremental quasi-full flow method by a large margin. The code will be open-sourced upon acceptance.

This paper introduces a robust and scalable Gaussian process regression (GPR) model via variational learning. This enables the application of Gaussian processes to a wide range of real data, which are often large-scale and contaminated by outliers. Towards this end, we employ a mixture likelihood model where outliers are assumed to be sampled from a uniform distribution. We next derive a variational formulation that jointly infers the mode of data, i.e., inlier or outlier, as well as hyperparameters by maximizing a lower bound of the true log marginal likelihood. Compared to previous robust GPR, our formulation approximates the exact posterior distribution. The inducing variable approximation and stochastic variational inference are further introduced to our variational framework, extending our model to large-scale data. We apply our model to two challenging real-world applications, namely feature matching and dense gene expression imputation. Extensive experiments demonstrate the superiority of our model in terms of robustness and speed. Notably, when matching 4k feature points, its inference is completed in milliseconds with almost no false matches. The code is at https://github.com/YifanLu2000/Robust-Scalable-GPR.

Integrating multiple sensors and addressing diverse tasks in an end-to-end algorithm are challenging yet critical topics for autonomous driving. To this end, we introduce BEVGuide, a novel Bird’s Eye-View (BEV) representation learning framework, representing the first attempt to unify a wide range of sensors under direct BEV guidance in an end-to-end fashion. Our architecture accepts input from a diverse sensor pool, including but not limited to Camera, Lidar and Radar sensors, and extracts BEV feature embeddings using a versatile and general transformer backbone. We design a BEV-guided multi-sensor attention block to take queries from BEV embeddings and learn the BEV representation from sensor-specific features. BEVGuide is efficient due to its lightweight backbone design and highly flexible as it supports almost any input sensor configurations. Extensive experiments demonstrate that our framework achieves exceptional performance in BEV perception tasks with a diverse sensor set. Project page is at https://yunzeman.github.io/BEVGuide.

Human-centric perceptions include a variety of vision tasks, which have widespread industrial applications, including surveillance, autonomous driving, and the metaverse. It is desirable to have a general pretrain model for versatile human-centric downstream tasks. This paper forges ahead along this path from the aspects of both benchmark and pretraining methods. Specifically, we propose a HumanBench based on existing datasets to comprehensively evaluate on the common ground the generalization abilities of different pretraining methods on 19 datasets from 6 diverse downstream tasks, including person ReID, pose estimation, human parsing, pedestrian attribute recognition, pedestrian detection, and crowd counting. To learn both coarse-grained and fine-grained knowledge in human bodies, we further propose a Projector AssisTed Hierarchical pretraining method (PATH) to learn diverse knowledge at different granularity levels. Comprehensive evaluations on HumanBench show that our PATH achieves new state-of-the-art results on 17 downstream datasets and on-par results on the other 2 datasets. The code will be publicly at https://github.com/OpenGVLab/HumanBench.

End-to-end autonomous driving has made impressive progress in recent years. Existing methods usually adopt the decoupled encoder-decoder paradigm, where the encoder extracts hidden features from raw sensor data, and the decoder outputs the ego-vehicle’s future trajectories or actions. Under such a paradigm, the encoder does not have access to the intended behavior of the ego agent, leaving the burden of finding out safety-critical regions from the massive receptive field and inferring about future situations to the decoder. Even worse, the decoder is usually composed of several simple multi-layer perceptrons (MLP) or GRUs while the encoder is delicately designed (e.g., a combination of heavy ResNets or Transformer). Such an imbalanced resource-task division hampers the learning process. In this work, we aim to alleviate the aforementioned problem by two principles: (1) fully utilizing the capacity of the encoder; (2) increasing the capacity of the decoder. Concretely, we first predict a coarse-grained future position and action based on the encoder features. Then, conditioned on the position and action, the future scene is imagined to check the ramification if we drive accordingly. We also retrieve the encoder features around the predicted coordinate to obtain fine-grained information about the safety-critical region. Finally, based on the predicted future and the retrieved salient feature, we refine the coarse-grained position and action by predicting its offset from ground-truth. The above refinement module could be stacked in a cascaded fashion, which extends the capacity of the decoder with spatial-temporal prior knowledge about the conditioned future. We conduct experiments on the CARLA simulator and achieve state-of-the-art performance in closed-loop benchmarks. Extensive ablation studies demonstrate the effectiveness of each proposed module. Code and models are available at https://github.com/opendrivelab/ThinkTwice.

Motion forecasting is a key module in an autonomous driving system. Due to the heterogeneous nature of multi-sourced input, multimodality in agent behavior, and low latency required by onboard deployment, this task is notoriously challenging. To cope with these difficulties, this paper proposes a novel agent-centric model with anchor-informed proposals for efficient multimodal motion forecasting. We design a modality-agnostic strategy to concisely encode the complex input in a unified manner. We generate diverse proposals, fused with anchors bearing goal-oriented context, to induce multimodal prediction that covers a wide range of future trajectories. The network architecture is highly uniform and succinct, leading to an efficient model amenable for real-world deployment. Experiments reveal that our agent-centric network compares favorably with the state-of-the-art methods in prediction accuracy, while achieving scene-centric level inference latency.

Spatial reasoning tasks in multi-agent environments such as event prediction, agent type identification, or missing data imputation are important for multiple applications (e.g., autonomous surveillance over sensor networks and subtasks for reinforcement learning (RL)). StarCraft II game replays encode intelligent (and adversarial) multi-agent behavior and could provide a testbed for these tasks; however, extracting simple and standardized representations for prototyping these tasks is laborious and hinders reproducibility. In contrast, MNIST and CIFAR10, despite their extreme simplicity, have enabled rapid prototyping and reproducibility of ML methods. Following the simplicity of these datasets, we construct a benchmark spatial reasoning dataset based on StarCraft II replays that exhibit complex multi-agent behaviors, while still being as easy to use as MNIST and CIFAR10. Specifically, we carefully summarize a window of 255 consecutive game states to create 3.6 million summary images from 60,000 replays, including all relevant metadata such as game outcome and player races. We develop three formats of decreasing complexity: Hyperspectral images that include one channel for every unit type (similar to multispectral geospatial images), RGB images that mimic CIFAR10, and grayscale images that mimic MNIST. We show how this dataset can be used for prototyping spatial reasoning methods. All datasets, code for extraction, and code for dataset loading can be found at https://starcraftdata.davidinouye.com/.

To comprehensively cover the uncertainty of the future, the common practice of multi-modal human trajectory prediction is to first generate a set/distribution of candidate future trajectories and then sample required numbers of trajectories from them as final predictions. Even though a large number of previous researches develop various strong models to predict candidate trajectories, how to effectively sample the final ones has not received much attention yet. In this paper, we propose stimulus verification, serving as a universal and effective sampling process to improve the multi-modal prediction capability, where stimulus refers to the factor in the observation that may affect the future movements such as social interaction and scene context. Stimulus verification introduces a probabilistic model, denoted as stimulus verifier, to verify the coherence between a predicted future trajectory and its corresponding stimulus. By highlighting prediction samples with better stimulus-coherence, stimulus verification ensures sampled trajectories plausible from the stimulus’ point of view and therefore aids in better multi-modal prediction performance. We implement stimulus verification on five representative prediction frameworks and conduct exhaustive experiments on three widely-used benchmarks. Superior results demonstrate the effectiveness of our approach.

Deep learning has had remarkable success in robotic perception, but its data-centric nature suffers when it comes to generalizing to ever-changing environments. By contrast, physics-based optimization generalizes better, but it does not perform as well in complicated tasks due to the lack of high-level semantic information and reliance on manual parametric tuning. To take advantage of these two complementary worlds, we present PyPose: a robotics-oriented, PyTorch-based library that combines deep perceptual models with physics-based optimization. PyPose’s architecture is tidy and well-organized, it has an imperative style interface and is efficient and user-friendly, making it easy to integrate into real-world robotic applications. Besides, it supports parallel computing of any order gradients of Lie groups and Lie algebras and 2nd-order optimizers, such as trust region methods. Experiments show that PyPose achieves more than 10× speedup in computation compared to the state-of-the-art libraries. To boost future research, we provide concrete examples for several fields of robot learning, including SLAM, planning, control, and inertial navigation.

Gaze estimation across domains has been explored recently because the training data are usually collected under controlled conditions while the trained gaze estimators are used in real and diverse environments. However, due to privacy and efficiency concerns, simultaneous access to annotated source data and to-be-predicted target data can be challenging. In light of this, we present an unsupervised source-free domain adaptation approach for gaze estimation, which adapts a source-trained gaze estimator to unlabeled target domains without source data. We propose the Uncertainty Reduction Gaze Adaptation (UnReGA) framework, which achieves adaptation by reducing both sample and model uncertainty. Sample uncertainty is mitigated by enhancing image quality and making them gaze-estimation-friendly, whereas model uncertainty is reduced by minimizing prediction variance on the same inputs. Extensive experiments are conducted on six cross-domain tasks, demonstrating the effectiveness of UnReGA and its components. Results show that UnReGA outperforms other state-of-the-art cross-domain gaze estimation methods under both protocols, with and without source data

Camouflaged object detection (COD) aims to address the tough issue of identifying camouflaged objects visually blended into the surrounding backgrounds. COD is a challenging task due to the intrinsic similarity of camouflaged objects with the background, as well as their ambiguous boundaries. Existing approaches to this problem have developed various techniques to mimic the human visual system. Albeit effective in many cases, these methods still struggle when camouflaged objects are so deceptive to the vision system. In this paper, we propose the FEature Decomposition and Edge Reconstruction (FEDER) model for COD. The FEDER model addresses the intrinsic similarity of foreground and background by decomposing the features into different frequency bands using learnable wavelets. It then focuses on the most informative bands to mine subtle cues that differentiate foreground and background. To achieve this, a frequency attention module and a guidance-based feature aggregation module are developed. To combat the ambiguous boundary problem, we propose to learn an auxiliary edge reconstruction task alongside the COD task. We design an ordinary differential equation-inspired edge reconstruction module that generates exact edges. By learning the auxiliary task in conjunction with the COD task, the FEDER model can generate precise prediction maps with accurate object boundaries. Experiments show that our FEDER model significantly outperforms state-of-the-art methods with cheaper computational and memory costs.

In this paper, we propose MOTRv2, a simple yet effective pipeline to bootstrap end-to-end multi-object tracking with a pretrained object detector. Existing end-to-end methods, e.g. MOTR and TrackFormer are inferior to their tracking-by-detection counterparts mainly due to their poor detection performance. We aim to improve MOTR by elegantly incorporating an extra object detector. We first adopt the anchor formulation of queries and then use an extra object detector to generate proposals as anchors, providing detection prior to MOTR. The simple modification greatly eases the conflict between joint learning detection and association tasks in MOTR. MOTRv2 keeps the end-to-end feature and scales well on large-scale benchmarks. MOTRv2 achieves the top performance (73.4% HOTA) among all existing methods on the DanceTrack dataset. Moreover, MOTRv2 reaches state-of-the-art performance on the BDD100K dataset. We hope this simple and effective pipeline can provide some new insights to the end-to-end MOT community. The code will be released in the near future.

Clothing-change person re-identification (CC Re-ID) aims to match the same person who changes clothes across cameras. Current methods are usually limited by the insufficient number and variation of clothing in training data, e.g. each person only has 2 outfits in the PRCC dataset. In this work, we propose a novel Clothing-Change Feature Augmentation (CCFA) model for CC Re-ID to largely expand clothing-change data in the feature space rather than visual image space. It automatically models the feature distribution expansion that reflects a person’s clothing colour and texture variations to augment model training. Specifically, to formulate meaningful clothing variations in the feature space, our method first estimates a clothing-change normal distribution with intra-ID cross-clothing variances. Then an augmentation generator learns to follow the estimated distribution to augment plausible clothing-change features. The augmented features are guaranteed to maximise the change of clothing and minimise the change of identity properties by adversarial learning to assure the effectiveness. Such augmentation is performed iteratively with an ID-correlated augmentation strategy to increase intra-ID clothing variations and reduce inter-ID clothing variations, enforcing the Re-ID model to learn clothing-independent features inherently. Extensive experiments demonstrate the effectiveness of our method with state-of-the-art results on CC Re-ID datasets.

Gait recognition is beneficial for a variety of applications, including video surveillance, crime scene investigation, and social security, to mention a few. However, gait recognition often suffers from multiple exterior factors in real scenes, such as carrying conditions, wearing overcoats, and diverse viewing angles. Recently, various deep learning-based gait recognition methods have achieved promising results, but they tend to extract one of the salient features using fixed-weighted convolutional networks, do not well consider the relationship within gait features in key regions, and ignore the aggregation of complete motion patterns. In this paper, we propose a new perspective that actual gait features include global motion patterns in multiple key regions, and each global motion pattern is composed of a series of local motion patterns. To this end, we propose a Dynamic Aggregation Network (DANet) to learn more discriminative gait features. Specifically, we create a dynamic attention mechanism between the features of neighboring pixels that not only adaptively focuses on key regions but also generates more expressive local motion patterns. In addition, we develop a self-attention mechanism to select representative local motion patterns and further learn robust global motion patterns. Extensive experiments on three popular public gait datasets, i.e., CASIA-B, OUMVLP, and Gait3D, demonstrate that the proposed method can provide substantial improvements over the current state-of-the-art methods.

Micro-expressions are spontaneous, rapid and subtle facial movements that can neither be forged nor suppressed. They are very important nonverbal communication clues, but are transient and of low intensity thus difficult to recognize. Recently deep learning based methods have been developed for micro-expression recognition using feature extraction and fusion techniques, however, targeted feature learning and efficient feature fusion still lack further study according to micro-expression characteristics. To address these issues, we propose a novel framework Feature Representation Learning with adaptive Displacement Generation and Transformer fusion (FRL-DGT), in which a convolutional Displacement Generation Module (DGM) with self-supervised learning is used to extract dynamic feature targeted to the subsequent ME recognition task, and a well-designed Transformer fusion mechanism composed of the Transformer-based local fusion module, global fusion module, and full-face fusion module is applied to extract the multi-level informative feature from the output of the DGM for the final micro-expression prediction. Extensive experiments with solid leave-one-subject-out (LOSO) evaluation results have strongly demonstrated the superiority of our proposed FRL-DGT to state-of-the-art methods.

In this work, we propose an ID-preserving talking head generation framework, which advances previous methods in two aspects. First, as opposed to interpolating from sparse flow, we claim that dense landmarks are crucial to achieving accurate geometry-aware flow fields. Second, inspired by face-swapping methods, we adaptively fuse the source identity during synthesis, so that the network better preserves the key characteristics of the image portrait. Although the proposed model surpasses prior generation fidelity on established benchmarks, personalized fine-tuning is still needed to further make the talking head generation qualified for real usage. However, this process is rather computationally demanding that is unaffordable to standard users. To alleviate this, we propose a fast adaptation model using a meta-learning approach. The learned model can be adapted to a high-quality personalized model as fast as 30 seconds. Last but not least, a spatial-temporal enhancement module is proposed to improve the fine details while ensuring temporal coherency. Extensive experiments prove the significant superiority of our approach over the state of the arts in both one-shot and personalized settings.

With the continuously thriving popularity around the world, fitness activity analytic has become an emerging research topic in computer vision. While a variety of new tasks and algorithms have been proposed recently, there are growing hunger for data resources involved in high-quality data, fine-grained labels, and diverse environments. In this paper, we present FLAG3D, a large-scale 3D fitness activity dataset with language instruction containing 180K sequences of 60 categories. FLAG3D features the following three aspects: 1) accurate and dense 3D human pose captured from advanced MoCap system to handle the complex activity and large movement, 2) detailed and professional language instruction to describe how to perform a specific activity, 3) versatile video resources from a high-tech MoCap system, rendering software, and cost-effective smartphones in natural environments. Extensive experiments and in-depth analysis show that FLAG3D contributes great research value for various challenges, such as cross-domain human action recognition, dynamic human mesh recovery, and language-guided human action generation. Our dataset and source code are publicly available at https://andytang15.github.io/FLAG3D.

Person re-identification (re-ID) via 3D skeleton data is an emerging topic with prominent advantages. Existing methods usually design skeleton descriptors with raw body joints or perform skeleton sequence representation learning. However, they typically cannot concurrently model different body-component relations, and rarely explore useful semantics from fine-grained representations of body joints. In this paper, we propose a generic Transformer-based Skeleton Graph prototype contrastive learning (TranSG) approach with structure-trajectory prompted reconstruction to fully capture skeletal relations and valuable spatial-temporal semantics from skeleton graphs for person re-ID. Specifically, we first devise the Skeleton Graph Transformer (SGT) to simultaneously learn body and motion relations within skeleton graphs, so as to aggregate key correlative node features into graph representations. Then, we propose the Graph Prototype Contrastive learning (GPC) to mine the most typical graph features (graph prototypes) of each identity, and contrast the inherent similarity between graph representations and different prototypes from both skeleton and sequence levels to learn discriminative graph representations. Last, a graph Structure-Trajectory Prompted Reconstruction (STPR) mechanism is proposed to exploit the spatial and temporal contexts of graph nodes to prompt skeleton graph reconstruction, which facilitates capturing more valuable patterns and graph semantics for person re-ID. Empirical evaluations demonstrate that TranSG significantly outperforms existing state-of-the-art methods. We further show its generality under different graph modeling, RGB-estimated skeletons, and unsupervised scenarios.

The task of reconstructing 3D human motion has wide-ranging applications. The gold standard Motion capture (MoCap) systems are accurate but inaccessible to the general public due to their cost, hardware, and space constraints. In contrast, monocular human mesh recovery (HMR) methods are much more accessible than MoCap as they take single-view videos as inputs. Replacing the multi-view MoCap systems with a monocular HMR method would break the current barriers to collecting accurate 3D motion thus making exciting applications like motion analysis and motion-driven animation accessible to the general public. However, the performance of existing HMR methods degrades when the video contains challenging and dynamic motion that is not in existing MoCap datasets used for training. This reduces its appeal as dynamic motion is frequently the target in 3D motion recovery in the aforementioned applications. Our study aims to bridge the gap between monocular HMR and multi-view MoCap systems by leveraging information shared across multiple video instances of the same action. We introduce the Neural Motion (NeMo) field. It is optimized to represent the underlying 3D motions across a set of videos of the same action. Empirically, we show that NeMo can recover 3D motion in sports using videos from the Penn Action dataset, where NeMo outperforms existing HMR methods in terms of 2D keypoint detection. To further validate NeMo using 3D metrics, we collected a small MoCap dataset mimicking actions in Penn Action, and show that NeMo achieves better 3D reconstruction compared to various baselines.

Motion segmentation is one of the main tasks in computer vision and is relevant for many applications. The optical flow (OF) is the input generally used to segment every frame of a video sequence into regions of coherent motion. Temporal consistency is a key feature of motion segmentation, but it is often neglected. In this paper, we propose an original unsupervised spatio-temporal framework for motion segmentation from optical flow that fully investigates the temporal dimension of the problem. More specifically, we have defined a 3D network for multiple motion segmentation that takes as input a sub-volume of successive optical flows and delivers accordingly a sub-volume of coherent segmentation maps. Our network is trained in a fully unsupervised way, and the loss function combines a flow reconstruction term involving spatio-temporal parametric motion models, and a regularization term enforcing temporal consistency on the masks. We have specified an easy temporal linkage of the predicted segments. Besides, we have proposed a flexible and efficient way of coding U-nets. We report experiments on several VOS benchmarks with convincing quantitative results, while not using appearance and not training with any ground-truth data. We also highlight through visual results the distinctive contribution of the short- and long-term temporal consistency brought by our OF segmentation method.

The polarization event camera PDAVIS is a novel bio-inspired neuromorphic vision sensor that reports both conventional polarization frames and asynchronous, continuously per-pixel polarization brightness changes (polarization events) with fast temporal resolution and large dynamic range. A deep neural network method (Polarization FireNet) was previously developed to reconstruct the polarization angle and degree from polarization events for bridging the gap between the polarization event camera and mainstream computer vision. However, Polarization FireNet applies a network pre-trained for normal event-based frame reconstruction independently on each of four channels of polarization events from four linear polarization angles, which ignores the correlations between channels and inevitably introduces content inconsistency between the four reconstructed frames, resulting in unsatisfactory polarization reconstruction performance. In this work, we strive to train an effective, yet efficient, DNN model that directly outputs polarization from the input raw polarization events. To this end, we constructed the first large-scale event-to-polarization dataset, which we subsequently employed to train our events-to-polarization network E2P. E2P extracts rich polarization patterns from input polarization events and enhances features through cross-modality context integration. We demonstrate that E2P outperforms Polarization FireNet by a significant margin with no additional computing cost. Experimental results also show that E2P produces more accurate measurement of polarization than the PDAVIS frames in challenging fast and high dynamic range scenes.

Continuous-time video frame interpolation is a fundamental technique in computer vision for its flexibility in synthesizing motion trajectories and novel video frames at arbitrary intermediate time steps. Yet, how to infer accurate intermediate motion and synthesize high-quality video frames are two critical challenges. In this paper, we present a novel VFI framework with improved treatment for these challenges. To address the former, we propose focalized trajectory fitting, which performs confidence-aware motion trajectory estimation by learning to pay focus to reliable optical flow candidates while suppressing the outliers. The second is range-nullspace synthesis, a novel frame renderer cast as solving an ill-posed problem addressed by learning decoupled components in orthogonal subspaces. The proposed framework sets new records on 7 of 10 public VFI benchmarks.

For video frame interpolation(VFI), existing deep-learning-based approaches strongly rely on the ground-truth (GT) intermediate frames, which sometimes ignore the non-unique nature of motion judging from the given adjacent frames. As a result, these methods tend to produce averaged solutions that are not clear enough. To alleviate this issue, we propose to relax the requirement of reconstructing an intermediate frame as close to the GT as possible. Towards this end, we develop a texture consistency loss (TCL) upon the assumption that the interpolated content should maintain similar structures with their counterparts in the given frames. Predictions satisfying this constraint are encouraged, though they may differ from the predefined GT. Without the bells and whistles, our plug-and-play TCL is capable of improving the performance of existing VFI frameworks consistently. On the other hand, previous methods usually adopt the cost volume or correlation map to achieve more accurate image or feature warping. However, the O(N^2) (N refers to the pixel count) computational complexity makes it infeasible for high-resolution cases. In this work, we design a simple, efficient O(N) yet powerful guided cross-scale pyramid alignment(GCSPA) module, where multi-scale information is highly exploited. Extensive experiments justify the efficiency and effectiveness of the proposed strategy.

Capturing high frame rate and high dynamic range (HFR&HDR) color videos in high-speed scenes with conventional frame-based cameras is very challenging. The increasing frame rate is usually guaranteed by using shorter exposure time so that the captured video is severely interfered by noise. Alternating exposures could alleviate the noise issue but sacrifice frame rate due to involving long-exposure frames. The neuromorphic spiking camera records high-speed scenes of high dynamic range without colors using a completely different sensing mechanism and visual representation. We introduce a hybrid camera system composed of a spiking and an alternating-exposure RGB camera to capture HFR&HDR scenes with high fidelity. Our insight is to bring each camera’s superiority into full play. The spike frames, with accurate fast motion information encoded, are first reconstructed for motion representation, from which the spike-based optical flows guide the recovery of missing temporal information for middle- and long-exposure RGB images while retaining their reliable color appearances. With the strong temporal constraint estimated from spike trains, both missing and distorted colors cross RGB frames are recovered to generate time-consistent and HFR color frames. We collect a new Spike-RGB dataset that contains 300 sequences of synthetic data and 20 groups of real-world data to demonstrate 1000 FPS HDR videos outperforming HDR video reconstruction methods and commercial high-speed cameras.

How to effectively explore spatial and temporal information is important for video deblurring. In contrast to existing methods that directly align adjacent frames without discrimination, we develop a deep discriminative spatial and temporal network to facilitate the spatial and temporal feature exploration for better video deblurring. We first develop a channel-wise gated dynamic network to adaptively explore the spatial information. As adjacent frames usually contain different contents, directly stacking features of adjacent frames without discrimination may affect the latent clear frame restoration. Therefore, we develop a simple yet effective discriminative temporal feature fusion module to obtain useful temporal features for latent frame restoration. Moreover, to utilize the information from long-range frames, we develop a wavelet-based feature propagation method that takes the discriminative temporal feature fusion module as the basic unit to effectively propagate main structures from long-range frames for better video deblurring. We show that the proposed method does not require additional alignment methods and performs favorably against state-of-the-art ones on benchmark datasets in terms of accuracy and model complexity.

Burst image processing is becoming increasingly popular in recent years. However, it is a challenging task since individual burst images undergo multiple degradations and often have mutual misalignments resulting in ghosting and zipper artifacts. Existing burst restoration methods usually do not consider the mutual correlation and non-local contextual information among burst frames, which tends to limit these approaches in challenging cases. Another key challenge lies in the robust up-sampling of burst frames. The existing up-sampling methods cannot effectively utilize the advantages of single-stage and progressive up-sampling strategies with conventional and/or recent up-samplers at the same time. To address these challenges, we propose a novel Gated Multi-Resolution Transfer Network (GMTNet) to reconstruct a spatially precise high-quality image from a burst of low-quality raw images. GMTNet consists of three modules optimized for burst processing tasks: Multi-scale Burst Feature Alignment (MBFA) for feature denoising and alignment, Transposed-Attention Feature Merging (TAFM) for multi-frame feature aggregation, and Resolution Transfer Feature Up-sampler (RTFU) to up-scale merged features and construct a high-quality output image. Detailed experimental analysis on five datasets validate our approach and sets a new state-of-the-art for burst super-resolution, burst denoising, and low-light burst enhancement. Our codes and models are available at https://github.com/nanmehta/GMTNet.

Mapping Low Dynamic Range (LDR) images with different exposures to High Dynamic Range (HDR) remains nontrivial and challenging on dynamic scenes due to ghosting caused by object motion or camera jitting. With the success of Deep Neural Networks (DNNs), several DNNs-based methods have been proposed to alleviate ghosting, they cannot generate approving results when motion and saturation occur. To generate visually pleasing HDR images in various cases, we propose a hybrid HDR deghosting network, called HyHDRNet, to learn the complicated relationship between reference and non-reference images. The proposed HyHDRNet consists of a content alignment subnetwork and a Transformer-based fusion subnetwork. Specifically, to effectively avoid ghosting from the source, the content alignment subnetwork uses patch aggregation and ghost attention to integrate similar content from other non-reference images with patch level and suppress undesired components with pixel level. To achieve mutual guidance between patch-level and pixel-level, we leverage a gating module to sufficiently swap useful information both in ghosted and saturated regions. Furthermore, to obtain a high-quality HDR image, the Transformer-based fusion subnetwork uses a Residual Deformable Transformer Block (RDTB) to adaptively merge information for different exposed regions. We examined the proposed method on four widely used public HDR image deghosting datasets. Experiments demonstrate that HyHDRNet outperforms state-of-the-art methods both quantitatively and qualitatively, achieving appealing HDR visualization with unified textures and colors.

The well-documented presence of texture bias in modern convolutional neural networks has led to a plethora of algorithms that promote an emphasis on shape cues, often to support generalization to new domains. Yet, common datasets, benchmarks and general model selection strategies are missing, and there is no agreed, rigorous evaluation protocol. In this paper, we investigate difficulties and limitations when training networks with reduced texture bias. In particular, we also show that proper evaluation and meaningful comparisons between methods are not trivial. We introduce BiasBed, a testbed for texture- and style-biased training, including multiple datasets and a range of existing algorithms. It comes with an extensive evaluation protocol that includes rigorous hypothesis testing to gauge the significance of the results, despite the considerable training instability of some style bias methods. Our extensive experiments, shed new light on the need for careful, statistically founded evaluation protocols for style bias (and beyond). E.g., we find that some algorithms proposed in the literature do not significantly mitigate the impact of style bias at all. With the release of BiasBed, we hope to foster a common understanding of consistent and meaningful comparisons, and consequently faster progress towards learning methods free of texture bias. Code is available at https://github.com/D1noFuzi/BiasBed.

In media industry, the demand of SDR-to-HDRTV up-conversion arises when users possess HDR-WCG (high dynamic range-wide color gamut) TVs while most off-the-shelf footage is still in SDR (standard dynamic range). The research community has started tackling this low-level vision task by learning-based approaches. When applied to real SDR, yet, current methods tend to produce dim and desaturated result, making nearly no improvement on viewing experience. Different from other network-oriented methods, we attribute such deficiency to training set (HDR-SDR pair). Consequently, we propose new HDRTV dataset (dubbed HDRTV4K) and new HDR-to-SDR degradation models. Then, it’s used to train a luminance-segmented network (LSN) consisting of a global mapping trunk, and two Transformer branches on bright and dark luminance range. We also update assessment criteria by tailored metrics and subjective experiment. Finally, ablation studies are conducted to prove the effectiveness. Our work is available at: https://github.com/AndreGuo/HDRTVDM.

Most well-established and widely used color difference (CD) metrics are handcrafted and subject-calibrated against uniformly colored patches, which do not generalize well to photographic images characterized by natural scene complexities. Constructing CD formulae for photographic images is still an active research topic in imaging/illumination, vision science, and color science communities. In this paper, we aim to learn a deep CD metric for photographic images with four desirable properties. First, it well aligns with the observations in vision science that color and form are linked inextricably in visual cortical processing. Second, it is a proper metric in the mathematical sense. Third, it computes accurate CDs between photographic images, differing mainly in color appearances. Fourth, it is robust to mild geometric distortions (e.g., translation or due to parallax), which are often present in photographic images of the same scene captured by different digital cameras. We show that all these properties can be satisfied at once by learning a multi-scale autoregressive normalizing flow for feature transform, followed by the Euclidean distance which is linearly proportional to the human perceptual CD. Quantitative and qualitative experiments on the large-scale SPCD dataset demonstrate the promise of the learned CD metric.

Low-light Image Enhancement (LIE) aims at improving contrast and restoring details for images captured in low-light conditions. Most of the previous LIE algorithms adjust illumination using a single input image with several handcrafted priors. Those solutions, however, often fail in revealing image details due to the limited information in a single image and the poor adaptability of handcrafted priors. To this end, we propose PairLIE, an unsupervised approach that learns adaptive priors from low-light image pairs. First, the network is expected to generate the same clean images as the two inputs share the same image content. To achieve this, we impose the network with the Retinex theory and make the two reflectance components consistent. Second, to assist the Retinex decomposition, we propose to remove inappropriate features in the raw image with a simple self-supervised mechanism. Extensive experiments on public datasets show that the proposed PairLIE achieves comparable performance against the state-of-the-art approaches with a simpler network and fewer handcrafted priors. Code is available at: https://github.com/zhenqifu/PairLIE.

To acquire a snapshot spectral image, coded aperture snapshot spectral imaging (CASSI) is proposed. A core problem of the CASSI system is to recover the reliable and fine underlying 3D spectral cube from the 2D measurement. By alternately solving a data subproblem and a prior subproblem, deep unfolding methods achieve good performance. However, in the data subproblem, the used sensing matrix is ill-suited for the real degradation process due to the device errors caused by phase aberration, distortion; in the prior subproblem, it is important to design a suitable model to jointly exploit both spatial and spectral priors. In this paper, we propose a Residual Degradation Learning Unfolding Framework (RDLUF), which bridges the gap between the sensing matrix and the degradation process. Moreover, a MixS2 Transformer is designed via mixing priors across spectral and spatial to strengthen the spectral-spatial representation capability. Finally, plugging the MixS2 Transformer into the RDLUF leads to an end-to-end trainable and interpretable neural network RDLUF-MixS2. Experimental results establish the superior performance of the proposed method over existing ones.

For real applications, existing HSI-SR methods are mostly not only limited to unstable performance under unknown scenarios but also suffer from high computation consumption. In this paper, we develop a new coordination optimization framework for stable, interpretable, and lightweight HSI-SR. Specifically, we create a positive cycle between fusion and degradation estimation under a new probabilistic framework. The estimated degradation is applied to fusion as guidance for a degradation-aware HSI-SR. Under the framework, we establish an explicit degradation estimation method to tackle the indeterminacy and unstable performance driven by black-box simulation in previous methods. Considering the interpretability in fusion, we integrate spectral mixing prior to the fusion process, which can be easily realized by a tiny autoencoder, leading to a dramatic release of the computation burden. We then develop a partial fine-tune strategy in inference to reduce the computation cost further. Comprehensive experiments demonstrate the superiority of our method against state-of-the-art under synthetic and real datasets. For instance, we achieve a 2.3 dB promotion on PSNR with 120x model size reduction and 4300x FLOPs reduction under the CAVE dataset. Code is available in https://github.com/WenjinGuo/DAEM.

Existing dehazing approaches struggle to process real-world hazy images owing to the lack of paired real data and robust priors. In this work, we present a new paradigm for real image dehazing from the perspectives of synthesizing more realistic hazy data and introducing more robust priors into the network. Specifically, (1) instead of adopting the de facto physical scattering model, we rethink the degradation of real hazy images and propose a phenomenological pipeline considering diverse degradation types. (2) We propose a Real Image Dehazing network via high-quality Codebook Priors (RIDCP). Firstly, a VQGAN is pre-trained on a large-scale high-quality dataset to obtain the discrete codebook, encapsulating high-quality priors (HQPs). After replacing the negative effects brought by haze with HQPs, the decoder equipped with a novel normalized feature alignment module can effectively utilize high-quality features and produce clean results. However, although our degradation pipeline drastically mitigates the domain gap between synthetic and real data, it is still intractable to avoid it, which challenges HQPs matching in the wild. Thus, we re-calculate the distance when matching the features to the HQPs by a controllable matching operation, which facilitates finding better counterparts. We provide a recommendation to control the matching based on an explainable solution. Users can also flexibly adjust the enhancement degree as per their preference. Extensive experiments verify the effectiveness of our data synthesis pipeline and the superior performance of RIDCP in real image dehazing. Code and data will be released.

GAN-based image restoration inverts the generative process to repair images corrupted by known degradations. Existing unsupervised methods must carefully be tuned for each task and degradation level. In this work, we make StyleGAN image restoration robust: a single set of hyperparameters works across a wide range of degradation levels. This makes it possible to handle combinations of several degradations, without the need to retune. Our proposed approach relies on a 3-phase progressive latent space extension and a conservative optimizer, which avoids the need for any additional regularization terms. Extensive experiments demonstrate robustness on inpainting, upsampling, denoising, and deartifacting at varying degradations levels, outperforming other StyleGAN-based inversion techniques. Our approach also favorably compares to diffusion-based restoration by yielding much more realistic inversion results. Code will be released upon publication.

Blind image quality assessment (BIQA) aims to automatically evaluate the perceived quality of a single image, whose performance has been improved by deep learning-based methods in recent years. However, the paucity of labeled data somewhat restrains deep learning-based BIQA methods from unleashing their full potential. In this paper, we propose to solve the problem by a pretext task customized for BIQA in a self-supervised learning manner, which enables learning representations from orders of magnitude more data. To constrain the learning process, we propose a quality-aware contrastive loss based on a simple assumption: the quality of patches from a distorted image should be similar, but vary from patches from the same image with different degradations and patches from different images. Further, we improve the existing degradation process and form a degradation space with the size of roughly 2x10^7. After pre-trained on ImageNet using our method, models are more sensitive to image quality and perform significantly better on downstream BIQA tasks. Experimental results show that our method obtains remarkable improvements on popular BIQA datasets.

The hardware image signal processing (ISP) pipeline is the intermediate layer between the imaging sensor and the downstream application, processing the sensor signal into an RGB image. The ISP is less programmable and consists of a series of processing modules. Each processing module handles a subtask and contains a set of tunable hyperparameters. A large number of hyperparameters form a complex mapping with the ISP output. The industry typically relies on manual and time-consuming hyperparameter tuning by image experts, biased towards human perception. Recently, several automatic ISP hyperparameter optimization methods using downstream evaluation metrics come into sight. However, existing methods for ISP tuning treat the high-dimensional parameter space as a global space for optimization and prediction all at once without inducing the structure knowledge of ISP. To this end, we propose a sequential ISP hyperparameter prediction framework that utilizes the sequential relationship within ISP modules and the similarity among parameters to guide the model sequence process. We validate the proposed method on object detection, image segmentation, and image quality tasks.

By optimizing the rate-distortion-realism trade-off, generative compression approaches produce detailed, realistic images, even at low bit rates, instead of the blurry reconstructions produced by rate-distortion optimized models. However, previous methods do not explicitly control how much detail is synthesized, which results in a common criticism of these methods: users might be worried that a misleading reconstruction far from the input image is generated. In this work, we alleviate these concerns by training a decoder that can bridge the two regimes and navigate the distortion-realism trade-off. From a single compressed representation, the receiver can decide to either reconstruct a low mean squared error reconstruction that is close to the input, a realistic reconstruction with high perceptual quality, or anything in between. With our method, we set a new state-of-the-art in distortion-realism, pushing the frontier of achievable distortion-realism pairs, i.e., our method achieves better distortions at high realism and better realism at low distortion than ever before.

Most neural networks for computer vision are designed to infer using RGB images. However, these RGB images are commonly encoded in JPEG before saving to disk; decoding them imposes an unavoidable overhead for RGB networks. Instead, our work focuses on training Vision Transformers (ViT) directly from the encoded features of JPEG. This way, we can avoid most of the decoding overhead, accelerating data load. Existing works have studied this aspect but they focus on CNNs. Due to how these encoded features are structured, CNNs require heavy modification to their architecture to accept such data. Here, we show that this is not the case for ViTs. In addition, we tackle data augmentation directly on these encoded features, which to our knowledge, has not been explored in-depth for training in this setting. With these two improvements -- ViT and data augmentation -- we show that our ViT-Ti model achieves up to 39.2% faster training and 17.9% faster inference with no accuracy loss compared to the RGB counterpart.

Deep learning based methods for super-resolution have become state-of-the-art and outperform traditional approaches by a significant margin. From the initial models designed for fixed integer scaling factors (e.g. x2 or x4), efforts were made to explore different directions such as modeling blur kernels or addressing non-integer scaling factors. However, existing works do not provide a sound framework to handle them jointly. In this paper we propose a framework for generic image resampling that not only addresses all the above mentioned issues but extends the sets of possible transforms from upscaling to generic transforms. A key aspect to unlock these capabilities is the faithful modeling of image warping and changes of the sampling rate during the training data preparation. This allows a localized representation of the implicit image degradation that takes into account the reconstruction kernel, the local geometric distortion and the anti-aliasing kernel. Using this spatially variant degradation map as conditioning for our resampling model, we can address with the same model both global transformations, such as upscaling or rotation, and locally varying transformations such lens distortion or undistortion. Another important contribution is the automatic estimation of the degradation map in this more complex resampling setting (i.e. blind image resampling). Finally, we show that state-of-the-art results can be achieved by predicting kernels to apply on the input image instead of direct color prediction. This renders our model applicable for different types of data not seen during the training such as normals.

Face super-resolution (FSR) aims to reconstruct high-resolution (HR) face images from the low-resolution (LR) ones. With the advent of deep learning, the FSR technique has achieved significant breakthroughs. However, existing FSR methods either have a fixed receptive field or fail to maintain facial structure, limiting the FSR performance. To circumvent this problem, Fourier transform is introduced, which can capture global facial structure information and achieve image-size receptive field. Relying on the Fourier transform, we devise a spatial-frequency mutual network (SFMNet) for FSR, which is the first FSR method to explore the correlations between spatial and frequency domains as far as we know. To be specific, our SFMNet is a two-branch network equipped with a spatial branch and a frequency branch. Benefiting from the property of Fourier transform, the frequency branch can achieve image-size receptive field and capture global dependency while the spatial branch can extract local dependency. Considering that these dependencies are complementary and both favorable for FSR, we further develop a frequency-spatial interaction block (FSIB) which mutually amalgamates the complementary spatial and frequency information to enhance the capability of the model. Quantitative and qualitative experimental results show that the proposed method outperforms state-of-the-art FSR methods in recovering face images. The implementation and model will be released at https://github.com/wcy-cs/SFMNet.

Transformer-based methods have shown impressive performance in low-level vision tasks, such as image super-resolution. However, we find that these networks can only utilize a limited spatial range of input information through attribution analysis. This implies that the potential of Transformer is still not fully exploited in existing networks. In order to activate more input pixels for better reconstruction, we propose a novel Hybrid Attention Transformer (HAT). It combines both channel attention and window-based self-attention schemes, thus making use of their complementary advantages of being able to utilize global statistics and strong local fitting capability. Moreover, to better aggregate the cross-window information, we introduce an overlapping cross-attention module to enhance the interaction between neighboring window features. In the training stage, we additionally adopt a same-task pre-training strategy to exploit the potential of the model for further improvement. Extensive experiments show the effectiveness of the proposed modules, and we further scale up the model to demonstrate that the performance of this task can be greatly improved. Our overall method significantly outperforms the state-of-the-art methods by more than 1dB. Codes and models are available at https://github.com/XPixelGroup/HAT.

While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) paradigm is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering a omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves record-high performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 x4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR.

Image aesthetics assessment (IAA) is a challenging task due to its highly subjective nature. Most of the current studies rely on large-scale datasets (e.g., AVA and AADB) to learn a general model for all kinds of photography images. However, little light has been shed on measuring the aesthetic quality of artistic images, and the existing datasets only contain relatively few artworks. Such a defect is a great obstacle to the aesthetic assessment of artistic images. To fill the gap in the field of artistic image aesthetics assessment (AIAA), we first introduce a large-scale AIAA dataset: Boldbrush Artistic Image Dataset (BAID), which consists of 60,337 artistic images covering various art forms, with more than 360,000 votes from online users. We then propose a new method, SAAN (Style-specific Art Assessment Network), which can effectively extract and utilize style-specific and generic aesthetic information to evaluate artistic images. Experiments demonstrate that our proposed approach outperforms existing IAA methods on the proposed BAID dataset according to quantitative comparisons. We believe the proposed dataset and method can serve as a foundation for future AIAA works and inspire more research in this field.

Automatic prohibited item detection in security inspection X-ray images is necessary for transportation.The abundance and diversity of the X-ray security images with prohibited item, termed as prohibited X-ray security images, are essential for training the detection model. In order to solve the data insufficiency, we propose a RegionWise Style-Controlled Fusion (RWSC-Fusion) network, which superimposes the prohibited items onto the normal X-ray security images, to synthesize the prohibited X-ray security images. The proposed RWSC-Fusion innovates both network structure and loss functions to generate more realistic X-ray security images. Specifically, a RWSCFusion module is designed to enable the region-wise fusion by controlling the appearance of the overlapping region with novel modulation parameters. In addition, an EdgeAttention (EA) module is proposed to effectively improve the sharpness of the synthetic images. As for the unsupervised loss function, we propose the Luminance loss in Logarithmic form (LL) and Correlation loss of Saturation Difference (CSD), to optimize the fused X-ray security images in terms of luminance and saturation. We evaluate the authenticity and the training effect of the synthetic X-ray security images on private and public SIXray dataset. The results confirm that our synthetic images are reliable enough to augment the prohibited Xray security images.

Any-scale image synthesis offers an efficient and scalable solution to synthesize photo-realistic images at any scale, even going beyond 2K resolution. However, existing GAN-based solutions depend excessively on convolutions and a hierarchical architecture, which introduce inconsistency and the “texture sticking” issue when scaling the output resolution. From another perspective, INR-based generators are scale-equivariant by design, but their huge memory footprint and slow inference hinder these networks from being adopted in large-scale or real-time systems. In this work, we propose Column-Row Entangled Pixel Synthesisthes (CREPS), a new generative model that is both efficient and scale-equivariant without using any spatial convolutions or coarse-to-fine design. To save memory footprint and make the system scalable, we employ a novel bi-line representation that decomposes layer-wise feature maps into separate “thick” column and row encodings. Experiments on standard datasets, including FFHQ, LSUN-Church, and MetFaces, confirm CREPS’ ability to synthesize scale-consistent and alias-free images up to 4K resolution with proper training and inference speed.

We present a novel framework for exemplar based image translation. Recent advanced methods for this task mainly focus on establishing cross-domain semantic correspondence, which sequentially dominates image generation in the manner of local style control. Unfortunately, cross domain semantic matching is challenging; and matching errors ultimately degrade the quality of generated images. To overcome this challenge, we improve the accuracy of matching on the one hand, and diminish the role of matching in image generation on the other hand. To achieve the former, we propose a masked and adaptive transformer (MAT) for learning accurate cross-domain correspondence, and executing context-aware feature augmentation. To achieve the latter, we use source features of the input and global style codes of the exemplar, as supplementary information, for decoding an image. Besides, we devise a novel contrastive style learning method, for acquire quality-discriminative style representations, which in turn benefit high-quality image generation. Experimental results show that our method, dubbed MATEBIT, performs considerably better than state-of-the-art methods, in diverse image translation tasks.

Generic image inpainting aims to complete a corrupted image by borrowing surrounding information, which barely generates novel content. By contrast, multi-modal inpainting provides more flexible and useful controls on the inpainted content, e.g., a text prompt can be used to describe an object with richer attributes, and a mask can be used to constrain the shape of the inpainted object rather than being only considered as a missing area. We propose a new diffusion-based model named SmartBrush for completing a missing region with an object using both text and shape-guidance. While previous work such as DALLE-2 and Stable Diffusion can do text-guided inapinting they do not support shape guidance and tend to modify background texture surrounding the generated object. Our model incorporates both text and shape guidance with precision control. To preserve the background better, we propose a novel training and sampling strategy by augmenting the diffusion U-net with object-mask prediction. Lastly, we introduce a multi-task training strategy by jointly training inpainting with text-to-image generation to leverage more training data. We conduct extensive experiments showing that our model outperforms all baselines in terms of visual quality, mask controllability, and background preservation.

Few-shot font generation (FFG), aiming at generating font images with a few samples, is an emerging topic in recent years due to the academic and commercial values. Typically, the FFG approaches follow the style-content disentanglement paradigm, which transfers the target font styles to characters by combining the content representations of source characters and the style codes of reference samples. Most existing methods attempt to increase font generation ability via exploring powerful style representations, which may be a sub-optimal solution for the FFG task due to the lack of modeling spatial transformation in transferring font styles. In this paper, we model font generation as a continuous transformation process from the source character image to the target font image via the creation and dissipation of font pixels, and embed the corresponding transformations into a neural transformation field. With the estimated transformation path, the neural transformation field generates a set of intermediate transformation results via the sampling process, and a font rendering formula is developed to accumulate them into the target font image. Extensive experiments show that our method achieves state-of-the-art performance on few-shot font generation task, which demonstrates the effectiveness of our proposed model. Our implementation is available at: https://github.com/fubinfb/NTF.

Different from conventional image matting, which either requires user-defined scribbles/trimap to extract a specific foreground object or directly extracts all the foreground objects in the image indiscriminately, we introduce a new task named Referring Image Matting (RIM) in this paper, which aims to extract the meticulous alpha matte of the specific object that best matches the given natural language description, thus enabling a more natural and simpler instruction for image matting. First, we establish a large-scale challenging dataset RefMatte by designing a comprehensive image composition and expression generation engine to automatically produce high-quality images along with diverse text attributes based on public datasets. RefMatte consists of 230 object categories, 47,500 images, 118,749 expression-region entities, and 474,996 expressions. Additionally, we construct a real-world test set with 100 high-resolution natural images and manually annotate complex phrases to evaluate the out-of-domain generalization abilities of RIM methods. Furthermore, we present a novel baseline method CLIPMat for RIM, including a context-embedded prompt, a text-driven semantic pop-up, and a multi-level details extractor. Extensive experiments on RefMatte in both keyword and expression settings validate the superiority of CLIPMat over representative methods. We hope this work could provide novel insights into image matting and encourage more follow-up studies. The dataset, code and models are available at https://github.com/JizhiziLi/RIM.

Generating synthetic images of handwritten text in a writer-specific style is a challenging task, especially in the case of unseen styles and new words, and even more when these latter contain characters that are rarely encountered during training. While emulating a writer’s style has been recently addressed by generative models, the generalization towards rare characters has been disregarded. In this work, we devise a Transformer-based model for Few-Shot styled handwritten text generation and focus on obtaining a robust and informative representation of both the text and the style. In particular, we propose a novel representation of the textual content as a sequence of dense vectors obtained from images of symbols written as standard GNU Unifont glyphs, which can be considered their visual archetypes. This strategy is more suitable for generating characters that, despite having been seen rarely during training, possibly share visual details with the frequently observed ones. As for the style, we obtain a robust representation of unseen writers’ calligraphy by exploiting specific pre-training on a large synthetic dataset. Quantitative and qualitative results demonstrate the effectiveness of our proposal in generating words in unseen styles and with rare characters more faithfully than existing approaches relying on independent one-hot encodings of the characters.

We propose a new framework for conditional image synthesis from semantic layouts of any precision levels, ranging from pure text to a 2D semantic canvas with precise shapes. More specifically, the input layout consists of one or more semantic regions with free-form text descriptions and adjustable precision levels, which can be set based on the desired controllability. The framework naturally reduces to text-to-image (T2I) at the lowest level with no shape information, and it becomes segmentation-to-image (S2I) at the highest level. By supporting the levels in-between, our framework is flexible in assisting users of different drawing expertise and at different stages of their creative workflow. We introduce several novel techniques to address the challenges coming with this new setup, including a pipeline for collecting training data; a precision-encoded mask pyramid and a text feature map representation to jointly encode precision level, semantics, and composition information; and a multi-scale guided diffusion model to synthesize images. To evaluate the proposed method, we collect a test dataset containing user-drawn layouts with diverse scenes and styles. Experimental results show that the proposed method can generate high-quality images following the layout at given precision, and compares favorably against existing methods. Project page https://zengxianyu.github.io/scenec/

Recent successes in image synthesis are powered by large-scale diffusion models. However, most methods are currently limited to either text- or image-conditioned generation for synthesizing an entire image, texture transfer or inserting objects into a user-specified region. In contrast, in this work we focus on synthesizing complex interactions (i.e., an articulated hand) with a given object. Given an RGB image of an object, we aim to hallucinate plausible images of a human hand interacting with it. We propose a two step generative approach that leverages a LayoutNet that samples an articulation-agnostic hand-object-interaction layout, and a ContentNet that synthesizes images of a hand grasping the object given the predicted layout. Both are built on top of a large-scale pretrained diffusion model to make use of its latent representation. Compared to baselines, the proposed method is shown to generalize better to novel objects and perform surprisingly well on out-of-distribution in-the-wild scenes. The resulting system allows us to predict descriptive affordance information, such as hand articulation and approaching orientation.

Recently, diffusion models have achieved great success in image synthesis. However, when it comes to the layout-to-image generation where an image often has a complex scene of multiple objects, how to make strong control over both the global layout map and each detailed object remains a challenging task. In this paper, we propose a diffusion model named LayoutDiffusion that can obtain higher generation quality and greater controllability than the previous works. To overcome the difficult multimodal fusion of image and layout, we propose to construct a structural image patch with region information and transform the patched image into a special layout to fuse with the normal layout in a unified form. Moreover, Layout Fusion Module (LFM) and Object-aware Cross Attention (OaCA) are proposed to model the relationship among multiple objects and designed to be object-aware and position-sensitive, allowing for precisely controlling the spatial related information. Extensive experiments show that our LayoutDiffusion outperforms the previous SOTA methods on FID, CAS by relatively 46.35%, 26.70% on COCO-stuff and 44.29%, 41.82% on VG. Code is available at https://github.com/ZGCTroy/LayoutDiffusion.

Large text-to-image models achieved a remarkable leap in the evolution of AI, enabling high-quality and diverse synthesis of images from a given text prompt. However, these models lack the ability to mimic the appearance of subjects in a given reference set and synthesize novel renditions of them in different contexts. In this work, we present a new approach for “personalization” of text-to-image diffusion models. Given as input just a few images of a subject, we fine-tune a pretrained text-to-image model such that it learns to bind a unique identifier with that specific subject. Once the subject is embedded in the output domain of the model, the unique identifier can be used to synthesize novel photorealistic images of the subject contextualized in different scenes. By leveraging the semantic prior embedded in the model with a new autogenous class-specific prior preservation loss, our technique enables synthesizing the subject in diverse scenes, poses, views and lighting conditions that do not appear in the reference images. We apply our technique to several previously-unassailable tasks, including subject recontextualization, text-guided view synthesis, and artistic rendering, all while preserving the subject’s key features. We also provide a new dataset and evaluation protocol for this new task of subject-driven generation. Project page: https://dreambooth.github.io/

Large-scale text-to-image diffusion models have made amazing advances. However, the status quo is to use text input alone, which can impede controllability. In this work, we propose GLIGEN: Open-Set Grounded Text-to-Image Generation, a novel approach that builds upon and extends the functionality of existing pre-trained text-to-image diffusion models by enabling them to also be conditioned on grounding inputs. To preserve the vast concept knowledge of the pre-trained model, we freeze all of its weights and inject the grounding information into new trainable layers via a gated mechanism. Our model achieves open-world grounded text2img generation with caption and bounding box condition inputs, and the grounding ability generalizes well to novel spatial configurations and concepts. GLIGEN’s zero-shot performance on COCO and LVIS outperforms existing supervised layout-to-image baselines by a large margin.

Text-conditioned image generation models have recently achieved astonishing results in image quality and text alignment and are consequently employed in a fast-growing number of applications. Since they are highly data-driven, relying on billion-sized datasets randomly scraped from the internet, they also suffer, as we demonstrate, from degenerated and biased human behavior. In turn, they may even reinforce such biases. To help combat these undesired side effects, we present safe latent diffusion (SLD). Specifically, to measure the inappropriate degeneration due to unfiltered and imbalanced training sets, we establish a novel image generation test bed - inappropriate image prompts (I2P) - containing dedicated, real-world image-to-text prompts covering concepts such as nudity and violence. As our exhaustive empirical evaluation demonstrates, the introduced SLD removes and suppresses inappropriate image parts during the diffusion process, with no additional training required and no adverse effect on overall image quality or text alignment.

Finding an initial noise vector that produces an input image when fed into the diffusion process (known as inversion) is an important problem in denoising diffusion models (DDMs), with applications for real image editing. The standard approach for real image editing with inversion uses denoising diffusion implicit models (DDIMs) to deterministically noise the image to the intermediate state along the path that the denoising would follow given the original conditioning. However, DDIM inversion for real images is unstable as it relies on local linearization assumptions, which result in the propagation of errors, leading to incorrect image reconstruction and loss of content. To alleviate these problems, we propose Exact Diffusion Inversion via Coupled Transformations (EDICT), an inversion method that draws inspiration from affine coupling layers. EDICT enables mathematically exact inversion of real and model-generated images by maintaining two coupled noise vectors which are used to invert each other in an alternating fashion. Using Stable Diffusion [25], a state-of-the-art latent diffusion model, we demonstrate that EDICT successfully reconstructs real images with high fidelity. On complex image datasets like MS-COCO, EDICT reconstruction significantly outperforms DDIM, improving the mean square error of reconstruction by a factor of two. Using noise vectors inverted from real images, EDICT enables a wide range of image edits--from local and global semantic edits to image stylization--while maintaining fidelity to the original image structure. EDICT requires no model training/finetuning, prompt tuning, or extra data and can be combined with any pretrained DDM.

Diffusion models have emerged as the new state-of-the-art generative model with high quality samples, with intriguing properties such as mode coverage and high flexibility. They have also been shown to be effective inverse problem solvers, acting as the prior of the distribution, while the information of the forward model can be granted at the sampling stage. Nonetheless, as the generative process remains in the same high dimensional (i.e. identical to data dimension) space, the models have not been extended to 3D inverse problems due to the extremely high memory and computational cost. In this paper, we combine the ideas from the conventional model-based iterative reconstruction with the modern diffusion models, which leads to a highly effective method for solving 3D medical image reconstruction tasks such as sparse-view tomography, limited angle tomography, compressed sensing MRI from pre-trained 2D diffusion models. In essence, we propose to augment the 2D diffusion prior with a model-based prior in the remaining direction at test time, such that one can achieve coherent reconstructions across all dimensions. Our method can be run in a single commodity GPU, and establishes the new state-of-the-art, showing that the proposed method can perform reconstructions of high fidelity and accuracy even in the most extreme cases (e.g. 2-view 3D tomography). We further reveal that the generalization capacity of the proposed method is surprisingly high, and can be used to reconstruct volumes that are entirely different from the training dataset. Code available: https://github.com/HJ-harry/DiffusionMBIR

Despite the visually-pleasing results achieved, the massive computational cost has been a long-standing flaw for diffusion probabilistic models~(DPMs), which, in turn, greatly limits their applications on resource-limited platforms. Prior methods towards efficient DPM, however, have largely focused on accelerating the testing yet overlooked their huge complexity and size. In this paper, we make a dedicated attempt to lighten DPM while striving to preserve its favourable performance. We start by training a small-sized latent diffusion model~(LDM) from scratch but observe a significant fidelity drop in the synthetic images. Through a thorough assessment, we find that DPM is intrinsically biased against high-frequency generation, and learns to recover different frequency components at different time-steps. These properties make compact networks unable to represent frequency dynamics with accurate high-frequency estimation. Towards this end, we introduce a customized design for slim DPM, which we term as Spectral Diffusion~(SD), for lightweight image synthesis. SD incorporates wavelet gating in its architecture to enable frequency dynamic feature extraction at every reverse steps, and conducts spectrum-aware distillation to promote high-frequency recovery by inverse weighting the objective based on spectrum magnitudes. Experimental results demonstrate that, SD achieves 8-18x computational complexity reduction as compared to the latent diffusion models on a series of conditional and unconditional image generation tasks while retaining competitive image fidelity.

Latent Diffusion Models (LDMs) enable high-quality image synthesis while avoiding excessive compute demands by training a diffusion model in a compressed lower-dimensional latent space. Here, we apply the LDM paradigm to high-resolution video generation, a particularly resource-intensive task. We first pre-train an LDM on images only; then, we turn the image generator into a video generator by introducing a temporal dimension to the latent space diffusion model and fine-tuning on encoded image sequences, i.e., videos. Similarly, we temporally align diffusion model upsamplers, turning them into temporally consistent video super resolution models. We focus on two relevant real-world applications: Simulation of in-the-wild driving data and creative content creation with text-to-video modeling. In particular, we validate our Video LDM on real driving videos of resolution 512x1024, achieving state-of-the-art performance. Furthermore, our approach can easily leverage off-the-shelf pre-trained image LDMs, as we only need to train a temporal alignment model in that case. Doing so, we turn the publicly available, state-of-the-art text-to-image LDM Stable Diffusion into an efficient and expressive text-to-video model with resolution up to 1280x2048. We show that the temporal layers trained in this way generalize to different fine-tuned text-to-image LDMs. Utilizing this property, we show the first results for personalized text-to-video generation, opening exciting directions for future content creation. Project page: https://nv-tlabs.github.io/VideoLDM/

In this paper, we show that a binary latent space can be explored for compact yet expressive image representations. We model the bi-directional mappings between an image and the corresponding latent binary representation by training an auto-encoder with a Bernoulli encoding distribution. On the one hand, the binary latent space provides a compact discrete image representation of which the distribution can be modeled more efficiently than pixels or continuous latent representations. On the other hand, we now represent each image patch as a binary vector instead of an index of a learned cookbook as in discrete image representations with vector quantization. In this way, we obtain binary latent representations that allow for better image quality and high-resolution image representations without any multi-stage hierarchy in the latent space. In this binary latent space, images can now be generated effectively using a binary latent diffusion model tailored specifically for modeling the prior over the binary image representations. We present both conditional and unconditional image generation experiments with multiple datasets, and show that the proposed method performs comparably to state-of-the-art methods while dramatically improving the sampling efficiency to as few as 16 steps without using any test-time acceleration. The proposed framework can also be seamlessly scaled to 1024 × 1024 high-resolution image generation without resorting to latent hierarchy or multi-stage refinements.

Deep learning-based video inpainting has yielded promising results and gained increasing attention from researchers. Generally, these methods usually assume that the corrupted region masks of each frame are known and easily obtained. However, the annotation of these masks are labor-intensive and expensive, which limits the practical application of current methods. Therefore, we expect to relax this assumption by defining a new semi-supervised inpainting setting, making the networks have the ability of completing the corrupted regions of the whole video using the annotated mask of only one frame. Specifically, in this work, we propose an end-to-end trainable framework consisting of completion network and mask prediction network, which are designed to generate corrupted contents of the current frame using the known mask and decide the regions to be filled of the next frame, respectively. Besides, we introduce a cycle consistency loss to regularize the training parameters of these two networks. In this way, the completion network and the mask prediction network can constrain each other, and hence the overall performance of the trained model can be maximized. Furthermore, due to the natural existence of prior knowledge (e.g., corrupted contents and clear borders), current video inpainting datasets are not suitable in the context of semi-supervised video inpainting. Thus, we create a new dataset by simulating the corrupted video of real-world scenarios. Extensive experimental results are reported to demonstrate the superiority of our model in the video inpainting task. Remarkably, although our model is trained in a semi-supervised manner, it can achieve comparable performance as fully-supervised methods.

Existing vector quantization (VQ) based autoregressive models follow a two-stage generation paradigm that first learns a codebook to encode images as discrete codes, and then completes generation based on the learned codebook. However, they encode fixed-size image regions into fixed-length codes and ignore their naturally different information densities, which results in insufficiency in important regions and redundancy in unimportant ones, and finally degrades the generation quality and speed. Moreover, the fixed-length coding leads to an unnatural raster-scan autoregressive generation. To address the problem, we propose a novel two-stage framework: (1) Dynamic-Quantization VAE (DQ-VAE) which encodes image regions into variable-length codes based on their information densities for an accurate & compact code representation. (2) DQ-Transformer which thereby generates images autoregressively from coarse-grained (smooth regions with fewer codes) to fine-grained (details regions with more codes) by modeling the position and content of codes in each granularity alternately, through a novel stacked-transformer architecture and shared-content, non-shared position input layers designs. Comprehensive experiments on various generation tasks validate our superiorities in both effectiveness and efficiency.

Video steganography is the art of unobtrusively concealing secret data in a cover video and then recovering the secret data through a decoding protocol at the receiver end. Although several attempts have been made, most of them are limited to low-capacity and fixed steganography. To rectify these weaknesses, we propose a Large-capacity and Flexible Video Steganography Network (LF-VSN) in this paper. For large-capacity, we present a reversible pipeline to perform multiple videos hiding and recovering through a single invertible neural network (INN). Our method can hide/recover 7 secret videos in/from 1 cover video with promising performance. For flexibility, we propose a key-controllable scheme, enabling different receivers to recover particular secret videos from the same cover video through specific keys. Moreover, we further improve the flexibility by proposing a scalable strategy in multiple videos hiding, which can hide variable numbers of secret videos in a cover video with a single model and a single training session. Extensive experiments demonstrate that with the significant improvement of the video steganography performance, our proposed LF-VSN has high security, large hiding capacity, and flexibility. The source code is available at https://github.com/MC-E/LF-VSN.

For any video codecs, the coding efficiency highly relies on whether the current signal to be encoded can find the relevant contexts from the previous reconstructed signals. Traditional codec has verified more contexts bring substantial coding gain, but in a time-consuming manner. However, for the emerging neural video codec (NVC), its contexts are still limited, leading to low compression ratio. To boost NVC, this paper proposes increasing the context diversity in both temporal and spatial dimensions. First, we guide the model to learn hierarchical quality patterns across frames, which enriches long-term and yet high-quality temporal contexts. Furthermore, to tap the potential of optical flow-based coding framework, we introduce a group-based offset diversity where the cross-group interaction is proposed for better context mining. In addition, this paper also adopts a quadtree-based partition to increase spatial context diversity when encoding the latent representation in parallel. Experiments show that our codec obtains 23.5% bitrate saving over previous SOTA NVC. Better yet, our codec has surpassed the under-developing next generation traditional codec/ECM in both RGB and YUV420 colorspaces, in terms of PSNR. The codes are at https://github.com/microsoft/DCVC.

Video semantic segmentation (VSS) is a computationally expensive task due to the per-frame prediction for videos of high frame rates. In recent work, compact models or adaptive network strategies have been proposed for efficient VSS. However, they did not consider a crucial factor that affects the computational cost from the input side: the input resolution. In this paper, we propose an altering resolution framework called AR-Seg for compressed videos to achieve efficient VSS. AR-Seg aims to reduce the computational cost by using low resolution for non-keyframes. To prevent the performance degradation caused by downsampling, we design a Cross Resolution Feature Fusion (CReFF) module, and supervise it with a novel Feature Similarity Training (FST) strategy. Specifically, CReFF first makes use of motion vectors stored in a compressed video to warp features from high-resolution keyframes to low-resolution non-keyframes for better spatial alignment, and then selectively aggregates the warped features with local attention mechanism. Furthermore, the proposed FST supervises the aggregated features with high-resolution features through an explicit similarity loss and an implicit constraint from the shared decoding layer. Extensive experiments on CamVid and Cityscapes show that AR-Seg achieves state-of-the-art performance and is compatible with different segmentation backbones. On CamVid, AR-Seg saves 67% computational cost (measured in GFLOPs) with the PSPNet18 backbone while maintaining high segmentation accuracy. Code: https://github.com/THU-LYJ-Lab/AR-Seg.

The high computational costs of video super-resolution (VSR) models hinder their deployment on resource-limited devices, e.g., smartphones and drones. Existing VSR models contain considerable redundant filters, which drag down the inference efficiency. To prune these unimportant filters, we develop a structured pruning scheme called Structured Sparsity Learning (SSL) according to the properties of VSR. In SSL, we design pruning schemes for several key components in VSR models, including residual blocks, recurrent networks, and upsampling networks. Specifically, we develop a Residual Sparsity Connection (RSC) scheme for residual blocks of recurrent networks to liberate pruning restrictions and preserve the restoration information. For upsampling networks, we design a pixel-shuffle pruning scheme to guarantee the accuracy of feature channel-space conversion. In addition, we observe that pruning error would be amplified as the hidden states propagate along with recurrent networks. To alleviate the issue, we design Temporal Finetuning (TF). Extensive experiments show that SSL can significantly outperform recent methods quantitatively and qualitatively. The code is available at https://github.com/Zj-BinXia/SSL.

We propose DisCo-CLIP, a distributed memory-efficient CLIP training approach, to reduce the memory consumption of contrastive loss when training contrastive learning models. Our approach decomposes the contrastive loss and its gradient computation into two parts, one to calculate the intra-GPU gradients and the other to compute the inter-GPU gradients. According to our decomposition, only the intra-GPU gradients are computed on the current GPU, while the inter-GPU gradients are collected via all_reduce from other GPUs instead of being repeatedly computed on every GPU. In this way, we can reduce the GPU memory consumption of contrastive loss computation from O(B^2) to O(B^2 / N), where B and N are the batch size and the number of GPUs used for training. Such a distributed solution is mathematically equivalent to the original non-distributed contrastive loss computation, without sacrificing any computation accuracy. It is particularly efficient for large-batch CLIP training. For instance, DisCo-CLIP can enable contrastive training of a ViT-B/32 model with a batch size of 32K or 196K using 8 or 64 A100 40GB GPUs, compared with the original CLIP solution which requires 128 A100 40GB GPUs to train a ViT-B/32 model with a batch size of 32K.

The transformer extends its success from the language to the vision domain. Because of the numerous stacked self-attention and cross-attention blocks in the transformer, which involve many high-dimensional tensor multiplication operations, the acceleration deployment of vision transformer on GPU hardware is challenging and also rarely studied. This paper thoroughly designs a compression scheme to maximally utilize the GPU-friendly 2:4 fine-grained structured sparsity and quantization. Specially, an original large model with dense weight parameters is first pruned into a sparse one by 2:4 structured pruning, which considers the GPU’s acceleration of 2:4 structured sparse pattern with FP16 data type, then the floating-point sparse model is further quantized into a fixed-point one by sparse-distillation-aware quantization aware training, which considers GPU can provide an extra speedup of 2:4 sparse calculation with integer tensors. A mixed-strategy knowledge distillation is used during the pruning and quantization process. The proposed compression scheme is flexible to support supervised and unsupervised learning styles. Experiment results show GPUSQ-ViT scheme achieves state-of-the-art compression by reducing vision transformer models 6.4-12.7 times on model size and 30.3-62 times on FLOPs with negligible accuracy degradation on ImageNet classification, COCO detection and ADE20K segmentation benchmarking tasks. Moreover, GPUSQ-ViT can boost actual deployment performance by 1.39-1.79 times and 3.22-3.43 times of latency and throughput on A100 GPU, and 1.57-1.69 times and 2.11-2.51 times improvement of latency and throughput on AGX Orin.

Vision transformers (ViT) have shown promise in various vision tasks while the U-Net based on a convolutional neural network (CNN) remains dominant in diffusion models. We design a simple and general ViT-based architecture (named U-ViT) for image generation with diffusion models. U-ViT is characterized by treating all inputs including the time, condition and noisy image patches as tokens and employing long skip connections between shallow and deep layers. We evaluate U-ViT in unconditional and class-conditional image generation, as well as text-to-image generation tasks, where U-ViT is comparable if not superior to a CNN-based U-Net of a similar size. In particular, latent diffusion models with U-ViT achieve record-breaking FID scores of 2.29 in class-conditional image generation on ImageNet 256x256, and 5.48 in text-to-image generation on MS-COCO, among methods without accessing large external datasets during the training of generative models. Our results suggest that, for diffusion-based image modeling, the long skip connection is crucial while the down-sampling and up-sampling operators in CNN-based U-Net are not always necessary. We believe that U-ViT can provide insights for future research on backbones in diffusion models and benefit generative modeling on large scale cross-modality datasets.

Vision Transformers (ViT) have shown competitive advantages in terms of performance compared to convolutional neural networks (CNNs), though they often come with high computational costs. To this end, previous methods explore different attention patterns by limiting a fixed number of spatially nearby tokens to accelerate the ViT’s multi-head self-attention (MHSA) operations. However, such structured attention patterns limit the token-to-token connections to their spatial relevance, which disregards learned semantic connections from a full attention mask. In this work, we propose an approach to learn instance-dependent attention patterns, by devising a lightweight connectivity predictor module that estimates the connectivity score of each pair of tokens. Intuitively, two tokens have high connectivity scores if the features are considered relevant either spatially or semantically. As each token only attends to a small number of other tokens, the binarized connectivity masks are often very sparse by nature and therefore provide the opportunity to reduce network FLOPs via sparse computations. Equipped with the learned unstructured attention pattern, sparse attention ViT (Sparsifiner) produces a superior Pareto frontier between FLOPs and top-1 accuracy on ImageNet compared to token sparsity. Our method reduces 48% ~ 69% FLOPs of MHSA while the accuracy drop is within 0.4%. We also show that combining attention and token sparsity reduces ViT FLOPs by over 60%.

Vision transformer has achieved impressive performance for many vision tasks. However, it may suffer from high redundancy in capturing local features for shallow layers. Local self-attention or early-stage convolutions are thus utilized, which sacrifice the capacity to capture long-range dependency. A challenge then arises: can we access efficient and effective global context modeling at the early stages of a neural network? To address this issue, we draw inspiration from the design of superpixels, which reduces the number of image primitives in subsequent processing, and introduce super tokens into vision transformer. Super tokens attempt to provide a semantically meaningful tessellation of visual content, thus reducing the token number in self-attention as well as preserving global modeling. Specifically, we propose a simple yet strong super token attention (STA) mechanism with three steps: the first samples super tokens from visual tokens via sparse association learning, the second performs self-attention on super tokens, and the last maps them back to the original token space. STA decomposes vanilla global attention into multiplications of a sparse association map and a low-dimensional attention, leading to high efficiency in capturing global dependencies. Based on STA, we develop a hierarchical vision transformer. Extensive experiments demonstrate its strong performance on various vision tasks. In particular, it achieves 86.4% top-1 accuracy on ImageNet-1K without any extra training data or label, 53.9 box AP and 46.8 mask AP on the COCO detection task, and 51.9 mIOU on the ADE20K semantic segmentation task.

In this paper, we focus on analyzing and improving the dropout technique for self-attention layers of Vision Transformer, which is important while surprisingly ignored by prior works. In particular, we conduct researches on three core questions: First, what to drop in self-attention layers? Different from dropping attention weights in literature, we propose to move dropout operations forward ahead of attention matrix calculation and set the Key as the dropout unit, yielding a novel dropout-before-softmax scheme. We theoretically verify that this scheme helps keep both regularization and probability features of attention weights, alleviating the overfittings problem to specific patterns and enhancing the model to globally capture vital information; Second, how to schedule the drop ratio in consecutive layers? In contrast to exploit a constant drop ratio for all layers, we present a new decreasing schedule that gradually decreases the drop ratio along the stack of self-attention layers. We experimentally validate the proposed schedule can avoid overfittings in low-level features and missing in high-level semantics, thus improving the robustness and stableness of model training; Third, whether need to perform structured dropout operation as CNN? We attempt patch-based block-version of dropout operation and find that this useful trick for CNN is not essential for ViT. Given exploration on the above three questions, we present the novel DropKey method that regards Key as the drop unit and exploits decreasing schedule for drop ratio, improving ViTs in a general way. Comprehensive experiments demonstrate the effectiveness of DropKey for various ViT architectures, e.g. T2T, VOLO, CeiT and DeiT, as well as for various vision tasks, e.g., image classification, object detection, human-object interaction detection and human body shape recovery.

Decoding visual stimuli from brain recordings aims to deepen our understanding of the human visual system and build a solid foundation for bridging human and computer vision through the Brain-Computer Interface. However, reconstructing high-quality images with correct semantics from brain recordings is a challenging problem due to the complex underlying representations of brain signals and the scarcity of data annotations. In this work, we present MinD-Vis: Sparse Masked Brain Modeling with Double-Conditioned Latent Diffusion Model for Human Vision Decoding. Firstly, we learn an effective self-supervised representation of fMRI data using mask modeling in a large latent space inspired by the sparse coding of information in the primary visual cortex. Then by augmenting a latent diffusion model with double-conditioning, we show that MinD-Vis can reconstruct highly plausible images with semantically matching details from brain recordings using very few paired annotations. We benchmarked our model qualitatively and quantitatively; the experimental results indicate that our method outperformed state-of-the-art in both semantic mapping (100-way semantic classification) and generation quality (FID) by 66% and 41% respectively. An exhaustive ablation study was also conducted to analyze our framework.

Vision Transformers (ViTs) have achieved overwhelming success, yet they suffer from vulnerable resolution scalability, i.e., the performance drops drastically when presented with input resolutions that are unseen during training. We introduce, ResFormer, a framework that is built upon the seminal idea of multi-resolution training for improved performance on a wide spectrum of, mostly unseen, testing resolutions. In particular, ResFormer operates on replicated images of different resolutions and enforces a scale consistency loss to engage interactive information across different scales. More importantly, to alternate among varying resolutions effectively, especially novel ones in testing, we propose a global-local positional embedding strategy that changes smoothly conditioned on input sizes. We conduct extensive experiments for image classification on ImageNet. The results provide strong quantitative evidence that ResFormer has promising scaling abilities towards a wide range of resolutions. For instance, ResFormer- B-MR achieves a Top-1 accuracy of 75.86% and 81.72% when evaluated on relatively low and high resolutions respectively (i.e., 96 and 640), which are 48% and 7.49% better than DeiT-B. We also demonstrate, moreover, ResFormer is flexible and can be easily extended to semantic segmentation, object detection and video action recognition.

Masked Autoencoders (MAE) have been prevailing paradigms for large-scale vision representation pre-training. By reconstructing masked image patches from a small portion of visible image regions, MAE forces the model to infer semantic correlation within an image. Recently, some approaches apply semantic-rich teacher models to extract image features as the reconstruction target, leading to better performance. However, unlike the low-level features such as pixel values, we argue the features extracted by powerful teacher models already encode rich semantic correlation across regions in an intact image. This raises one question: is reconstruction necessary in Masked Image Modeling (MIM) with a teacher model? In this paper, we propose an efficient MIM paradigm named MaskAlign. MaskAlign simply learns the consistency of visible patch feature extracted by the student model and intact image features extracted by the teacher model. To further advance the performance and tackle the problem of input inconsistency between the student and teacher model, we propose a Dynamic Alignment (DA) module to apply learnable alignment. Our experimental results demonstrate that masked modeling does not lose effectiveness even without reconstruction on masked regions. Combined with Dynamic Alignment, MaskAlign can achieve state-of-the-art performance with much higher efficiency.

Masked Autoencoder (MAE) has demonstrated superior performance on various vision tasks via randomly masking image patches and reconstruction. However, effective data augmentation strategies for MAE still remain open questions, different from those in contrastive learning that serve as the most important part. This paper studies the prevailing mixing augmentation for MAE. We first demonstrate that naive mixing will in contrast degenerate model performance due to the increase of mutual information (MI). To address, we propose homologous recognition, an auxiliary pretext task, not only to alleviate the MI increasement by explicitly requiring each patch to recognize homologous patches, but also to perform object-aware self-supervised pre-training for better downstream dense perception performance. With extensive experiments, we demonstrate that our proposed Mixed Autoencoder (MixedAE) achieves the state-of-the-art transfer results among masked image modeling (MIM) augmentations on different downstream tasks with significant efficiency. Specifically, our MixedAE outperforms MAE by +0.3% accuracy, +1.7 mIoU and +0.9 AP on ImageNet-1K, ADE20K and COCO respectively with a standard ViT-Base. Moreover, MixedAE surpasses iBOT, a strong MIM method combined with instance discrimination, while accelerating training by 2x. To our best knowledge, this is the very first work to consider mixing for MIM from the perspective of pretext task design. Code will be made available.

Due to the modality gap between visible and infrared images with high visual ambiguity, learning diverse modality-shared semantic concepts for visible-infrared person re-identification (VI-ReID) remains a challenging problem. Body shape is one of the significant modality-shared cues for VI-ReID. To dig more diverse modality-shared cues, we expect that erasing body-shape-related semantic concepts in the learned features can force the ReID model to extract more and other modality-shared features for identification. To this end, we propose shape-erased feature learning paradigm that decorrelates modality-shared features in two orthogonal subspaces. Jointly learning shape-related feature in one subspace and shape-erased features in the orthogonal complement achieves a conditional mutual information maximization between shape-erased feature and identity discarding body shape information, thus enhancing the diversity of the learned representation explicitly. Extensive experiments on SYSU-MM01, RegDB, and HITSZ-VCM datasets demonstrate the effectiveness of our method.

We present G-MSM (Graph-based Multi-Shape Matching), a novel unsupervised learning approach for non-rigid shape correspondence. Rather than treating a collection of input poses as an unordered set of samples, we explicitly model the underlying shape data manifold. To this end, we propose an adaptive multi-shape matching architecture that constructs an affinity graph on a given set of training shapes in a self-supervised manner. The key idea is to combine putative, pairwise correspondences by propagating maps along shortest paths in the underlying shape graph. During training, we enforce cycle-consistency between such optimal paths and the pairwise matches which enables our model to learn topology-aware shape priors. We explore different classes of shape graphs and recover specific settings, like template-based matching (star graph) or learnable ranking/sorting (TSP graph), as special cases in our framework. Finally, we demonstrate state-of-the-art performance on several recent shape correspondence benchmarks, including real-world 3D scan meshes with topological noise and challenging inter-class pairs.

The goal of click-based interactive image segmentation is to extract target masks with the input of positive/negative clicks. Every time a new click is placed, existing methods run the whole segmentation network to obtain a corrected mask, which is inefficient since several clicks may be needed to reach satisfactory accuracy. To this end, we propose an efficient method to correct the mask with a lightweight mask correction network. The whole network remains a low computational cost from the second click, even if we have a large backbone. However, a simple correction network with limited capacity is not likely to achieve comparable performance with a classic segmentation network. Thus, we propose a click-guided self-attention module and a click-guided correlation module to effectively exploits the click information to boost performance. First, several templates are selected based on the semantic similarity with click features. Then the self-attention module propagates the template information to other pixels, while the correlation module directly uses the templates to obtain target outlines. With the efficient architecture and two click-guided modules, our method shows preferable performance and efficiency compared to existing methods. The code will be released at https://github.com/feiaxyt/EMC-Click.

Current Scene Graph Generation (SGG) methods explore contextual information to predict relationships among entity pairs. However, due to the diverse visual appearance of numerous possible subject-object combinations, there is a large intra-class variation within each predicate category, e.g., “man-eating-pizza, giraffe-eating-leaf”, and the severe inter-class similarity between different classes, e.g., “man-holding-plate, man-eating-pizza”, in model’s latent space. The above challenges prevent current SGG methods from acquiring robust features for reliable relation prediction. In this paper, we claim that predicate’s categoryinherent semantics can serve as class-wise prototypes in the semantic space for relieving the above challenges caused by the diverse visual appearances. To the end, we propose the Prototype-based Embedding Network (PE-Net), which models entities/predicates with prototype-aligned compact and distinctive representations and establishes matching between entity pairs and predicates in a common embedding space for relation recognition. Moreover, Prototypeguided Learning (PL) is introduced to help PE-Net efficiently learn such entity-predicate matching, and Prototype Regularization (PR) is devised to relieve the ambiguous entity-predicate matching caused by the predicate’s semantic overlap. Extensive experiments demonstrate that our method gains superior relation recognition capability on SGG, achieving new state-of-the-art performances on both Visual Genome and Open Images datasets.

This paper proposes a new visual reasoning formulation that aims at discovering changes between image pairs and their temporal orders. Recognizing scene dynamics and their chronological orders is a fundamental aspect of human cognition. The aforementioned abilities make it possible to follow step-by-step instructions, reason about and analyze events, recognize abnormal dynamics, and restore scenes to their previous states. However, it remains unclear how well current AI systems perform in these capabilities. Although a series of studies have focused on identifying and describing changes from image pairs, they mainly consider those changes that occur synchronously, thus neglecting potential orders within those changes. To address the above issue, we first propose a visual transformation graph structure for conveying order-aware changes. Then, we benchmarked previous methods on our newly generated dataset and identified the issues of existing methods for change order recognition. Finally, we show a significant improvement in order-aware change recognition by introducing a new model that explicitly associates different changes and then identifies changes and their orders in a graph representation.

The task of dynamic scene graph generation (SGG) from videos is complicated and challenging due to the inherent dynamics of a scene, temporal fluctuation of model predictions, and the long-tailed distribution of the visual relationships in addition to the already existing challenges in image-based SGG. Existing methods for dynamic SGG have primarily focused on capturing spatio-temporal context using complex architectures without addressing the challenges mentioned above, especially the long-tailed distribution of relationships. This often leads to the generation of biased scene graphs. To address these challenges, we introduce a new framework called TEMPURA: TEmporal consistency and Memory Prototype guided UnceRtainty Attenuation for unbiased dynamic SGG. TEMPURA employs object-level temporal consistencies via transformer-based sequence modeling, learns to synthesize unbiased relationship representations using memory-guided training, and attenuates the predictive uncertainty of visual relations using a Gaussian Mixture Model (GMM). Extensive experiments demonstrate that our method achieves significant (up to 10% in some cases) performance gain over existing methods highlight- ing its superiority in generating more unbiased scene graphs. Code: https://github.com/sayaknag/unbiasedSGG.git

Cascaded computation, whereby predictions are recurrently refined over several stages, has been a persistent theme throughout the development of landmark detection models. In this work, we show that the recently proposed Deep Equilibrium Model (DEQ) can be naturally adapted to this form of computation. Our Landmark DEQ (LDEQ) achieves state-of-the-art performance on the challenging WFLW facial landmark dataset, reaching 3.92 NME with fewer parameters and a training memory cost of O(1) in the number of recurrent modules. Furthermore, we show that DEQs are particularly suited for landmark detection in videos. In this setting, it is typical to train on still images due to the lack of labelled videos. This can lead to a “flickering” effect at inference time on video, whereby a model can rapidly oscillate between different plausible solutions across consecutive frames. By rephrasing DEQs as a constrained optimization, we emulate recurrence at inference time, despite not having access to temporal data at training time. This Recurrence without Recurrence (RwR) paradigm helps in reducing landmark flicker, which we demonstrate by introducing a new metric, normalized mean flicker (NMF), and contributing a new facial landmark video dataset (WFLW-V) targeting landmark uncertainty. On the WFLW-V hard subset made up of 500 videos, our LDEQ with RwR improves the NME and NMF by 10 and 13% respectively, compared to the strongest previously published model using a hand-tuned conventional filter.

Existing Siamese tracking methods, which are built on pair-wise matching between two single frames, heavily rely on additional sophisticated mechanism to exploit temporal information among successive video frames, hindering them from high efficiency and industrial deployments. In this work, we resort to sequence-level target matching that can encode temporal contexts into the spatial features through a neat feedforward video model. Specifically, we adapt the standard video transformer architecture to visual tracking by enabling spatiotemporal feature learning directly from frame-level patch sequences. To better adapt to the tracking task, we carefully blend the spatiotemporal information in the video clips through sequential multi-branch triplet blocks, which formulates a video transformer backbone. Our experimental study compares different model variants, such as tokenization strategies, hierarchical structures, and video attention schemes. Then, we propose a disentangled dual-template mechanism that decouples static and dynamic appearance changes over time, and reduces the temporal redundancy in video frames. Extensive experiments show that our method, named as VideoTrack, achieves state-of-the-art results while running in real-time.

The appearance of an object can be fleeting when it transforms. As eggs are broken or paper is torn, their color, shape, and texture can change dramatically, preserving virtually nothing of the original except for the identity itself. Yet, this important phenomenon is largely absent from existing video object segmentation (VOS) benchmarks. In this work, we close the gap by collecting a new dataset for Video Object Segmentation under Transformations (VOST). It consists of more than 700 high-resolution videos, captured in diverse environments, which are 20 seconds long on average and densely labeled with instance masks. A careful, multi-step approach is adopted to ensure that these videos focus on complex object transformations, capturing their full temporal extent. We then extensively evaluate state-of-the-art VOS methods and make a number of important discoveries. In particular, we show that existing methods struggle when applied to this novel task and that their main limitation lies in over-reliance on static, appearance cues. This motivates us to propose a few modifications for the top-performing baseline that improve its performance by better capturing spatio-temporal information. But more broadly, the hope is to stimulate discussion on learning more robust video object representations.

Increasing scene-awareness is a key challenge in video anomaly detection (VAD). In this work, we propose a hierarchical semantic contrast (HSC) method to learn a scene-aware VAD model from normal videos. We first incorporate foreground object and background scene features with high-level semantics by taking advantage of pre-trained video parsing models. Then, building upon the autoencoder-based reconstruction framework, we introduce both scene-level and object-level contrastive learning to enforce the encoded latent features to be compact within the same semantic classes while being separable across different classes. This hierarchical semantic contrast strategy helps to deal with the diversity of normal patterns and also increases their discrimination ability. Moreover, for the sake of tackling rare normal activities, we design a skeleton-based motion augmentation to increase samples and refine the model further. Extensive experiments on three public datasets and scene-dependent mixture datasets validate the effectiveness of our proposed method.

The recent advancement in Video Instance Segmentation (VIS) has largely been driven by the use of deeper and increasingly data-hungry transformer-based models. However, video masks are tedious and expensive to annotate, limiting the scale and diversity of existing VIS datasets. In this work, we aim to remove the mask-annotation requirement. We propose MaskFreeVIS, achieving highly competitive VIS performance, while only using bounding box annotations for the object state. We leverage the rich temporal mask consistency constraints in videos by introducing the Temporal KNN-patch Loss (TK-Loss), providing strong mask supervision without any labels. Our TK-Loss finds one-to-many matches across frames, through an efficient patch-matching step followed by a K-nearest neighbor selection. A consistency loss is then enforced on the found matches. Our mask-free objective is simple to implement, has no trainable parameters, is computationally efficient, yet outperforms baselines employing, e.g., state-of-the-art optical flow to enforce temporal mask consistency. We validate MaskFreeVIS on the YouTube-VIS 2019/2021, OVIS and BDD100K MOTS benchmarks. The results clearly demonstrate the efficacy of our method by drastically narrowing the gap between fully and weakly-supervised VIS performance. Our code and trained models are available at http://vis.xyz/pub/maskfreevis.

This paper proposes a low latency neural network architecture for event-based dense prediction tasks. Conventional architectures encode entire scene contents at a fixed rate regardless of their temporal characteristics. Instead, the proposed network encodes contents at a proper temporal scale depending on its movement speed. We achieve this by constructing temporal hierarchy using stacked latent memories that operate at different rates. Given low latency event steams, the multi-level memories gradually extract dynamic to static scene contents by propagating information from the fast to the slow memory modules. The architecture not only reduces the redundancy of conventional architectures but also exploits long-term dependencies. Furthermore, an attention-based event representation efficiently encodes sparse event streams into the memory cells. We conduct extensive evaluations on three event-based dense prediction tasks, where the proposed approach outperforms the existing methods on accuracy and latency, while demonstrating effective event and image fusion capabilities. The code is available at https://hamarh.github.io/hmnet/

Tracking objects over long videos effectively means solving a spectrum of problems, from short-term association for un-occluded objects to long-term association for objects that are occluded and then reappear in the scene. Methods tackling these two tasks are often disjoint and crafted for specific scenarios, and top-performing approaches are often a mix of techniques, which yields engineering-heavy solutions that lack generality. In this work, we question the need for hybrid approaches and introduce SUSHI, a unified and scalable multi-object tracker. Our approach processes long clips by splitting them into a hierarchy of subclips, which enables high scalability. We leverage graph neural networks to process all levels of the hierarchy, which makes our model unified across temporal scales and highly general. As a result, we obtain significant improvements over state-of-the-art on four diverse datasets. Our code and models are available at bit.ly/sushi-mot.

Autoregressive transformers have shown remarkable success in video generation. However, the transformers are prohibited from directly learning the long-term dependency in videos due to the quadratic complexity of self-attention, and inherently suffering from slow inference time and error propagation due to the autoregressive process. In this paper, we propose Memory-efficient Bidirectional Transformer (MeBT) for end-to-end learning of long-term dependency in videos and fast inference. Based on recent advances in bidirectional transformers, our method learns to decode the entire spatio-temporal volume of a video in parallel from partially observed patches. The proposed transformer achieves a linear time complexity in both encoding and decoding, by projecting observable context tokens into a fixed number of latent tokens and conditioning them to decode the masked tokens through the cross-attention. Empowered by linear complexity and bidirectional modeling, our method demonstrates significant improvement over the autoregressive Transformers for generating moderately long videos in both quality and speed.

Masked visual modeling (MVM) has been recently proven effective for visual pre-training. While similar reconstructive objectives on video inputs (e.g., masked frame modeling) have been explored in video-language (VidL) pre-training, previous studies fail to find a truly effective MVM strategy that can largely benefit the downstream performance. In this work, we systematically examine the potential of MVM in the context of VidL learning. Specifically, we base our study on a fully end-to-end VIdeO-LanguagE Transformer (VIOLET), where the supervision from MVM training can be backpropagated to the video pixel space. In total, eight different reconstructive targets of MVM are explored, from low-level pixel values and oriented gradients to high-level depth maps, optical flow, discrete visual tokens, and latent visual features. We conduct comprehensive experiments and provide insights into the factors leading to effective MVM training, resulting in an enhanced model VIOLETv2. Empirically, we show VIOLETv2 pre-trained with MVM objective achieves notable improvements on 13 VidL benchmarks, ranging from video question answering, video captioning, to text-to-video retrieval.

Humans naturally perceive surrounding scenes by unifying sound and sight in a first-person view. Likewise, machines are advanced to approach human intelligence by learning with multisensory inputs from an egocentric perspective. In this paper, we explore the challenging egocentric audio-visual object localization task and observe that 1) egomotion commonly exists in first-person recordings, even within a short duration; 2) The out-of-view sound components can be created while wearers shift their attention. To address the first problem, we propose a geometry-aware temporal aggregation module to handle the egomotion explicitly. The effect of egomotion is mitigated by estimating the temporal geometry transformation and exploiting it to update visual representations. Moreover, we propose a cascaded feature enhancement module to tackle the second issue. It improves cross-modal localization robustness by disentangling visually-indicated audio representation. During training, we take advantage of the naturally available audio-visual temporal synchronization as the “free” self-supervision to avoid costly labeling. We also annotate and create the Epic Sounding Object dataset for evaluation purposes. Extensive experiments show that our method achieves state-of-the-art localization performance in egocentric videos and can be generalized to diverse audio-visual scenes.

Audiovisual automatic speech recognition (AV-ASR) aims to improve the robustness of a speech recognition system by incorporating visual information. Training fully supervised multimodal models for this task from scratch, however is limited by the need for large labelled audiovisual datasets (in each downstream domain of interest). We present AVFormer, a simple method for augmenting audioonly models with visual information, at the same time performing lightweight domain adaptation. We do this by (i) injecting visual embeddings into a frozen ASR model using lightweight trainable adaptors. We show that these can be trained on a small amount of weakly labelled video data with minimum additional training time and parameters. (ii) We also introduce a simple curriculum scheme during training which we show is crucial to enable the model to jointly process audio and visual information effectively; and finally (iii) we show that our model achieves state of the art zero-shot results on three different AV-ASR benchmarks (How2, VisSpeech and Ego4D), while also crucially preserving decent performance on traditional audio-only speech recognition benchmarks (LibriSpeech). Qualitative results show that our model effectively leverages visual information for robust speech recognition.

Active speaker detection is a challenging task in audio-visual scenarios, with the aim to detect who is speaking in one or more speaker scenarios. This task has received considerable attention because it is crucial in many applications. Existing studies have attempted to improve the performance by inputting multiple candidate information and designing complex models. Although these methods have achieved excellent performance, their high memory and computational power consumption render their application to resource-limited scenarios difficult. Therefore, in this study, a lightweight active speaker detection architecture is constructed by reducing the number of input candidates, splitting 2D and 3D convolutions for audio-visual feature extraction, and applying gated recurrent units with low computational complexity for cross-modal modeling. Experimental results on the AVA-ActiveSpeaker dataset reveal that the proposed framework achieves competitive mAP performance (94.1% vs. 94.2%), while the resource costs are significantly lower than the state-of-the-art method, particularly in model parameters (1.0M vs. 22.5M, approximately 23x) and FLOPs (0.6G vs. 2.6G, approximately 4x). Additionally, the proposed framework also performs well on the Columbia dataset, thus demonstrating good robustness. The code and model weights are available at https://github.com/Junhua-Liao/Light-ASD.

Existing audio-visual event localization (AVE) handles manually trimmed videos with only a single instance in each of them. However, this setting is unrealistic as natural videos often contain numerous audio-visual events with different categories. To better adapt to real-life applications, in this paper we focus on the task of dense-localizing audio-visual events, which aims to jointly localize and recognize all audio-visual events occurring in an untrimmed video. The problem is challenging as it requires fine-grained audio-visual scene and context understanding. To tackle this problem, we introduce the first Untrimmed Audio-Visual (UnAV-100) dataset, which contains 10K untrimmed videos with over 30K audio-visual events. Each video has 2.8 audio-visual events on average, and the events are usually related to each other and might co-occur as in real-life scenes. Next, we formulate the task using a new learning-based framework, which is capable of fully integrating audio and visual modalities to localize audio-visual events with various lengths and capture dependencies between them in a single pass. Extensive experiments demonstrate the effectiveness of our method as well as the significance of multi-scale cross-modal perception and dependency modeling for this task.

Although action recognition systems can achieve top performance when evaluated on in-distribution test points, they are vulnerable to unanticipated distribution shifts in test data. However, test-time adaptation of video action recognition models against common distribution shifts has so far not been demonstrated. We propose to address this problem with an approach tailored to spatio-temporal models that is capable of adaptation on a single video sample at a step. It consists in a feature distribution alignment technique that aligns online estimates of test set statistics towards the training statistics. We further enforce prediction consistency over temporally augmented views of the same test video sample. Evaluations on three benchmark action recognition datasets show that our proposed technique is architecture-agnostic and able to significantly boost the performance on both, the state of the art convolutional architecture TANet and the Video Swin Transformer. Our proposed method demonstrates a substantial performance gain over existing test-time adaptation approaches in both evaluations of a single distribution shift and the challenging case of random distribution shifts.

This paper simultaneously addresses three limitations associated with conventional skeleton-based action recognition; skeleton detection and tracking errors, poor variety of the targeted actions, as well as person-wise and frame-wise action recognition. A point cloud deep-learning paradigm is introduced to the action recognition, and a unified framework along with a novel deep neural network architecture called Structured Keypoint Pooling is proposed. The proposed method sparsely aggregates keypoint features in a cascaded manner based on prior knowledge of the data structure (which is inherent in skeletons), such as the instances and frames to which each keypoint belongs, and achieves robustness against input errors. Its less constrained and tracking-free architecture enables time-series keypoints consisting of human skeletons and nonhuman object contours to be efficiently treated as an input 3D point cloud and extends the variety of the targeted action. Furthermore, we propose a Pooling-Switching Trick inspired by Structured Keypoint Pooling. This trick switches the pooling kernels between the training and inference phases to detect person-wise and frame-wise actions in a weakly supervised manner using only video-level action labels. This trick enables our training scheme to naturally introduce novel data augmentation, which mixes multiple point clouds extracted from different videos. In the experiments, we comprehensively verify the effectiveness of the proposed method against the limitations, and the method outperforms state-of-the-art skeleton-based action recognition and spatio-temporal action localization methods.

Object discovery -- separating objects from the background without manual labels -- is a fundamental open challenge in computer vision. Previous methods struggle to go beyond clustering of low-level cues, whether handcrafted (e.g., color, texture) or learned (e.g., from auto-encoders). In this work, we augment the auto-encoder representation learning framework with two key components: motion-guidance and mid-level feature tokenization. Although both have been separately investigated, we introduce a new transformer decoder showing that their benefits can compound thanks to motion-guided vector quantization. We show that our architecture effectively leverages the synergy between motion and tokenization, improving upon the state of the art on both synthetic and real datasets. Our approach enables the emergence of interpretable object-specific mid-level features, demonstrating the benefits of motion-guidance (no labeling) and quantization (interpretability, memory efficiency).

Existing methods for open set action recognition focus on novelty detection that assumes video clips show a single action, which is unrealistic in the real world. We propose a new method for open set action recognition and novelty detection via MUlti-Label Evidential learning (MULE), that goes beyond previous novel action detection methods by addressing the more general problems of single or multiple actors in the same scene, with simultaneous action(s) by any actor. Our Beta Evidential Neural Network estimates multi-action uncertainty with Beta densities based on actor-context-object relation representations. An evidence debiasing constraint is added to the objective func- tion for optimization to reduce the static bias of video representations, which can incorrectly correlate predictions and static cues. We develop a primal-dual average scheme update-based learning algorithm to optimize the proposed problem and provide corresponding theoretical analysis. Besides, uncertainty and belief-based novelty estimation mechanisms are formulated to detect novel actions. Extensive experiments on two real-world video datasets show that our proposed approach achieves promising performance in single/multi-actor, single/multi-action settings. Our code and models are released at https://github.com/charliezhaoyinpeng/mule.

Weakly-supervised Temporal Action Localization (WTAL) attempts to localize the actions in untrimmed videos using only video-level supervision. Most recent works approach WTAL from a localization-by-classification perspective where these methods try to classify each video frame followed by a manually-designed post-processing pipeline to aggregate these per-frame action predictions into action snippets. Due to this perspective, the model lacks any explicit understanding of action boundaries and tends to focus only on the most discriminative parts of the video resulting in incomplete action localization. To address this, we present PivoTAL, Prior-driven Supervision for Weakly-supervised Temporal Action Localization, to approach WTAL from a localization-by-localization perspective by learning to localize the action snippets directly. To this end, PivoTAL leverages the underlying spatio-temporal regularities in videos in the form of action-specific scene prior, action snippet generation prior, and learnable Gaussian prior to supervise the localization-based training. PivoTAL shows significant improvement (of at least 3% avg mAP) over all existing methods on the benchmark datasets, THUMOS-14 and ActivitNet-v1.3.

The task of weakly supervised temporal action localization targets at generating temporal boundaries for actions of interest, meanwhile the action category should also be classified. Pseudo-label-based methods, which serve as an effective solution, have been widely studied recently. However, existing methods generate pseudo labels during training and make predictions during testing under different pipelines or settings, resulting in a gap between training and testing. In this paper, we propose to generate high-quality pseudo labels from the predicted action boundaries. Nevertheless, we note that existing post-processing, like NMS, would lead to information loss, which is insufficient to generate high-quality action boundaries. More importantly, transforming action boundaries into pseudo labels is quite challenging, since the predicted action instances are generally overlapped and have different confidence scores. Besides, the generated pseudo-labels can be fluctuating and inaccurate at the early stage of training. It might repeatedly strengthen the false predictions if there is no mechanism to conduct self-correction. To tackle these issues, we come up with an effective pipeline for learning better pseudo labels. Firstly, we propose a Gaussian weighted fusion module to preserve information of action instances and obtain high-quality action boundaries. Second, we formulate the pseudo-label generation as an optimization problem under the constraints in terms of the confidence scores of action instances. Finally, we introduce the idea of Delta pseudo labels, which enables the model with the ability of self-correction. Our method achieves superior performance to existing methods on two benchmarks, THUMOS14 and ActivityNet1.3, achieving gains of 1.9% on THUMOS14 and 3.7% on ActivityNet1.3 in terms of average mAP.

Recent research on video moment retrieval has mostly focused on enhancing the performance of accuracy, efficiency, and robustness, all of which largely rely on the abundance of high-quality annotations. While the precise frame-level annotations are time-consuming and cost-expensive, few attentions have been paid to the labeling process. In this work, we explore a new interactive manner to stimulate the process of human-in-the-loop annotation in video moment retrieval task. The key challenge is to select “ambiguous” frames and videos for binary annotations to facilitate the network training. To be specific, we propose a new hierarchical uncertainty-based modeling that explicitly considers modeling the uncertainty of each frame within the entire video sequence corresponding to the query description, and selecting the frame with the highest uncertainty. Only selected frame will be annotated by the human experts, which can largely reduce the workload. After obtaining a small number of labels provided by the expert, we show that it is sufficient to learn a competitive video moment retrieval model in such a harsh environment. Moreover, we treat the uncertainty score of frames in a video as a whole, and estimate the difficulty of each video, which can further relieve the burden of video selection. In general, our active learning strategy for video moment retrieval works not only at the frame level but also at the sequence level. Experiments on two public datasets validate the effectiveness of our proposed method.

Recently, video moment retrieval and highlight detection (MR/HD) are being spotlighted as the demand for video understanding is drastically increased. The key objective of MR/HD is to localize the moment and estimate clip-wise accordance level, i.e., saliency score, to the given text query. Although the recent transformer-based models brought some advances, we found that these methods do not fully exploit the information of a given query. For example, the relevance between text query and video contents is sometimes neglected when predicting the moment and its saliency. To tackle this issue, we introduce Query-Dependent DETR (QD-DETR), a detection transformer tailored for MR/HD. As we observe the insignificant role of a given query in transformer architectures, our encoding module starts with cross-attention layers to explicitly inject the context of text query into video representation. Then, to enhance the model’s capability of exploiting the query information, we manipulate the video-query pairs to produce irrelevant pairs. Such negative (irrelevant) video-query pairs are trained to yield low saliency scores, which in turn, encourages the model to estimate precise accordance between query-video pairs. Lastly, we present an input-adaptive saliency predictor which adaptively defines the criterion of saliency scores for the given video-query pairs. Our extensive studies verify the importance of building the query-dependent representation for MR/HD. Specifically, QD-DETR outperforms state-of-the-art methods on QVHighlights, TVSum, and Charades-STA datasets. Codes are available at github.com/wjun0830/QD-DETR.

Adopting contrastive image-text pretrained models like CLIP towards video classification has gained attention due to its cost-effectiveness and competitive performance. However, recent works in this area face a trade-off. Finetuning the pretrained model to achieve strong supervised performance results in low zero-shot generalization. Similarly, freezing the backbone to retain zero-shot capability causes significant drop in supervised accuracy. Because of this, recent works in literature typically train separate models for supervised and zero-shot action recognition. In this work, we propose a multimodal prompt learning scheme that works to balance the supervised and zero-shot performance under a single unified training. Our prompting approach on the vision side caters for three aspects: 1) Global video-level prompts to model the data distribution; 2) Local frame-level prompts to provide per-frame discriminative conditioning; and 3) a summary prompt to extract a condensed video representation. Additionally, we define a prompting scheme on the text side to augment the textual context. Through this prompting scheme, we can achieve state-of-the-art zero-shot performance on Kinetics-600, HMDB51 and UCF101 while remaining competitive in the supervised setting. By keeping the pretrained backbone frozen, we optimize a much lower number of parameters and retain the existing general representation which helps achieve the strong zero-shot performance. Our codes and models will be publicly released.

The correlation between the vision and text is essential for video moment retrieval (VMR), however, existing methods heavily rely on separate pre-training feature extractors for visual and textual understanding. Without sufficient temporal boundary annotations, it is non-trivial to learn universal video-text alignments. In this work, we explore multi-modal correlations derived from large-scale image-text data to facilitate generalisable VMR. To address the limitations of image-text pre-training models on capturing the video changes, we propose a generic method, referred to as Visual-Dynamic Injection (VDI), to empower the model’s understanding of video moments. Whilst existing VMR methods are focusing on building temporal-aware video features, being aware of the text descriptions about the temporal changes is also critical but originally overlooked in pre-training by matching static images with sentences. Therefore, we extract visual context and spatial dynamic information from video frames and explicitly enforce their alignments with the phrases describing video changes (e.g. verb). By doing so, the potentially relevant visual and motion patterns in videos are encoded in the corresponding text embeddings (injected) so to enable more accurate video-text alignments. We conduct extensive experiments on two VMR benchmark datasets (Charades-STA and ActivityNet-Captions) and achieve state-of-the-art performances. Especially, VDI yields notable advantages when being tested on the out-of-distribution splits where the testing samples involve novel scenes and vocabulary.

There is growing interest in searching for information from large video corpora. Prior works have studied relevant tasks, such as text-based video retrieval, moment retrieval, video summarization, and video captioning in isolation, without an end-to-end setup that can jointly search from video corpora and generate summaries. Such an end-to-end setup would allow for many interesting applications, e.g., a text-based search that finds a relevant video from a video corpus, extracts the most relevant moment from that video, and segments the moment into important steps with captions. To address this, we present the HiREST (HIerarchical REtrieval and STep-captioning) dataset and propose a new benchmark that covers hierarchical information retrieval and visual/textual stepwise summarization from an instructional video corpus. HiREST consists of 3.4K text-video pairs from an instructional video dataset, where 1.1K videos have annotations of moment spans relevant to text query and breakdown of each moment into key instruction steps with caption and timestamps (totaling 8.6K step captions). Our hierarchical benchmark consists of video retrieval, moment retrieval, and two novel moment segmentation and step captioning tasks. In moment segmentation, models break down a video moment into instruction steps and identify start-end boundaries. In step captioning, models generate a textual summary for each step. We also present starting point task-specific and end-to-end joint baseline models for our new benchmark. While the baseline models show some promising results, there still exists large room for future improvement by the community.

Video-language embeddings are a promising avenue for injecting semantics into visual representations, but existing methods capture only short-term associations between seconds-long video clips and their accompanying text. We propose HierVL, a novel hierarchical video-language embedding that simultaneously accounts for both long-term and short-term associations. As training data, we take videos accompanied by timestamped text descriptions of human actions, together with a high-level text summary of the activity throughout the long video (as are available in Ego4D). We introduce a hierarchical contrastive training objective that encourages text-visual alignment at both the clip level and video level. While the clip-level constraints use the step-by-step descriptions to capture what is happening in that instant, the video-level constraints use the summary text to capture why it is happening, i.e., the broader context for the activity and the intent of the actor. Our hierarchical scheme yields a clip representation that outperforms its single-level counterpart, as well as a long-term video representation that achieves SotA results on tasks requiring long-term video modeling. HierVL successfully transfers to multiple challenging downstream tasks (in EPIC-KITCHENS-100, Charades-Ego, HowTo100M) in both zero-shot and fine-tuned settings.

Pre-training on large-scale video data has become a common recipe for learning transferable spatiotemporal representations in recent years. Despite some progress, existing methods are mostly limited to highly curated datasets (e.g., K400) and exhibit unsatisfactory out-of-the-box representations. We argue that it is due to the fact that they only capture pixel-level knowledge rather than spatiotemporal semantics, which hinders further progress in video understanding. Inspired by the great success of image-text pre-training (e.g., CLIP), we take the first step to exploit language semantics to boost transferable spatiotemporal representation learning. We introduce a new pretext task, Turning to Video for Transcript Sorting (TVTS), which sorts shuffled ASR scripts by attending to learned video representations. We do not rely on descriptive captions and learn purely from video, i.e., leveraging the natural transcribed speech knowledge to provide noisy but useful semantics over time. Our method enforces the vision model to contextualize what is happening over time so that it can re-organize the narrative transcripts, and can seamlessly apply to large-scale uncurated video data in the real world. Our method demonstrates strong out-of-the-box spatiotemporal representations on diverse benchmarks, e.g., +13.6% gains over VideoMAE on SSV2 via linear probing. The code is available at https://github.com/TencentARC/TVTS.

Spatio-temporal video grounding aims to localize the aligned visual tube corresponding to a language query. Existing techniques achieve such alignment by exploiting dense boundary and bounding box annotations, which can be prohibitively expensive. To bridge the gap, we investigate the weakly-supervised setting, where models learn from easily accessible video-language data without annotations. We identify that intra-sample spurious correlations among video-language components can be alleviated if the model captures the decomposed structures of video and language data. In this light, we propose a novel framework, namely WINNER, for hierarchical video-text understanding. WINNER first builds the language decomposition tree in a bottom-up manner, upon which the structural attention mechanism and top-down feature backtracking jointly build a multi-modal decomposition tree, permitting a hierarchical understanding of unstructured videos. The multi-modal decomposition tree serves as the basis for multi-hierarchy language-tube matching. A hierarchical contrastive learning objective is proposed to learn the multi-hierarchy correspondence and distinguishment with intra-sample and inter-sample video-text decomposition structures, achieving video-language decomposition structure alignment. Extensive experiments demonstrate the rationality of our design and its effectiveness beyond state-of-the-art weakly supervised methods, even some supervised methods.

Spatio-Temporal Video Grounding (STVG) aims to localize the target object spatially and temporally according to the given language query. It is a challenging task in which the model should well understand dynamic visual cues (e.g., motions) and static visual cues (e.g., object appearances) in the language description, which requires effective joint modeling of spatio-temporal visual-linguistic dependencies. In this work, we propose a novel framework in which a static vision-language stream and a dynamic vision-language stream are developed to collaboratively reason the target tube. The static stream performs cross-modal understanding in a single frame and learns to attend to the target object spatially according to intra-frame visual cues like object appearances. The dynamic stream models visual-linguistic dependencies across multiple consecutive frames to capture dynamic cues like motions. We further design a novel cross-stream collaborative block between the two streams, which enables the static and dynamic streams to transfer useful and complementary information from each other to achieve collaborative reasoning. Experimental results show the effectiveness of the collaboration of the two streams and our overall framework achieves new state-of-the-art performance on both HCSTVG and VidSTG datasets.

The goal of this work is to understand the way actions are performed in videos. That is, given a video, we aim to predict an adverb indicating a modification applied to the action (e.g. cut “finely”). We cast this problem as a regression task. We measure textual relationships between verbs and adverbs to generate a regression target representing the action change we aim to learn. We test our approach on a range of datasets and achieve state-of-the-art results on both adverb prediction and antonym classification. Furthermore, we outperform previous work when we lift two commonly assumed conditions: the availability of action labels during testing and the pairing of adverbs as antonyms. Existing datasets for adverb recognition are either noisy, which makes learning difficult, or contain actions whose appearance is not influenced by adverbs, which makes evaluation less reliable. To address this, we collect a new high quality dataset: Adverbs in Recipes (AIR). We focus on instructional recipes videos, curating a set of actions that exhibit meaningful visual changes when performed differently. Videos in AIR are more tightly trimmed and were manually reviewed by multiple annotators to ensure high labelling quality. Results show that models learn better from AIR given its cleaner videos. At the same time, adverb prediction on AIR is challenging, demonstrating that there is considerable room for improvement.

Unified vision-language frameworks have greatly advanced in recent years, most of which adopt an encoder-decoder architecture to unify image-text tasks as sequence-to-sequence generation. However, existing video-language (VidL) models still require task-specific designs in model architecture and training objectives for each task. In this work, we explore a unified VidL framework LAVENDER, where Masked Language Modeling (MLM) is used as the common interface for all pre-training and downstream tasks. Such unification leads to a simplified model architecture, where only a lightweight MLM head, instead of a decoder with much more parameters, is needed on top of the multimodal encoder. Surprisingly, experimental results show that this unified framework achieves competitive performance on 14 VidL benchmarks, covering video question answering, text-to-video retrieval and video captioning. Extensive analyses further demonstrate LAVENDER can (i) seamlessly support all downstream tasks with just a single set of parameter values when multi-task finetuned; (ii) generalize to various downstream tasks with limited training samples; and (iii) enable zero-shot evaluation on video question answering tasks.

Understanding dense action in videos is a fundamental challenge towards the generalization of vision models. Several works show that compositionality is key to achieving generalization by combining known primitive elements, especially for handling novel composited structures. Compositional temporal grounding is the task of localizing dense action by using known words combined in novel ways in the form of novel query sentences for the actual grounding. In recent works, composition is assumed to be learned from pairs of whole videos and language embeddings through large scale self-supervised pre-training. Alternatively, one can process the video and language into word-level primitive elements, and then only learn fine-grained semantic correspondences. Both approaches do not consider the granularity of the compositions, where different query granularity corresponds to different video segments. Therefore, a good compositional representation should be sensitive to different video and query granularity. We propose a method to learn a coarse-to-fine compositional representation by decomposing the original query sentence into different granular levels, and then learning the correct correspondences between the video and recombined queries through a contrastive ranking constraint. Additionally, we run temporal boundary prediction in a coarse-to-fine manner for precise grounding boundary detection. Experiments are performed on two datasets Charades-CG and ActivityNet-CG showing the superior compositional generalizability of our approach.

Sign language recognition (SLR) is a weakly supervised task that annotates sign videos as textual glosses. Recent studies show that insufficient training caused by the lack of large-scale available sign datasets becomes the main bottleneck for SLR. Most SLR works thereby adopt pretrained visual modules and develop two mainstream solutions. The multi-stream architectures extend multi-cue visual features, yielding the current SOTA performances but requiring complex designs and might introduce potential noise. Alternatively, the advanced single-cue SLR frameworks using explicit cross-modal alignment between visual and textual modalities are simple and effective, potentially competitive with the multi-cue framework. In this work, we propose a novel contrastive visual-textual transformation for SLR, CVT-SLR, to fully explore the pretrained knowledge of both the visual and language modalities. Based on the single-cue cross-modal alignment framework, we propose a variational autoencoder (VAE) for pretrained contextual knowledge while introducing the complete pretrained language module. The VAE implicitly aligns visual and textual modalities while benefiting from pretrained contextual knowledge as the traditional contextual module. Meanwhile, a contrastive cross-modal alignment algorithm is designed to explicitly enhance the consistency constraints. Extensive experiments on public datasets (PHOENIX-2014 and PHOENIX-2014T) demonstrate that our proposed CVT-SLR consistently outperforms existing single-cue methods and even outperforms SOTA multi-cue methods.

Tracking by natural language specification aims to locate the referred target in a sequence based on the natural language description. Existing algorithms solve this issue in two steps, visual grounding and tracking, and accordingly deploy the separated grounding model and tracking model to implement these two steps, respectively. Such a separated framework overlooks the link between visual grounding and tracking, which is that the natural language descriptions provide global semantic cues for localizing the target for both two steps. Besides, the separated framework can hardly be trained end-to-end. To handle these issues, we propose a joint visual grounding and tracking framework, which reformulates grounding and tracking as a unified task: localizing the referred target based on the given visual-language references. Specifically, we propose a multi-source relation modeling module to effectively build the relation between the visual-language references and the test image. In addition, we design a temporal modeling module to provide a temporal clue with the guidance of the global semantic information for our model, which effectively improves the adaptability to the appearance variations of the target. Extensive experimental results on TNL2K, LaSOT, OTB99, and RefCOCOg demonstrate that our method performs favorably against state-of-the-art algorithms for both tracking and grounding. Code is available at https://github.com/lizhou-cs/JointNLT.

The state of the arts in vision-language pretraining (VLP) achieves exemplary performance but suffers from high training costs resulting from slow convergence and long training time, especially on large-scale web datasets. An essential obstacle to training efficiency lies in the entangled prediction rate (percentage of tokens for reconstruction) and corruption rate (percentage of corrupted tokens) in masked language modeling (MLM), that is, a proper corruption rate is achieved at the cost of a large portion of output tokens being excluded from prediction loss. To accelerate the convergence of VLP, we propose a new pretraining task, namely, free language modeling (FLM), that enables a 100% prediction rate with arbitrary corruption rates. FLM successfully frees the prediction rate from the tie-up with the corruption rate while allowing the corruption spans to be customized for each token to be predicted. FLM-trained models are encouraged to learn better and faster given the same GPU time by exploiting bidirectional contexts more flexibly. Extensive experiments show FLM could achieve an impressive 2.5x pretraining time reduction in comparison to the MLM-based methods, while keeping competitive performance on both vision-language understanding and generation tasks.

For robots to be generally useful, they must be able to find arbitrary objects described by people (i.e., be language-driven) even without expensive navigation training on in-domain data (i.e., perform zero-shot inference). We explore these capabilities in a unified setting: language-driven zero-shot object navigation (L-ZSON). Inspired by the recent success of open-vocabulary models for image classification, we investigate a straightforward framework, CLIP on Wheels (CoW), to adapt open-vocabulary models to this task without fine-tuning. To better evaluate L-ZSON, we introduce the Pasture benchmark, which considers finding uncommon objects, objects described by spatial and appearance attributes, and hidden objects described relative to visible objects. We conduct an in-depth empirical study by directly deploying 22 CoW baselines across Habitat, RoboTHOR, and Pasture. In total we evaluate over 90k navigation episodes and find that (1) CoW baselines often struggle to leverage language descriptions, but are surprisingly proficient at finding uncommon objects. (2) A simple CoW, with CLIP-based object localization and classical exploration---and no additional training---matches the navigation efficiency of a state-of-the-art ZSON method trained for 500M steps on Habitat MP3D data. This same CoW provides a 15.6 percentage point improvement in success over a state-of-the-art RoboTHOR ZSON model.

Determining the exact latitude and longitude that a photo was taken is a useful and widely applicable task, yet it remains exceptionally difficult despite the accelerated progress of other computer vision tasks. Most previous approaches have opted to learn single representations of query images, which are then classified at different levels of geographic granularity. These approaches fail to exploit the different visual cues that give context to different hierarchies, such as the country, state, and city level. To this end, we introduce an end-to-end transformer-based architecture that exploits the relationship between different geographic levels (which we refer to as hierarchies) and the corresponding visual scene information in an image through hierarchical cross-attention. We achieve this by learning a query for each geographic hierarchy and scene type. Furthermore, we learn a separate representation for different environmental scenes, as different scenes in the same location are often defined by completely different visual features. We achieve state of the art accuracy on 4 standard geo-localization datasets : Im2GPS, Im2GPS3k, YFCC4k, and YFCC26k, as well as qualitatively demonstrate how our method learns different representations for different visual hierarchies and scenes, which has not been demonstrated in the previous methods. Above previous testing datasets mostly consist of iconic landmarks or images taken from social media, which makes the dataset a simple memory task, or makes it biased towards certain places. To address this issue we introduce a much harder testing dataset, Google-World-Streets-15k, comprised of images taken from Google Streetview covering the whole planet and present state of the art results. Our code can be found at https://github.com/AHKerrigan/GeoGuessNet.

Building benchmarks to systemically analyze different capabilities of video question answering (VideoQA) models is challenging yet crucial. Existing benchmarks often use non-compositional simple questions and suffer from language biases, making it difficult to diagnose model weaknesses incisively. A recent benchmark AGQA poses a promising paradigm to generate QA pairs automatically from pre-annotated scene graphs, enabling it to measure diverse reasoning abilities with granular control. However, its questions have limitations in reasoning about the fine-grained semantics in videos as such information is absent in its scene graphs. To this end, we present ANetQA, a large-scale benchmark that supports fine-grained compositional reasoning over the challenging untrimmed videos from ActivityNet. Similar to AGQA, the QA pairs in ANetQA are automatically generated from annotated video scene graphs. The fine-grained properties of ANetQA are reflected in the following: (i) untrimmed videos with fine-grained semantics; (ii) spatio-temporal scene graphs with fine-grained taxonomies; and (iii) diverse questions generated from fine-grained templates. ANetQA attains 1.4 billion unbalanced and 13.4 million balanced QA pairs, which is an order of magnitude larger than AGQA with a similar number of videos. Comprehensive experiments are performed for state-of-the-art methods. The best model achieves 44.5% accuracy while human performance tops out at 84.5%, leaving sufficient room for improvement.

Creativity is an indispensable part of human cognition and also an inherent part of how we make sense of the world. Metaphorical abstraction is fundamental in communicating creative ideas through nuanced relationships between abstract concepts such as feelings. While computer vision benchmarks and approaches predominantly focus on understanding and generating literal interpretations of images, metaphorical comprehension of images remains relatively unexplored. Towards this goal, we introduce MetaCLUE, a set of vision tasks on visual metaphor. We also collect high-quality and rich metaphor annotations (abstract objects, concepts, relationships along with their corresponding object boxes) as there do not exist any datasets that facilitate the evaluation of these tasks. We perform a comprehensive analysis of state-of-the-art models in vision and language based on our annotations, highlighting strengths and weaknesses of current approaches in visual metaphor Classification, Localization, Understanding (retrieval, question answering, captioning) and gEneration (text-to-image synthesis) tasks. We hope this work provides a concrete step towards systematically developing AI systems with human-like creative capabilities. Project page: https://metaclue.github.io

Most existing works solving Room-to-Room VLN problem only utilize RGB images and do not consider local context around candidate views, which lack sufficient visual cues about surrounding environment. Moreover, natural language contains complex semantic information thus its correlations with visual inputs are hard to model merely with cross attention. In this paper, we propose GeoVLN, which learns Geometry-enhanced visual representation based on slot attention for robust Visual-and-Language Navigation. The RGB images are compensated with the corresponding depth maps and normal maps predicted by Omnidata as visual inputs. Technically, we introduce a two-stage module that combine local slot attention and CLIP model to produce geometry-enhanced representation from such input. We employ V&L BERT to learn a cross-modal representation that incorporate both language and vision informations. Additionally, a novel multiway attention module is designed, encouraging different phrases of input instruction to exploit the most related features from visual input. Extensive experiments demonstrate the effectiveness of our newly designed modules and show the compelling performance of the proposed method.

Text-to-motion generation is an emerging and challenging problem, which aims to synthesize motion with the same semantics as the input text. However, due to the lack of diverse labeled training data, most approaches either limit to specific types of text annotations or require online optimizations to cater to the texts during inference at the cost of efficiency and stability. In this paper, we investigate offline open-vocabulary text-to-motion generation in a zero-shot learning manner that neither requires paired training data nor extra online optimization to adapt for unseen texts. Inspired by the prompt learning in NLP, we pretrain a motion generator that learns to reconstruct the full motion from the masked motion. During inference, instead of changing the motion generator, our method reformulates the input text into a masked motion as the prompt for the motion generator to “reconstruct” the motion. In constructing the prompt, the unmasked poses of the prompt are synthesized by a text-to-pose generator. To supervise the optimization of the text-to-pose generator, we propose the first text-pose alignment model for measuring the alignment between texts and 3D poses. And to prevent the pose generator from overfitting to limited training texts, we further propose a novel wordless training mechanism that optimizes the text-to-pose generator without any training texts. The comprehensive experimental results show that our method obtains a significant improvement against the baseline methods. The code is available at https://github.com/junfanlin/oohmg.

Soft prompt learning has recently emerged as one of the methods of choice for adapting V&L models to a downstream task using a few training examples. However, current methods significantly overfit the training data, suffering from large accuracy degradation when tested on unseen classes from the same domain. To this end, in this paper, we make the following 4 contributions: (1) To alleviate base class overfitting, we propose a novel Language-Aware Soft Prompting (LASP) learning method by means of a text-to-text cross-entropy loss that maximizes the probability of the learned prompts to be correctly classified with respect to pre-defined hand-crafted textual prompts. (2) To increase the representation capacity of the prompts, we propose grouped LASP where each group of prompts is optimized with respect to a separate subset of textual prompts. (3) We identify a visual-language misalignment introduced by prompt learning and LASP, and more importantly, propose a re-calibration mechanism to address it. (4) We show that LASP is inherently amenable to including, during training, virtual classes, i.e. class names for which no visual samples are available, further increasing the robustness of the learned prompts. Through evaluations on 11 datasets, we show that our approach (a) significantly outperforms all prior works on soft prompting, and (b) matches and surpasses, for the first time, the accuracy on novel classes obtained by hand-crafted prompts and CLIP for 8 out of 11 test datasets. Code will be made available.

Vision-Language Pre-Training (VLP) has shown promising capabilities to align image and text pairs, facilitating a broad variety of cross-modal learning tasks. However, we observe that VLP models often lack the visual grounding/localization capability which is critical for many downstream tasks such as visual reasoning. In this work, we propose a novel Position-guided Text Prompt (PTP) paradigm to enhance the visual grounding ability of cross-modal models trained with VLP. Specifically, in the VLP phase, PTP divides the image into NxN blocks, and identifies the objects in each block through the widely used object detector in VLP. It then reformulates the visual grounding task into a fill-in-the-blank problem given a PTP by encouraging the model to predict the objects in the given blocks or regress the blocks of a given object, e.g. filling “P” or “O” in a PTP “The block P has a O”. This mechanism improves the visual grounding capability of VLP models and thus helps them better handle various downstream tasks. By introducing PTP into several state-of-the-art VLP frameworks, we observe consistently significant improvements across representative cross-modal learning model architectures and several benchmarks, e.g. zero-shot Flickr30K Retrieval (+4.8 in average recall@1) for ViLT baseline, and COCO Captioning (+5.3 in CIDEr) for SOTA BLIP baseline. Moreover, PTP achieves comparable results with object-detector based methods, and much faster inference speed since PTP discards its object detector for inference while the later cannot. Our code and pre-trained weight will be released.

The ability to discover abstract physical concepts and understand how they work in the world through observing lies at the core of human intelligence. The acquisition of this ability is based on compositionally perceiving the environment in terms of objects and relations in an unsupervised manner. Recent approaches learn object-centric representations and capture visually observable concepts of objects, e.g., shape, size, and location. In this paper, we take a step forward and try to discover and represent intrinsic physical concepts such as mass and charge. We introduce the PHYsical Concepts Inference NEtwork (PHYCINE), a system that infers physical concepts in different abstract levels without supervision. The key insights underlining PHYCINE are two-fold, commonsense knowledge emerges with prediction, and physical concepts of different abstract levels should be reasoned in a bottom-to-up fashion. Empirical evaluation demonstrates that variables inferred by our system work in accordance with the properties of the corresponding physical concepts. We also show that object representations containing the discovered physical concepts variables could help achieve better performance in causal reasoning tasks, i.e., COMPHY.

Multimodal semantic understanding often has to deal with uncertainty, which means the obtained messages tend to refer to multiple targets. Such uncertainty is problematic for our interpretation, including inter- and intra-modal uncertainty. Little effort has studied the modeling of this uncertainty, particularly in pre-training on unlabeled datasets and fine-tuning in task-specific downstream datasets. In this paper, we project the representations of all modalities as probabilistic distributions via a Probability Distribution Encoder (PDE) by utilizing sequence-level interactions. Compared to the exiting deterministic methods, such uncertainty modeling can convey richer multimodal semantic information and more complex relationships. Furthermore, we integrate uncertainty modeling with popular pre-training frameworks and propose suitable pre-training tasks: Distribution-based Vision-Language Contrastive learning (D-VLC), Distribution-based Masked Language Modeling (D-MLM), and Distribution-based Image-Text Matching (D-ITM). The fine-tuned models are applied to challenging downstream tasks, including image-text retrieval, visual question answering, visual reasoning, and visual entailment, and achieve state-of-the-art results.

Animal pose estimation is challenging for existing image-based methods because of limited training data and large intra- and inter-species variances. Motivated by the progress of visual-language research, we propose that pre-trained language models (eg, CLIP) can facilitate animal pose estimation by providing rich prior knowledge for describing animal keypoints in text. However, we found that building effective connections between pre-trained language models and visual animal keypoints is non-trivial since the gap between text-based descriptions and keypoint-based visual features about animal pose can be significant. To address this issue, we introduce a novel prompt-based Contrastive learning scheme for connecting Language and AniMal Pose (CLAMP) effectively. The CLAMP attempts to bridge the gap by adapting the text prompts to the animal keypoints during network training. The adaptation is decomposed into spatial-aware and feature-aware processes, and two novel contrastive losses are devised correspondingly. In practice, the CLAMP enables the first cross-modal animal pose estimation paradigm. Experimental results show that our method achieves state-of-the-art performance under the supervised, few-shot, and zero-shot settings, outperforming image-based methods by a large margin. The code is available at https://github.com/xuzhang1199/CLAMP.

Human spatial attention conveys information about theregions of visual scenes that are important for perform-ing visual tasks. Prior work has shown that the informa-tion about human attention can be leveraged to benefit var-ious supervised vision tasks. Might providing this weakform of supervision be useful for self-supervised represen-tation learning? Addressing this question requires collect-ing large datasets with human attention labels. Yet, col-lecting such large scale data is very expensive. To addressthis challenge, we construct an auxiliary teacher model topredict human attention, trained on a relatively small la-beled dataset. This teacher model allows us to generate im-age (pseudo) attention labels for ImageNet. We then traina model with a primary contrastive objective; to this stan-dard configuration, we add a simple output head trained topredict the attentional map for each image, guided by thepseudo labels from teacher model. We measure the qual-ity of learned representations by evaluating classificationperformance from the frozen learned embeddings as wellas performance on image retrieval tasks. We find that thespatial-attention maps predicted from the contrastive modeltrained with teacher guidance aligns better with human at-tention compared to vanilla contrastive models. Moreover,we find that our approach improves classification accuracyand robustness of the contrastive models on ImageNet andImageNet-C. Further, we find that model representationsbecome more useful for image retrieval task as measuredby precision-recall performance on ImageNet, ImageNet-C,CIFAR10, and CIFAR10-C datasets.

Self-supervised pretraining has achieved remarkable success in high-level vision, but its application in low-level vision remains ambiguous and not well-established. What is the primitive intention of pretraining? What is the core problem of pretraining in low-level vision? In this paper, we aim to answer these essential questions and establish a new pretraining scheme for low-level vision. Specifically, we examine previous pretraining methods in both high-level and low-level vision, and categorize current low-level vision tasks into two groups based on the difficulty of data acquisition: low-cost and high-cost tasks. Existing literature has mainly focused on pretraining for low-cost tasks, where the observed performance improvement is often limited. However, we argue that pretraining is more significant for high-cost tasks, where data acquisition is more challenging. To learn a general low-level vision representation that can improve the performance of various tasks, we propose a new pretraining paradigm called degradation autoencoder (DegAE). DegAE follows the philosophy of designing pretext task for self-supervised pretraining and is elaborately tailored to low-level vision. With DegAE pretraining, SwinIR achieves a 6.88dB performance gain on image dehaze task, while Uformer obtains 3.22dB and 0.54dB improvement on dehaze and derain tasks, respectively.

Both masked image modeling (MIM) and natural language supervision have facilitated the progress of transferable visual pre-training. In this work, we seek the synergy between two paradigms and study the emerging properties when MIM meets natural language supervision. To this end, we present a novel masked visual Reconstruction In Language semantic Space (RILS) pre-training framework, in which sentence representations, encoded by the text encoder, serve as prototypes to transform the vision-only signals into patch-sentence probabilities as semantically meaningful MIM reconstruction targets. The vision models can therefore capture useful components with structured information by predicting proper semantic of masked tokens. Better visual representations could, in turn, improve the text encoder via the image-text alignment objective, which is essential for the effective MIM target transformation. Extensive experimental results demonstrate that our method not only enjoys the best of previous MIM and CLIP but also achieves further improvements on various tasks due to their mutual benefits. RILS exhibits advanced transferability on downstream classification, detection, and segmentation, especially for low-shot regimes. Code is available at https://github.com/hustvl/RILS.

Understanding geometric concepts, such as distance and shape, is essential for understanding the real world and also for many vision tasks. To incorporate such information into a visual representation of a scene, we propose learning to represent the scene by sketching, inspired by human behavior. Our method, coined Learning by Sketching (LBS), learns to convert an image into a set of colored strokes that explicitly incorporate the geometric information of the scene in a single inference step without requiring a sketch dataset. A sketch is then generated from the strokes where CLIP-based perceptual loss maintains a semantic similarity between the sketch and the image. We show theoretically that sketching is equivariant with respect to arbitrary affine transformations and thus provably preserves geometric information. Experimental results show that LBS substantially improves the performance of object attribute classification on the unlabeled CLEVR dataset, domain transfer between CLEVR and STL-10 datasets, and for diverse downstream tasks, confirming that LBS provides rich geometric information.

This paper, for the very first time, introduces human sketches to the landscape of XAI (Explainable Artificial Intelligence). We argue that sketch as a “human-centred” data form, represents a natural interface to study explainability. We focus on cultivating sketch-specific explainability designs. This starts by identifying strokes as a unique building block that offers a degree of flexibility in object construction and manipulation impossible in photos. Following this, we design a simple explainability-friendly sketch encoder that accommodates the intrinsic properties of strokes: shape, location, and order. We then move on to define the first ever XAI task for sketch, that of stroke location inversion SLI. Just as we have heat maps for photos, and correlation matrices for text, SLI offers an explainability angle to sketch in terms of asking a network how well it can recover stroke locations of an unseen sketch. We offer qualitative results for readers to interpret as snapshots of the SLI process in the paper, and as GIFs on the project page. A minor but interesting note is that thanks to its sketch-specific design, our sketch encoder also yields the best sketch recognition accuracy to date while having the smallest number of parameters. The code is available at https://sketchxai.github.io.

While generative modeling on multimodal image-text data has been actively developed with large-scale paired datasets, there have been limited attempts to generate both image and text data by a single model rather than a generation of one fixed modality conditioned on the other modality. In this paper, we explore a unified generative vision-and-language (VL) model that can produce both images and text sequences. Especially, we propose a generative VL transformer based on the non-autoregressive mask prediction, named MAGVLT, and compare it with an autoregressive generative VL transformer (ARGVLT). In comparison to ARGVLT, the proposed MAGVLT enables bidirectional context encoding, fast decoding by parallel token predictions in an iterative refinement, and extended editing capabilities such as image and text infilling. For rigorous training of our MAGVLT with image-text pairs from scratch, we combine the image-to-text, text-to image, and joint image-and-text mask prediction tasks. Moreover, we devise two additional tasks based on the step-unrolled mask prediction and the selective prediction on the mixture of two image-text pairs. Experimental results on various downstream generation tasks of VL benchmarks show that our MAGVLT outperforms ARGVLT by a large margin even with significant inference speedup. Particularly, MAGVLT achieves competitive results on both zero-shot image-to-text and text-to-image generation tasks from MS-COCO by one moderate-sized model (fewer than 500M parameters) even without the use of monomodal data and networks.

This paper studies the problem of zero-short sketch-based image retrieval (ZS-SBIR), however with two significant differentiators to prior art (i) we tackle all variants (inter-category, intra-category, and cross datasets) of ZS-SBIR with just one network (“everything”), and (ii) we would really like to understand how this sketch-photo matching operates (“explainable”). Our key innovation lies with the realization that such a cross-modal matching problem could be reduced to comparisons of groups of key local patches -- akin to the seasoned “bag-of-words” paradigm. Just with this change, we are able to achieve both of the aforementioned goals, with the added benefit of no longer requiring external semantic knowledge. Technically, ours is a transformer-based cross-modal network, with three novel components (i) a self-attention module with a learnable tokenizer to produce visual tokens that correspond to the most informative local regions, (ii) a cross-attention module to compute local correspondences between the visual tokens across two modalities, and finally (iii) a kernel-based relation network to assemble local putative matches and produce an overall similarity metric for a sketch-photo pair. Experiments show ours indeed delivers superior performances across all ZS-SBIR settings. The all important explainable goal is elegantly achieved by visualizing cross-modal token correspondences, and for the first time, via sketch to photo synthesis by universal replacement of all matched photo patches.

Recent advances on text-to-image generation have witnessed the rise of diffusion models which act as powerful generative models. Nevertheless, it is not trivial to exploit such latent variable models to capture the dependency among discrete words and meanwhile pursue complex visual-language alignment in image captioning. In this paper, we break the deeply rooted conventions in learning Transformer-based encoder-decoder, and propose a new diffusion model based paradigm tailored for image captioning, namely Semantic-Conditional Diffusion Networks (SCD-Net). Technically, for each input image, we first search the semantically relevant sentences via cross-modal retrieval model to convey the comprehensive semantic information. The rich semantics are further regarded as semantic prior to trigger the learning of Diffusion Transformer, which produces the output sentence in a diffusion process. In SCD-Net, multiple Diffusion Transformer structures are stacked to progressively strengthen the output sentence with better visional-language alignment and linguistical coherence in a cascaded manner. Furthermore, to stabilize the diffusion process, a new self-critical sequence training strategy is designed to guide the learning of SCD-Net with the knowledge of a standard autoregressive Transformer model. Extensive experiments on COCO dataset demonstrate the promising potential of using diffusion models in the challenging image captioning task. Source code is available at https://github.com/YehLi/xmodaler/tree/master/configs/image_caption/scdnet.

In this paper, we propose an end-to-end Retrieval-Augmented Visual Language Model (REVEAL) that learns to encode world knowledge into a large-scale memory, and to retrieve from it to answer knowledge-intensive queries. REVEAL consists of four key components: the memory, the encoder, the retriever and the generator. The large-scale memory encodes various sources of multimodal world knowledge (e.g. image-text pairs, question answering pairs, knowledge graph triplets, etc.) via a unified encoder. The retriever finds the most relevant knowledge entries in the memory, and the generator fuses the retrieved knowledge with the input query to produce the output. A key novelty in our approach is that the memory, encoder, retriever and generator are all pre-trained end-to-end on a massive amount of data. Furthermore, our approach can use a diverse set of multimodal knowledge sources, which is shown to result in significant gains. We show that REVEAL achieves state-of-the-art results on visual question answering and image captioning.

We propose a text-to-image generation algorithm based on deep neural networks when text captions for images are unavailable during training. In this work, instead of simply generating pseudo-ground-truth sentences of training images using existing image captioning methods, we employ a pretrained CLIP model, which is capable of properly aligning embeddings of images and corresponding texts in a joint space and, consequently, works well on zero-shot recognition tasks. We optimize a text-to-image generation model by maximizing the data log-likelihood conditioned on pairs of image-text CLIP embeddings. To better align data in the two domains, we employ a principled way based on a variational inference, which efficiently estimates an approximate posterior of the hidden text embedding given an image and its CLIP feature. Experimental results validate that the proposed framework outperforms existing approaches by large margins under unsupervised and semi-supervised text-to-image generation settings.

We present Fast Language-Image Pre-training (FLIP), a simple and more efficient method for training CLIP. Our method randomly masks out and removes a large portion of image patches during training. Masking allows us to learn from more image-text pairs given the same wall-clock time and contrast more samples per iteration with similar memory footprint. It leads to a favorable trade-off between accuracy and training time. In our experiments on 400 million image-text pairs, FLIP improves both accuracy and speed over the no-masking baseline. On a large diversity of downstream tasks, FLIP dominantly outperforms the CLIP counterparts trained on the same data. Facilitated by the speedup, we explore the scaling behavior of increasing the model size, data size, or training length, and report encouraging results and comparisons. We hope that our work will foster future research on scaling vision-language learning.

Existing techniques for image-to-image translation commonly have suffered from two critical problems: heavy reliance on per-sample domain annotation and/or inability to handle multiple attributes per image. Recent truly-unsupervised methods adopt clustering approaches to easily provide per-sample one-hot domain labels. However, they cannot account for the real-world setting: one sample may have multiple attributes. In addition, the semantics of the clusters are not easily coupled to human understanding. To overcome these, we present LANguage-driven Image-to-image Translation model, dubbed LANIT. We leverage easy-to-obtain candidate attributes given in texts for a dataset: the similarity between images and attributes indicates per-sample domain labels. This formulation naturally enables multi-hot labels so that users can specify the target domain with a set of attributes in language. To account for the case that the initial prompts are inaccurate, we also present prompt learning. We further present domain regularization loss that enforces translated images to be mapped to the corresponding domain. Experiments on several standard benchmarks demonstrate that LANIT achieves comparable or superior performance to existing models. The code is available at github.com/KU-CVLAB/LANIT.

Despite advances in global image representation, existing image retrieval approaches rarely consider geometric structure during the global retrieval stage. In this work, we revisit the conventional self-similarity descriptor from a convolutional perspective, to encode both the visual and structural cues of the image to global image representation. Our proposed network, named Structural Embedding Network (SENet), captures the internal structure of the images and gradually compresses them into dense self-similarity descriptors while learning diverse structures from various images. These self-similarity descriptors and original image features are fused and then pooled into global embedding, so that global embedding can represent both geometric and visual cues of the image. Along with this novel structural embedding, our proposed network sets new state-of-the-art performances on several image retrieval benchmarks, convincing its robustness to look-alike distractors. The code and models are available: https://github.com/sungonce/SENet.

Cross-modal retrieval across image and text modalities is a challenging task due to its inherent ambiguity: An image often exhibits various situations, and a caption can be coupled with diverse images. Set-based embedding has been studied as a solution to this problem. It seeks to encode a sample into a set of different embedding vectors that capture different semantics of the sample. In this paper, we present a novel set-based embedding method, which is distinct from previous work in two aspects. First, we present a new similarity function called smooth-Chamfer similarity, which is designed to alleviate the side effects of existing similarity functions for set-based embedding. Second, we propose a novel set prediction module to produce a set of embedding vectors that effectively captures diverse semantics of input by the slot attention mechanism. Our method is evaluated on the COCO and Flickr30K datasets across different visual backbones, where it outperforms existing methods including ones that demand substantially larger computation at inference.

Self supervision and natural language supervision have emerged as two exciting ways to train general purpose image encoders which excel at a variety of downstream tasks. Recent works such as M3AE (Geng et al 2022) and SLIP (Mu et al 2022) have suggested that these approaches can be effectively combined, but most notably their results use small (<20M examples) pre-training datasets and don’t effectively reflect the large-scale regime (>100M samples) that is commonly used for these approaches. Here we investigate whether a similar approach can be effective when trained with a much larger amount of data. We find that a combination of two state of the art approaches: masked auto-encoders, MAE (He et al 2021) and contrastive language image pre-training, CLIP (Radford et al 2021) provides a benefit over CLIP when trained on a corpus of 11.3M image-text pairs, but little to no benefit (as evaluated on a suite of common vision tasks) over CLIP when trained on a large corpus of 1.4B images. Our work provides some much needed clarity into the effectiveness (or lack thereof) of self supervision for large-scale image-text training.

Pre-trained vision-language models have inspired much research on few-shot learning. However, with only a few training images, there exist two crucial problems: (1) the visual feature distributions are easily distracted by class-irrelevant information in images, and (2) the alignment between the visual and language feature distributions is difficult. To deal with the distraction problem, we propose a Selective Attack module, which consists of trainable adapters that generate spatial attention maps of images to guide the attacks on class-irrelevant image areas. By messing up these areas, the critical features are captured and the visual distributions of image features are calibrated. To better align the visual and language feature distributions that describe the same object class, we propose a cross-modal distribution alignment module, in which we introduce a vision-language prototype for each class to align the distributions, and adopt the Earth Mover’s Distance (EMD) to optimize the prototypes. For efficient computation, the upper bound of EMD is derived. In addition, we propose an augmentation strategy to increase the diversity of the images and the text prompts, which can reduce overfitting to the few-shot training images. Extensive experiments on 11 datasets demonstrate that our method consistently outperforms prior arts in few-shot learning.

Deep hashing is an appealing approach for large-scale image retrieval. Most existing supervised deep hashing methods learn hash functions using pairwise or triple image similarities in randomly sampled mini-batches. They suffer from low training efficiency, insufficient coverage of data distribution, and pair imbalance problems. Recently, central similarity quantization (CSQ) attacks the above problems by using “hash centers” as a global similarity metric, which encourages the hash codes of similar images to approach their common hash center and distance themselves from other hash centers. Although achieving SOTA retrieval performance, CSQ falls short of a worst-case guarantee on the minimal distance between its constructed hash centers, i.e. the hash centers can be arbitrarily close. This paper presents an optimization method that finds hash centers with a constraint on the minimal distance between any pair of hash centers, which is non-trivial due to the non-convex nature of the problem. More importantly, we adopt the Gilbert-Varshamov bound from coding theory, which helps us to obtain a large minimal distance while ensuring the empirical feasibility of our optimization approach. With these clearly-separated hash centers, each is assigned to one image class, we propose several effective loss functions to train deep hashing networks. Extensive experiments on three datasets for image retrieval demonstrate that the proposed method achieves superior retrieval performance over the state-of-the-art deep hashing methods.

Zero-shot capability has been considered as a new revolution of deep learning, letting machines work on tasks without curated training data. As a good start and the only existing outcome of zero-shot image captioning (IC), ZeroCap abandons supervised training and sequentially searching every word in the caption using the knowledge of large-scale pre-trained models. Though effective, its autoregressive generation and gradient-directed searching mechanism limit the diversity of captions and inference speed, respectively. Moreover, ZeroCap does not consider the controllability issue of zero-shot IC. To move forward, we propose a framework for Controllable Zero-shot IC, named ConZIC. The core of ConZIC is a novel sampling-based non-autoregressive language model named GibbsBERT, which can generate and continuously polish every word. Extensive quantitative and qualitative results demonstrate the superior performance of our proposed ConZIC for both zero-shot IC and controllable zero-shot IC. Especially, ConZIC achieves about 5× faster generation speed than ZeroCap, and about 1.5× higher diversity scores, with accurate generation given different control signals.

Large scale vision and language models can achieve impressive zero-shot recognition performance by mapping class specific text queries to image content. Two distinct challenges that remain however, are high sensitivity to the choice of handcrafted class names that define queries, and the difficulty of adaptation to new, smaller datasets. Towards addressing these problems, we propose to leverage available data to learn, for each class, an optimal word embedding as a function of the visual content. By learning new word embeddings on an otherwise frozen model, we are able to retain zero-shot capabilities for new classes, easily adapt models to new datasets, and adjust potentially erroneous, non-descriptive or ambiguous class names. We show that our solution can easily be integrated in image classification and object detection pipelines, yields significant performance gains in multiple scenarios and provides insights into model biases and labelling errors.

Visual place recognition (VPR) is a fundamental task of computer vision for visual localization. Existing methods are trained using image pairs that either depict the same place or not. Such a binary indication does not consider continuous relations of similarity between images of the same place taken from different positions, determined by the continuous nature of camera pose. The binary similarity induces a noisy supervision signal into the training of VPR methods, which stall in local minima and require expensive hard mining algorithms to guarantee convergence. Motivated by the fact that two images of the same place only partially share visual cues due to camera pose differences, we deploy an automatic re-annotation strategy to re-label VPR datasets. We compute graded similarity labels for image pairs based on available localization metadata. Furthermore, we propose a new Generalized Contrastive Loss (GCL) that uses graded similarity labels for training contrastive networks. We demonstrate that the use of the new labels and GCL allow to dispense from hard-pair mining, and to train image descriptors that perform better in VPR by nearest neighbor search, obtaining superior or comparable results than methods that require expensive hard-pair mining and re-ranking techniques.

This paper presents DetCLIPv2, an efficient and scalable training framework that incorporates large-scale image-text pairs to achieve open-vocabulary object detection (OVD). Unlike previous OVD frameworks that typically rely on a pre-trained vision-language model (e.g., CLIP) or exploit image-text pairs via a pseudo labeling process, DetCLIPv2 directly learns the fine-grained word-region alignment from massive image-text pairs in an end-to-end manner. To accomplish this, we employ a maximum word-region similarity between region proposals and textual words to guide the contrastive objective. To enable the model to gain localization capability while learning broad concepts, DetCLIPv2 is trained with a hybrid supervision from detection, grounding and image-text pair data under a unified data formulation. By jointly training with an alternating scheme and adopting low-resolution input for image-text pairs, DetCLIPv2 exploits image-text pair data efficiently and effectively: DetCLIPv2 utilizes 13× more image-text pairs than DetCLIP with a similar training time and improves performance. With 13M image-text pairs for pre-training, DetCLIPv2 demonstrates superior open-vocabulary detection performance, e.g., DetCLIPv2 with Swin-T backbone achieves 40.4% zero-shot AP on the LVIS benchmark, which outperforms previous works GLIP/GLIPv2/DetCLIP by 14.4/11.4/4.5% AP, respectively, and even beats its fully-supervised counterpart by a large margin.

Human-Object Interaction (HOI) detection aims to localize human-object pairs and recognize their interactions. Recently, Contrastive Language-Image Pre-training (CLIP) has shown great potential in providing interaction prior for HOI detectors via knowledge distillation. However, such approaches often rely on large-scale training data and suffer from inferior performance under few/zero-shot scenarios. In this paper, we propose a novel HOI detection framework that efficiently extracts prior knowledge from CLIP and achieves better generalization. In detail, we first introduce a novel interaction decoder to extract informative regions in the visual feature map of CLIP via a cross-attention mechanism, which is then fused with the detection backbone by a knowledge integration block for more accurate human-object pair detection. In addition, prior knowledge in CLIP text encoder is leveraged to generate a classifier by embedding HOI descriptions. To distinguish fine-grained interactions, we build a verb classifier from training data via visual semantic arithmetic and a lightweight verb representation adapter. Furthermore, we propose a training-free enhancement to exploit global HOI predictions from CLIP. Extensive experiments demonstrate that our method outperforms the state of the art by a large margin on various settings, e.g. +4.04 mAP on HICO-Det. The source code is available in https://github.com/Artanic30/HOICLIP.

In this paper, we consider the problem of simultaneously detecting objects and inferring their visual attributes in an image, even for those with no manual annotations provided at the training stage, resembling an open-vocabulary scenario. To achieve this goal, we make the following contributions: (i) we start with a naive two-stage approach for open-vocabulary object detection and attribute classification, termed CLIP-Attr. The candidate objects are first proposed with an offline RPN and later classified for semantic category and attributes; (ii) we combine all available datasets and train with a federated strategy to finetune the CLIP model, aligning the visual representation with attributes, additionally, we investigate the efficacy of leveraging freely available online image-caption pairs under weakly supervised learning; (iii) in pursuit of efficiency, we train a Faster-RCNN type model end-to-end with knowledge distillation, that performs class-agnostic object proposals and classification on semantic categories and attributes with classifiers generated from a text encoder; Finally, (iv) we conduct extensive experiments on VAW, MS-COCO, LSA, and OVAD datasets, and show that recognition of semantic category and attributes is complementary for visual scene understanding, i.e., jointly training object detection and attributes prediction largely outperform existing approaches that treat the two tasks independently, demonstrating strong generalization ability to novel attributes and categories.

This paper presents the first significant object detection framework, NeRF-RPN, which directly operates on NeRF. Given a pre-trained NeRF model, NeRF-RPN aims to detect all bounding boxes of objects in a scene. By exploiting a novel voxel representation that incorporates multi-scale 3D neural volumetric features, we demonstrate it is possible to regress the 3D bounding boxes of objects in NeRF directly without rendering the NeRF at any viewpoint. NeRF-RPN is a general framework and can be applied to detect objects without class labels. We experimented NeRF-RPN with various backbone architectures, RPN head designs, and loss functions. All of them can be trained in an end-to-end manner to estimate high quality 3D bounding boxes. To facilitate future research in object detection for NeRF, we built a new benchmark dataset which consists of both synthetic and real-world data with careful labeling and clean up. Code and dataset are available at https://github.com/lyclyc52/NeRF_RPN.

Existing instance segmentation models learn task-specific information using manual mask annotations from base (training) categories. These mask annotations require tremendous human effort, limiting the scalability to annotate novel (new) categories. To alleviate this problem, Open-Vocabulary (OV) methods leverage large-scale image-caption pairs and vision-language models to learn novel categories. In summary, an OV method learns task-specific information using strong supervision from base annotations and novel category information using weak supervision from image-captions pairs. This difference between strong and weak supervision leads to overfitting on base categories, resulting in poor generalization towards novel categories. In this work, we overcome this issue by learning both base and novel categories from pseudo-mask annotations generated by the vision-language model in a weakly supervised manner using our proposed Mask-free OVIS pipeline. Our method automatically generates pseudo-mask annotations by leveraging the localization ability of a pre-trained vision-language model for objects present in image-caption pairs. The generated pseudo-mask annotations are then used to supervise an instance segmentation model, freeing the entire pipeline from any labour-expensive instance-level annotations and overfitting. Our extensive experiments show that our method trained with just pseudo-masks significantly improves the mAP scores on the MS-COCO dataset and OpenImages dataset compared to the recent state-of-the-art methods trained with manual masks. Codes and models are provided in https://vibashan.github.io/ovis-web/.

Image-based Virtual Try-ON aims to transfer an in-shop garment onto a specific person. Existing methods employ a global warping module to model the anisotropic deformation for different garment parts, which fails to preserve the semantic information of different parts when receiving challenging inputs (e.g, intricate human poses, difficult garments). Moreover, most of them directly warp the input garment to align with the boundary of the preserved region, which usually requires texture squeezing to meet the boundary shape constraint and thus leads to texture distortion. The above inferior performance hinders existing methods from real-world applications. To address these problems and take a step towards real-world virtual try-on, we propose a General-Purpose Virtual Try-ON framework, named GP-VTON, by developing an innovative Local-Flow Global-Parsing (LFGP) warping module and a Dynamic Gradient Truncation (DGT) training strategy. Specifically, compared with the previous global warping mechanism, LFGP employs local flows to warp garments parts individually, and assembles the local warped results via the global garment parsing, resulting in reasonable warped parts and a semantic-correct intact garment even with challenging inputs.On the other hand, our DGT training strategy dynamically truncates the gradient in the overlap area and the warped garment is no more required to meet the boundary constraint, which effectively avoids the texture squeezing problem. Furthermore, our GP-VTON can be easily extended to multi-category scenario and jointly trained by using data from different garment categories. Extensive experiments on two high-resolution benchmarks demonstrate our superiority over the existing state-of-the-art methods.

Compositional Zero-Shot Learning (CZSL) aims to recognize novel concepts formed by known states and objects during training. Existing methods either learn the combined state-object representation, challenging the generalization of unseen compositions, or design two classifiers to identify state and object separately from image features, ignoring the intrinsic relationship between them. To jointly eliminate the above issues and construct a more robust CZSL system, we propose a novel framework termed Decomposed Fusion with Soft Prompt (DFSP), by involving vision-language models (VLMs) for unseen composition recognition. Specifically, DFSP constructs a vector combination of learnable soft prompts with state and object to establish the joint representation of them. In addition, a cross-modal decomposed fusion module is designed between the language and image branches, which decomposes state and object among language features instead of image features. Notably, being fused with the decomposed features, the image features can be more expressive for learning the relationship with states and objects, respectively, to improve the response of unseen compositions in the pair space, hence narrowing the domain gap between seen and unseen sets. Experimental results on three challenging benchmarks demonstrate that our approach significantly outperforms other state-of-the-art methods by large margins.

Referring image segmentation aims to segment the target referent in an image conditioning on a natural language expression. Existing one-stage methods employ per-pixel classification frameworks, which attempt straightforwardly to align vision and language at the pixel level, thus failing to capture critical object-level information. In this paper, we propose a mask classification framework, Contrastive Grouping with Transformer network (CGFormer), which explicitly captures object-level information via token-based querying and grouping strategy. Specifically, CGFormer first introduces learnable query tokens to represent objects and then alternately queries linguistic features and groups visual features into the query tokens for object-aware cross-modal reasoning. In addition, CGFormer achieves cross-level interaction by jointly updating the query tokens and decoding masks in every two consecutive layers. Finally, CGFormer cooperates contrastive learning to the grouping strategy to identify the token and its mask corresponding to the referent. Experimental results demonstrate that CGFormer outperforms state-of-the-art methods in both segmentation and generalization settings consistently and significantly. Code is available at https://github.com/Toneyaya/CGFormer.

Few-shot learning is a challenging problem since only a few examples are provided to recognize a new class. Several recent studies exploit additional semantic information, e.g. text embeddings of class names, to address the issue of rare samples through combining semantic prototypes with visual prototypes. However, these methods still suffer from the spurious visual features learned from the rare support samples, resulting in limited benefits. In this paper, we propose a novel Semantic Prompt (SP) approach for few-shot learning. Instead of the naive exploitation of semantic information for remedying classifiers, we explore leveraging semantic information as prompts to tune the visual feature extraction network adaptively. Specifically, we design two complementary mechanisms to insert semantic prompts into the feature extractor: one is to enable the interaction between semantic prompts and patch embeddings along the spatial dimension via self-attention, another is to supplement visual features with the transformed semantic prompts along the channel dimension. By combining these two mechanisms, the feature extractor presents a better ability to attend to the class-specific features and obtains more generalized image representations with merely a few support samples. Through extensive experiments on four datasets, the proposed approach achieves promising results, improving the 1-shot learning accuracy by 3.67% on average.

Referring Expression Segmentation (RES) aims to generate a segmentation mask for the object described by a given language expression. Existing classic RES datasets and methods commonly support single-target expressions only, i.e., one expression refers to one target object. Multi-target and no-target expressions are not considered. This limits the usage of RES in practice. In this paper, we introduce a new benchmark called Generalized Referring Expression Segmentation (GRES), which extends the classic RES to allow expressions to refer to an arbitrary number of target objects. Towards this, we construct the first large-scale GRES dataset called gRefCOCO that contains multi-target, no-target, and single-target expressions. GRES and gRefCOCO are designed to be well-compatible with RES, facilitating extensive experiments to study the performance gap of the existing RES methods on the GRES task. In the experimental study, we find that one of the big challenges of GRES is complex relationship modeling. Based on this, we propose a region-based GRES baseline ReLA that adaptively divides the image into regions with sub-instance clues, and explicitly models the region-region and region-language dependencies. The proposed approach ReLA achieves new state-of-the-art performance on the both newly proposed GRES and classic RES tasks. The proposed gRefCOCO dataset and method are available at https://henghuiding.github.io/GRES.

We derive a method that yields highly accurate semantic segmentation maps without the use of any additional neural network, layers, manually annotated training data, or supervised training. Our method is based on the observation that the correlation of a set of pixels belonging to the same semantic segment do not change when generating synthetic variants of an image using the style mixing approach in GANs. We show how we can use GAN inversion to accurately semantically segment synthetic and real photos as well as generate large training image-semantic segmentation mask pairs for downstream tasks.

Semantic image synthesis aims to generate photo realistic images given a semantic segmentation map. Despite much recent progress, training them still requires large datasets of images annotated with per-pixel label maps that are extremely tedious to obtain. To alleviate the high annotation cost, we propose a transfer method that leverages a model trained on a large source dataset to improve the learning ability on small target datasets via estimated pairwise relations between source and target classes. The class affinity matrix is introduced as a first layer to the source model to make it compatible with the target label maps, and the source model is then further fine-tuned for the target domain. To estimate the class affinities we consider different approaches to leverage prior knowledge: semantic segmentation on the source domain, textual label embeddings, and self-supervised vision features. We apply our approach to GAN-based and diffusion-based architectures for semantic synthesis. Our experiments show that the different ways to estimate class affinity can effectively combined, and that our approach significantly improves over existing state-of-the-art transfer approaches for generative image models.

With the increasing interest and rapid development of methods for Ultra-High Resolution (UHR) segmentation, a large-scale benchmark covering a wide range of scenes with full fine-grained dense annotations is urgently needed to facilitate the field. To this end, the URUR dataset is introduced, in the meaning of Ultra-High Resolution dataset with Ultra-Rich Context. As the name suggests, URUR contains amounts of images with high enough resolution (3,008 images of size 5,120×5,120), a wide range of complex scenes (from 63 cities), rich-enough context (1 million instances with 8 categories) and fine-grained annotations (about 80 billion manually annotated pixels), which is far superior to all the existing UHR datasets including DeepGlobe, Inria Aerial, UDD, etc.. Moreover, we also propose WSDNet, a more efficient and effective framework for UHR segmentation especially with ultra-rich context. Specifically, multi-level Discrete Wavelet Transform (DWT) is naturally integrated to release computation burden while preserve more spatial details, along with a Wavelet Smooth Loss (WSL) to reconstruct original structured context and texture with a smooth constrain. Experiments on several UHR datasets demonstrate its state-of-the-art performance. The dataset is available at https://github.com/jankyee/URUR.

This paper introduces Content-aware Token Sharing (CTS), a token reduction approach that improves the computational efficiency of semantic segmentation networks that use Vision Transformers (ViTs). Existing works have proposed token reduction approaches to improve the efficiency of ViT-based image classification networks, but these methods are not directly applicable to semantic segmentation, which we address in this work. We observe that, for semantic segmentation, multiple image patches can share a token if they contain the same semantic class, as they contain redundant information. Our approach leverages this by employing an efficient, class-agnostic policy network that predicts if image patches contain the same semantic class, and lets them share a token if they do. With experiments, we explore the critical design choices of CTS and show its effectiveness on the ADE20K, Pascal Context and Cityscapes datasets, various ViT backbones, and different segmentation decoders. With Content-aware Token Sharing, we are able to reduce the number of processed tokens by up to 44%, without diminishing the segmentation quality.

Few-shot semantic segmentation (FSS) aims to form class-agnostic models segmenting unseen classes with only a handful of annotations. Previous methods limited to the semantic feature and prototype representation suffer from coarse segmentation granularity and train-set overfitting. In this work, we design Hierarchically Decoupled Matching Network (HDMNet) mining pixel-level support correlation based on the transformer architecture. The self-attention modules are used to assist in establishing hierarchical dense features, as a means to accomplish the cascade matching between query and support features. Moreover, we propose a matching module to reduce train-set overfitting and introduce correlation distillation leveraging semantic correspondence from coarse resolution to boost fine-grained segmentation. Our method performs decently in experiments. We achieve 50.0% mIoU on COCO-5i dataset one-shot setting and 56.0% on five-shot segmentation, respectively. The code is available on the project website.

We study the problem of semantic segmentation calibration. Lots of solutions have been proposed to approach model miscalibration of confidence in image classification. However, to date, confidence calibration research on semantic segmentation is still limited. We provide a systematic study on the calibration of semantic segmentation models and propose a simple yet effective approach. First, we find that model capacity, crop size, multi-scale testing, and prediction correctness have impact on calibration. Among them, prediction correctness, especially misprediction, is more important to miscalibration due to over-confidence. Next, we propose a simple, unifying, and effective approach, namely selective scaling, by separating correct/incorrect prediction for scaling and more focusing on misprediction logit smoothing. Then, we study popular existing calibration methods and compare them with selective scaling on semantic segmentation calibration. We conduct extensive experiments with a variety of benchmarks on both in-domain and domain-shift calibration and show that selective scaling consistently outperforms other methods.

Recent attention in instance segmentation has focused on query-based models. Despite being non-maximum suppression (NMS)-free and end-to-end, the superiority of these models on high-accuracy real-time benchmarks has not been well demonstrated. In this paper, we show the strong potential of query-based models on efficient instance segmentation algorithm designs. We present FastInst, a simple, effective query-based framework for real-time instance segmentation. FastInst can execute at a real-time speed (i.e., 32.5 FPS) while yielding an AP of more than 40 (i.e., 40.5 AP) on COCO test-dev without bells and whistles. Specifically, FastInst follows the meta-architecture of recently introduced Mask2Former. Its key designs include instance activation-guided queries, dual-path update strategy, and ground truth mask-guided learning, which enable us to use lighter pixel decoders, fewer Transformer decoder layers, while achieving better performance. The experiments show that FastInst outperforms most state-of-the-art real-time counterparts, including strong fully convolutional baselines, in both speed and accuracy. Code can be found at https://github.com/junjiehe96/FastInst.

Weakly supervised semantic segmentation is typically inspired by class activation maps, which serve as pseudo masks with class-discriminative regions highlighted. Although tremendous efforts have been made to recall precise and complete locations for each class, existing methods still commonly suffer from the unsolicited Out-of-Candidate (OC) error predictions that do not belong to the label candidates, which could be avoidable since the contradiction with image-level class tags is easy to be detected. In this paper, we develop a group ranking-based Out-of-Candidate Rectification (OCR) mechanism in a plug-and-play fashion. Firstly, we adaptively split the semantic categories into In-Candidate (IC) and OC groups for each OC pixel according to their prior annotation correlation and posterior prediction correlation. Then, we derive a differentiable rectification loss to force OC pixels to shift to the IC group. Incorporating OCR with seminal baselines (e.g., AffinityNet, SEAM, MCTformer), we can achieve remarkable performance gains on both Pascal VOC (+3.2%, +3.3%, +0.8% mIoU) and MS COCO (+1.0%, +1.3%, +0.5% mIoU) datasets with negligible extra training overhead, which justifies the effectiveness and generality of OCR.

Modern incremental learning for semantic segmentation methods usually learn new categories based on dense annotations. Although achieve promising results, pixel-by-pixel labeling is costly and time-consuming. Weakly incremental learning for semantic segmentation (WILSS) is a novel and attractive task, which aims at learning to segment new classes from cheap and widely available image-level labels. Despite the comparable results, the image-level labels can not provide details to locate each segment, which limits the performance of WILSS. This inspires us to think how to improve and effectively utilize the supervision of new classes given image-level labels while avoiding forgetting old ones. In this work, we propose a novel and data-efficient framework for WILSS, named FMWISS. Specifically, we propose pre-training based co-segmentation to distill the knowledge of complementary foundation models for generating dense pseudo labels. We further optimize the noisy pseudo masks with a teacher-student architecture, where a plug-in teacher is optimized with a proposed dense contrastive loss. Moreover, we introduce memory-based copy-paste augmentation to improve the catastrophic forgetting problem of old classes. Extensive experiments on Pascal VOC and COCO datasets demonstrate the superior performance of our framework, e.g., FMWISS achieves 70.7% and 73.3% in the 15-5 VOC setting, outperforming the state-of-the-art method by 3.4% and 6.1%, respectively.

Deep neural networks have made huge progress in the last few decades. However, as the real-world data often exhibits a long-tailed distribution, vanilla deep models tend to be heavily biased toward the majority classes. To address this problem, state-of-the-art methods usually adopt a mixture of experts (MoE) to focus on different parts of the long-tailed distribution. Experts in these methods are with the same model depth, which neglects the fact that different classes may have different preferences to be fit by models with different depths. To this end, we propose a novel MoE-based method called Self-Heterogeneous Integration with Knowledge Excavation (SHIKE). We first propose Depth-wise Knowledge Fusion (DKF) to fuse features between different shallow parts and the deep part in one network for each expert, which makes experts more diverse in terms of representation. Based on DKF, we further propose Dynamic Knowledge Transfer (DKT) to reduce the influence of the hardest negative class that has a non-negligible impact on the tail classes in our MoE framework. As a result, the classification accuracy of long-tailed data can be significantly improved, especially for the tail classes. SHIKE achieves the state-of-the-art performance of 56.3%, 60.3%, 75.4%, and 41.9% on CIFAR100-LT (IF100), ImageNet-LT, iNaturalist 2018, and Places-LT, respectively. The source code is available at https://github.com/jinyan-06/SHIKE.

Recently, semi-supervised semantic segmentation has achieved promising performance with a small fraction of labeled data. However, most existing studies treat all unlabeled data equally and barely consider the differences and training difficulties among unlabeled instances. Differentiating unlabeled instances can promote instance-specific supervision to adapt to the model’s evolution dynamically. In this paper, we emphasize the cruciality of instance differences and propose an instance-specific and model-adaptive supervision for semi-supervised semantic segmentation, named iMAS. Relying on the model’s performance, iMAS employs a class-weighted symmetric intersection-over-union to evaluate quantitative hardness of each unlabeled instance and supervises the training on unlabeled data in a model-adaptive manner. Specifically, iMAS learns from unlabeled instances progressively by weighing their corresponding consistency losses based on the evaluated hardness. Besides, iMAS dynamically adjusts the augmentation for each instance such that the distortion degree of augmented instances is adapted to the model’s generalization capability across the training course. Not integrating additional losses and training procedures, iMAS can obtain remarkable performance gains against current state-of-the-art approaches on segmentation benchmarks under different semi-supervised partition protocols.

Given the large-scale data and the high annotation cost, pretraining-finetuning becomes a popular paradigm in multiple computer vision tasks. Previous research has covered both the unsupervised pretraining and supervised finetuning in this paradigm, while little attention is paid to exploiting the annotation budget for finetuning. To fill in this gap, we formally define this new active finetuning task focusing on the selection of samples for annotation in the pretraining-finetuning paradigm. We propose a novel method called ActiveFT for active finetuning task to select a subset of data distributing similarly with the entire unlabeled pool and maintaining enough diversity by optimizing a parametric model in the continuous space. We prove that the Earth Mover’s distance between the distributions of the selected subset and the entire data pool is also reduced in this process. Extensive experiments show the leading performance and high efficiency of ActiveFT superior to baselines on both image classification and semantic segmentation.

Integrated Gradients (IG) as well as its variants are well-known techniques for interpreting the decisions of deep neural networks. While IG-based approaches attain state-of-the-art performance, they often integrate noise into their explanation saliency maps, which reduce their interpretability. To minimize the noise, we examine the source of the noise analytically and propose a new approach to reduce the explanation noise based on our analytical findings. We propose the Important Direction Gradient Integration (IDGI) framework, which can be easily incorporated into any IG-based method that uses the Reimann Integration for integrated gradient computation. Extensive experiments with three IG-based methods show that IDGI improves them drastically on numerous interpretability metrics.

Because the subtle differences between the different sub-categories of common visual categories such as bird species, fine-grained classification has been seen as a challenging task for many years. Most previous works focus towards the features in the single discriminative region isolatedly, while neglect the connection between the different discriminative regions in the whole image. However, the relationship between different discriminative regions contains rich posture information and by adding the posture information, model can learn the behavior of the object which attribute to improve the classification performance. In this paper, we propose a novel fine-grained framework named PMRC (posture mining and reverse cross-entropy), which is able to combine with different backbones to good effect. In PMRC, we use the Deep Navigator to generate the discriminative regions from the images, and then use them to construct the graph. We aggregate the graph by message passing and get the classification results. Specifically, in order to force PMRC to learn how to mine the posture information, we design a novel training paradigm, which makes the Deep Navigator and message passing communicate and train together. In addition, we propose the reverse cross-entropy (RCE) and demomenstate that compared to the cross-entropy (CE), RCE can not only promote the accurracy of our model but also generalize to promote the accuracy of other kinds of fine-grained classification models. Experimental results on benchmark datasets confirm that PMRC can achieve state-of-the-art.

We propose Mask Auto-Labeler (MAL), a high-quality Transformer-based mask auto-labeling framework for instance segmentation using only box annotations. MAL takes box-cropped images as inputs and conditionally generates their mask pseudo-labels.We show that Vision Transformers are good mask auto-labelers. Our method significantly reduces the gap between auto-labeling and human annotation regarding mask quality. Instance segmentation models trained using the MAL-generated masks can nearly match the performance of their fully-supervised counterparts, retaining up to 97.4% performance of fully supervised models. The best model achieves 44.1% mAP on COCO instance segmentation (test-dev 2017), outperforming state-of-the-art box-supervised methods by significant margins. Qualitative results indicate that masks produced by MAL are, in some cases, even better than human annotations.

This paper investigates a phenomenon where query-based object detectors mispredict at the last decoding stage while predicting correctly at an intermediate stage. We review the training process and attribute the overlooked phenomenon to two limitations: lack of training emphasis and cascading errors from decoding sequence. We design and present Selective Query Recollection (SQR), a simple and effective training strategy for query-based object detectors. It cumulatively collects intermediate queries as decoding stages go deeper and selectively forwards the queries to the downstream stages aside from the sequential structure. Such-wise, SQR places training emphasis on later stages and allows later stages to work with intermediate queries from earlier stages directly. SQR can be easily plugged into various query-based object detectors and significantly enhances their performance while leaving the inference pipeline unchanged. As a result, we apply SQR on Adamixer, DAB-DETR, and Deformable-DETR across various settings (backbone, number of queries, schedule) and consistently brings 1.4 ~ 2.8 AP improvement.

Active learning selects informative samples for annotation within budget, which has proven efficient recently on object detection. However, the widely used active detection benchmarks conduct image-level evaluation, which is unrealistic in human workload estimation and biased towards crowded images. Furthermore, existing methods still perform image-level annotation, but equally scoring all targets within the same image incurs waste of budget and redundant labels. Having revealed above problems and limitations, we introduce a box-level active detection framework that controls a box-based budget per cycle, prioritizes informative targets and avoids redundancy for fair comparison and efficient application. Under the proposed box-level setting, we devise a novel pipeline, namely Complementary Pseudo Active Strategy (ComPAS). It exploits both human annotations and the model intelligence in a complementary fashion: an efficient input-end committee queries labels for informative objects only; meantime well-learned targets are identified by the model and compensated with pseudo-labels. ComPAS consistently outperforms 10 competitors under 4 settings in a unified codebase. With supervision from labeled data only, it achieves 100% supervised performance of VOC0712 with merely 19% box annotations. On the COCO dataset, it yields up to 4.3% mAP improvement over the second-best method. ComPAS also supports training with the unlabeled pool, where it surpasses 90% COCO supervised performance with 85% label reduction. Our source code is publicly available at https://github.com/lyumengyao/blad.

Unsupervised domain adaptive object detection (UDA-OD) aims to learn a detector by generalizing knowledge from a labeled source domain to an unlabeled target domain. Though the existing graph-based methods for UDA-OD perform well in some cases, they cannot learn a proper node set for the graph. In addition, these methods build the graph solely based on the visual features and do not consider the linguistic knowledge carried by the semantic prototypes, e.g., dataset labels. To overcome these problems, we propose a cross-modality graph reasoning adaptation (CIGAR) method to take advantage of both visual and linguistic knowledge. Specifically, our method performs cross-modality graph reasoning between the linguistic modality graph and visual modality graphs to enhance their representations. We also propose a discriminative feature selector to find the most discriminative features and take them as the nodes of the visual graph for both efficiency and effectiveness. In addition, we employ the linguistic graph matching loss to regulate the update of linguistic graphs and maintain their semantic representation during the training process. Comprehensive experiments validate the effectiveness of our proposed CIGAR.

The recent detection transformer (DETR) simplifies the object detection pipeline by removing hand-crafted designs and hyperparameters as employed in conventional two-stage object detectors. However, how to leverage the simple yet effective DETR architecture in domain adaptive object detection is largely neglected. Inspired by the unique DETR attention mechanisms, we design DA-DETR, a domain adaptive object detection transformer that introduces information fusion for effective transfer from a labeled source domain to an unlabeled target domain. DA-DETR introduces a novel CNN-Transformer Blender (CTBlender) that fuses the CNN features and Transformer features ingeniously for effective feature alignment and knowledge transfer across domains. Specifically, CTBlender employs the Transformer features to modulate the CNN features across multiple scales where the high-level semantic information and the low-level spatial information are fused for accurate object identification and localization. Extensive experiments show that DA-DETR achieves superior detection performance consistently across multiple widely adopted domain adaptation benchmarks.

Incremental object detection (IOD) aims to train an object detector in phases, each with annotations for new object categories. As other incremental settings, IOD is subject to catastrophic forgetting, which is often addressed by techniques such as knowledge distillation (KD) and exemplar replay (ER). However, KD and ER do not work well if applied directly to state-of-the-art transformer-based object detectors such as Deformable DETR and UP-DETR. In this paper, we solve these issues by proposing a ContinuaL DEtection TRansformer (CL-DETR), a new method for transformer-based IOD which enables effective usage of KD and ER in this context. First, we introduce a Detector Knowledge Distillation (DKD) loss, focusing on the most informative and reliable predictions from old versions of the model, ignoring redundant background predictions, and ensuring compatibility with the available ground-truth labels. We also improve ER by proposing a calibration strategy to preserve the label distribution of the training set, therefore better matching training and testing statistics. We conduct extensive experiments on COCO 2017 and demonstrate that CL-DETR achieves state-of-the-art results in the IOD setting.

We analyze the DETR-based framework on semi-supervised object detection (SSOD) and observe that (1) the one-to-one assignment strategy generates incorrect matching when the pseudo ground-truth bounding box is inaccurate, leading to training inefficiency; (2) DETR-based detectors lack deterministic correspondence between the input query and its prediction output, which hinders the applicability of the consistency-based regularization widely used in current SSOD methods. We present Semi-DETR, the first transformer-based end-to-end semi-supervised object detector, to tackle these problems. Specifically, we propose a Stage-wise Hybrid Matching strategy that com- bines the one-to-many assignment and one-to-one assignment strategies to improve the training efficiency of the first stage and thus provide high-quality pseudo labels for the training of the second stage. Besides, we introduce a Cross-view Query Consistency method to learn the semantic feature invariance of object queries from different views while avoiding the need to find deterministic query correspondence. Furthermore, we propose a Cost-based Pseudo Label Mining module to dynamically mine more pseudo boxes based on the matching cost of pseudo ground truth bounding boxes for consistency training. Extensive experiments on all SSOD settings of both COCO and Pascal VOC benchmark datasets show that our Semi-DETR method outperforms all state-of-the-art methods by clear margins.

State-of-the-art 3D object detectors are usually trained on large-scale datasets with high-quality 3D annotations. However, such 3D annotations are often expensive and time-consuming, which may not be practical for real applications. A natural remedy is to adopt semi-supervised learning (SSL) by leveraging a limited amount of labeled samples and abundant unlabeled samples. Current pseudo-labeling-based SSL object detection methods mainly adopt a teacher-student framework, with a single fixed threshold strategy to generate supervision signals, which inevitably brings confused supervision when guiding the student network training. Besides, the data augmentation of the point cloud in the typical teacher-student framework is too weak, and only contains basic down sampling and flip-and-shift (i.e., rotate and scaling), which hinders the effective learning of feature information. Hence, we address these issues by introducing a novel approach of Hierarchical Supervision and Shuffle Data Augmentation (HSSDA), which is a simple yet effective teacher-student framework. The teacher network generates more reasonable supervision for the student network by designing a dynamic dual-threshold strategy. Besides, the shuffle data augmentation strategy is designed to strengthen the feature representation ability of the student network. Extensive experiments show that HSSDA consistently outperforms the recent state-of-the-art methods on different datasets. The code will be released at https://github.com/azhuantou/HSSDA.

Self-training approaches recently achieved promising results in cross-domain object detection, where people iteratively generate pseudo labels for unlabeled target domain samples with a model, and select high-confidence samples to refine the model. In this work, we reveal that the consistency of classification and localization predictions are crucial to measure the quality of pseudo labels, and propose a new Harmonious Teacher approach to improve the self-training for cross-domain object detection. In particular, we first propose to enhance the quality of pseudo labels by regularizing the consistency of the classification and localization scores when training the detection model. The consistency losses are defined for both labeled source samples and the unlabeled target samples. Then, we further remold the traditional sample selection method by a sample reweighing strategy based on the consistency of classification and localization scores to improve the ranking of predictions. This allows us to fully exploit all instance predictions from the target domain without abandoning valuable hard examples. Without bells and whistles, our method shows superior performance in various cross-domain scenarios compared with the state-of-the-art baselines, which validates the effectiveness of our Harmonious Teacher. Our codes will be available at https://github.com/kinredon/Harmonious-Teacher.

Object detectors often suffer from the domain gap between training (source domain) and real-world applications (target domain). Mean-teacher self-training is a powerful paradigm in unsupervised domain adaptation for object detection, but it struggles with low-quality pseudo-labels. In this work, we identify the intriguing alignment and synergy between mean-teacher self-training and contrastive learning. Motivated by this, we propose Contrastive Mean Teacher (CMT) -- a unified, general-purpose framework with the two paradigms naturally integrated to maximize beneficial learning signals. Instead of using pseudo-labels solely for final predictions, our strategy extracts object-level features using pseudo-labels and optimizes them via contrastive learning, without requiring labels in the target domain. When combined with recent mean-teacher self-training methods, CMT leads to new state-of-the-art target-domain performance: 51.9% mAP on Foggy Cityscapes, outperforming the previously best by 2.1% mAP. Notably, CMT can stabilize performance and provide more significant gains as pseudo-label noise increases.

Recent advances in robust semi-supervised learning (SSL) typical filters out-of-distribution (OOD) information at the sample level. We argue that an overlooked problem of robust SSL is its corrupted information on semantic level, practically limiting the development of the field. In this paper, we take an initiative step to explore and propose a unified framework termed as OOD Semantic Pruning (OSP), aims at pruning OOD semantics out from the in-distribution (ID) features. Specifically, (i) we propose an aliasing OOD matching module to pair each ID sample with an OOD sample with semantic overlap. (ii) We design a soft orthogonality regularization, which first transforms each ID feature by suppressing its semantic component that is collinear with paired OOD sample. It then forces the predictions before and after soft orthogonality transformation to be consistent. Being practically simple, our method shows a strong performance in OOD detection and ID classification on challenging benchmarks. In particular, OSP surpasses the previous state-of-the-art by 13.7% on accuracy for ID classification and 5.9% on AUROC for OOD detection on TinyImageNet dataset. Codes are available in the supplementary material.

Recent studies usually approach multi-label zero-shot learning (MLZSL) with visual-semantic mapping on spatial-class correlation, which can be computationally costly, and worse still, fails to capture fine-grained class-specific semantics. We observe that different channels may usually have different sensitivities on classes, which can correspond to specific semantics. Such an intrinsic channel-class correlation suggests a potential alternative for the more accurate and class-harmonious feature representations. In this paper, our interest is to fully explore the power of channel-class correlation as the unique base for MLZSL. Specifically, we propose a light yet efficient Multi-Label Multi-Layer Perceptron-based Encoder, dubbed (ML)^2P-Encoder, to extract and preserve channel-wise semantics. We reorganize the generated feature maps into several groups, of which each of them can be trained independently with (ML)^2P-Encoder. On top of that, a global group-wise attention module is further designed to build the multi-label specific class relationships among different classes, which eventually fulfills a novel Channel-Class Correlation MLZSL framework (C^3-MLZSL). Extensive experiments on large-scale MLZSL benchmarks including NUS-WIDE and Open-Images-V4 demonstrate the superiority of our model against other representative state-of-the-art models.

We propose a novel teacher-student model for semi-supervised multi-organ segmentation. In the teacher-student model, data augmentation is usually adopted on unlabeled data to regularize the consistent training between teacher and student. We start from a key perspective that fixed relative locations and variable sizes of different organs can provide distribution information where a multi-organ CT scan is drawn. Thus, we treat the prior anatomy as a strong tool to guide the data augmentation and reduce the mismatch between labeled and unlabeled images for semi-supervised learning. More specifically, we propose a data augmentation strategy based on partition-and-recovery N^3 cubes cross- and within- labeled and unlabeled images. Our strategy encourages unlabeled images to learn organ semantics in relative locations from the labeled images (cross-branch) and enhances the learning ability for small organs (within-branch). For within-branch, we further propose to refine the quality of pseudo labels by blending the learned representations from small cubes to incorporate local attributes. Our method is termed as MagicNet, since it treats the CT volume as a magic-cube and N^3-cube partition-and-recovery process matches with the rule of playing a magic-cube. Extensive experiments on two public CT multi-organ datasets demonstrate the effectiveness of MagicNet, and noticeably outperforms state-of-the-art semi-supervised medical image segmentation approaches, with +7% DSC improvement on MACT dataset with 10% labeled images.

Real-world medical image segmentation has tremendous long-tailed complexity of objects, among which tail conditions correlate with relatively rare diseases and are clinically significant. A trustworthy medical AI algorithm should demonstrate its effectiveness on tail conditions to avoid clinically dangerous damage in these out-of-distribution (OOD) cases. In this paper, we adopt the concept of object queries in Mask transformers to formulate semantic segmentation as a soft cluster assignment. The queries fit the feature-level cluster centers of inliers during training. Therefore, when performing inference on a medical image in real-world scenarios, the similarity between pixels and the queries detects and localizes OOD regions. We term this OOD localization as MaxQuery. Furthermore, the foregrounds of real-world medical images, whether OOD objects or inliers, are lesions. The difference between them is obviously less than that between the foreground and background, resulting in the object queries may focus redundantly on the background. Thus, we propose a query-distribution (QD) loss to enforce clear boundaries between segmentation targets and other regions at the query level, improving the inlier segmentation and OOD indication. Our proposed framework is tested on two real-world segmentation tasks, i.e., segmentation of pancreatic and liver tumors, outperforming previous leading algorithms by an average of 7.39% on AUROC, 14.69% on AUPR, and 13.79% on FPR95 for OOD localization. On the other hand, our framework improves the performance of inlier segmentation by an average of 5.27% DSC compared with nnUNet.

Radiography imaging protocols focus on particular body regions, therefore producing images of great similarity and yielding recurrent anatomical structures across patients. To exploit this structured information, we propose the use of Space-aware Memory Queues for In-painting and Detecting anomalies from radiography images (abbreviated as SQUID). We show that SQUID can taxonomize the ingrained anatomical structures into recurrent patterns; and in the inference, it can identify anomalies (unseen/modified patterns) in the image. SQUID surpasses 13 state-of-the-art methods in unsupervised anomaly detection by at least 5 points on two chest X-ray benchmark datasets measured by the Area Under the Curve (AUC). Additionally, we have created a new dataset (DigitAnatomy), which synthesizes the spatial correlation and consistent shape in chest anatomy. We hope DigitAnatomy can prompt the development, evaluation, and interpretability of anomaly detection methods.

Cell detection is a fundamental task in computational pathology that can be used for extracting high-level medical information from whole-slide images. For accurate cell detection, pathologists often zoom out to understand the tissue-level structures and zoom in to classify cells based on their morphology and the surrounding context. However, there is a lack of efforts to reflect such behaviors by pathologists in the cell detection models, mainly due to the lack of datasets containing both cell and tissue annotations with overlapping regions. To overcome this limitation, we propose and publicly release OCELOT, a dataset purposely dedicated to the study of cell-tissue relationships for cell detection in histopathology. OCELOT provides overlapping cell and tissue annotations on images acquired from multiple organs. Within this setting, we also propose multi-task learning approaches that benefit from learning both cell and tissue tasks simultaneously. When compared against a model trained only for the cell detection task, our proposed approaches improve cell detection performance on 3 datasets: proposed OCELOT, public TIGER, and internal CARP datasets. On the OCELOT test set in particular, we show up to 6.79 improvement in F1-score. We believe the contributions of this paper, including the release of the OCELOT dataset at https://lunit-io.github.io/research/publications/ocelot are a crucial starting point toward the important research direction of incorporating cell-tissue relationships in computation pathology.

Multi-class cell detection and counting is an essential task for many pathological diagnoses. Manual counting is tedious and often leads to inter-observer variations among pathologists. While there exist multiple, general-purpose, deep learning-based object detection and counting methods, they may not readily transfer to detecting and counting cells in medical images, due to the limited data, presence of tiny overlapping objects, multiple cell types, severe class-imbalance, minute differences in size/shape of cells, etc. In response, we propose guided posterior regularization DeGPR, which assists an object detector by guiding it to exploit discriminative features among cells. The features may be pathologist-provided or inferred directly from visual data. We validate our model on two publicly available datasets (CoNSeP and MoNuSAC), and on MuCeD, a novel dataset that we contribute. MuCeD consists of 55 biopsy images of the human duodenum for predicting celiac disease. We perform extensive experimentation with three object detection baselines on three datasets to show that DeGPR is model-agnostic, and consistently improves baselines obtaining up to 9% (absolute) mAP gains.

Medical datasets and especially biobanks, often contain extensive tabular data with rich clinical information in addition to images. In practice, clinicians typically have less data, both in terms of diversity and scale, but still wish to deploy deep learning solutions. Combined with increasing medical dataset sizes and expensive annotation costs, the necessity for unsupervised methods that can pretrain multimodally and predict unimodally has risen. To address these needs, we propose the first self-supervised contrastive learning framework that takes advantage of images and tabular data to train unimodal encoders. Our solution combines SimCLR and SCARF, two leading contrastive learning strategies, and is simple and effective. In our experiments, we demonstrate the strength of our framework by predicting risks of myocardial infarction and coronary artery disease (CAD) using cardiac MR images and 120 clinical features from 40,000 UK Biobank subjects. Furthermore, we show the generalizability of our approach to natural images using the DVM car advertisement dataset. We take advantage of the high interpretability of tabular data and through attribution and ablation experiments find that morphometric tabular features, describing size and shape, have outsized importance during the contrastive learning process and improve the quality of the learned embeddings. Finally, we introduce a novel form of supervised contrastive learning, label as a feature (LaaF), by appending the ground truth label as a tabular feature during multimodal pretraining, outperforming all supervised contrastive baselines.

Whole Slide Images (WSIs) are usually gigapixel in size and lack pixel-level annotations. The WSI datasets are also imbalanced in categories. These unique characteristics, significantly different from the ones in natural images, pose the challenge of classifying WSI images as a kind of weakly supervise learning problems. In this study, we propose, RankMix, a data augmentation method of mixing ranked features in a pair of WSIs. RankMix introduces the concepts of pseudo labeling and ranking in order to extract key WSI regions in contributing to the WSI classification task. A two-stage training is further proposed to boost stable training and model performance. To our knowledge, the study of weakly supervised learning from the perspective of data augmentation to deal with the WSI classification problem that suffers from lack of training data and imbalance of categories is relatively unexplored.

Out-of-distribution (OOD) detection has been extensively studied in order to successfully deploy neural networks, in particular, for safety-critical applications. Moreover, performing OOD detection on large-scale datasets is closer to reality, but is also more challenging. Several approaches need to either access the training data for score design or expose models to outliers during training. Some post-hoc methods are able to avoid the aforementioned constraints, but are less competitive. In this work, we propose Generalized ENtropy score (GEN), a simple but effective entropy-based score function, which can be applied to any pre-trained softmax-based classifier. Its performance is demonstrated on the large-scale ImageNet-1k OOD detection benchmark. It consistently improves the average AUROC across six commonly-used CNN-based and visual transformer classifiers over a number of state-of-the-art post-hoc methods. The average AUROC improvement is at least 3.5%. Furthermore, we used GEN on top of feature-based enhancing methods as well as methods using training statistics to further improve the OOD detection performance. The code is available at: https://github.com/XixiLiu95/GEN.

Discriminating known from unknown objects is an important essential ability for human beings. To simulate this ability, a task of unsupervised out-of-distribution object detection (OOD-OD) is proposed to detect the objects that are never-seen-before during model training, which is beneficial for promoting the safe deployment of object detectors. Due to lacking unknown data for supervision, for this task, the main challenge lies in how to leverage the known in-distribution (ID) data to improve the detector’s discrimination ability. In this paper, we first propose a method of Structure-Enhanced Recurrent Variational AutoEncoder (SR-VAE), which mainly consists of two dedicated recurrent VAE branches. Specifically, to boost the performance of object localization, we explore utilizing the classical Laplacian of Gaussian (LoG) operator to enhance the structure information in the extracted low-level features. Meanwhile, we design a VAE branch that recurrently generates the augmentation of the classification features to strengthen the discrimination ability of the object classifier. Finally, to alleviate the impact of lacking unknown data, another cycle-consistent conditional VAE branch is proposed to synthesize virtual OOD features that deviate from the distribution of ID features, which improves the capability of distinguishing OOD objects. In the experiments, our method is evaluated on OOD-OD, open-vocabulary detection, and incremental object detection. The significant performance gains over baselines show the superiorities of our method. The code will be released at https://github.com/AmingWu/SR-VAE.

Multi-view clustering have hitherto been studied due to their effectiveness in dealing with heterogeneous data. Despite the empirical success made by recent works, there still exists several severe challenges. Particularly, previous multi-view clustering algorithms seldom consider the topological structure in data, which is essential for clustering data on manifold. Moreover, existing methods cannot fully consistency the consistency of local structures between different views as they explore the clustering structure in a view-wise manner. In this paper, we propose to exploit the implied data manifold by learning the topological structure of data. Besides, considering that the consistency of multiple views is manifested in the generally similar local structure while the inconsistent structures are minority, we further explore the intersections of multiple views in the sample level such that the cross-view consistency can be better maintained. We model the above concerns in a unified framework and design an efficient algorithm to solve the corresponding optimization problem. Experimental results on various multi-view datasets certificate the effectiveness of the proposed method and verify its superiority over other SOTA approaches.

Self-supervised learning is a central component in recent approaches to deep multi-view clustering (MVC). However, we find large variations in the development of self-supervision-based methods for deep MVC, potentially slowing the progress of the field. To address this, we present DeepMVC, a unified framework for deep MVC that includes many recent methods as instances. We leverage our framework to make key observations about the effect of self-supervision, and in particular, drawbacks of aligning representations with contrastive learning. Further, we prove that contrastive alignment can negatively influence cluster separability, and that this effect becomes worse when the number of views increases. Motivated by our findings, we develop several new DeepMVC instances with new forms of self-supervision. We conduct extensive experiments and find that (i) in line with our theoretical findings, contrastive alignments decreases performance on datasets with many views; (ii) all methods benefit from some form of self-supervision; and (iii) our new instances outperform previous methods on several datasets. Based on our results, we suggest several promising directions for future research. To enhance the openness of the field, we provide an open-source implementation of DeepMVC, including recent models and our new instances. Our implementation includes a consistent evaluation protocol, facilitating fair and accurate evaluation of methods and components.

Fair clustering aims to divide data into distinct clusters while preventing sensitive attributes (e.g., gender, race, RNA sequencing technique) from dominating the clustering. Although a number of works have been conducted and achieved huge success recently, most of them are heuristical, and there lacks a unified theory for algorithm design. In this work, we fill this blank by developing a mutual information theory for deep fair clustering and accordingly designing a novel algorithm, dubbed FCMI. In brief, through maximizing and minimizing mutual information, FCMI is designed to achieve four characteristics highly expected by deep fair clustering, i.e., compact, balanced, and fair clusters, as well as informative features. Besides the contributions to theory and algorithm, another contribution of this work is proposing a novel fair clustering metric built upon information theory as well. Unlike existing evaluation metrics, our metric measures the clustering quality and fairness as a whole instead of separate manner. To verify the effectiveness of the proposed FCMI, we conduct experiments on six benchmarks including a single-cell RNA-seq atlas compared with 11 state-of-the-art methods in terms of five metrics. The code could be accessed from https://pengxi.me.

Few-shot learning (FSL) is popular due to its ability to adapt to novel classes. Compared with inductive few-shot learning, transductive models typically perform better as they leverage all samples of the query set. The two existing classes of methods, prototype-based and graph-based, have the disadvantages of inaccurate prototype estimation and sub-optimal graph construction with kernel functions, respectively. %, which hurt the performance. In this paper, we propose a novel prototype-based label propagation to solve these issues. Specifically, our graph construction is based on the relation between prototypes and samples rather than between samples. As prototypes are being updated, the graph changes.We also estimate the label of each prototype instead of considering a prototype be the class centre. On mini-ImageNet, tiered-ImageNet, CIFAR-FS and CUB datasets, we show the proposed method outperforms other state-of-the-art methods in transductive FSL and semi-supervised FSL when some unlabeled data accompanies the novel few-shot task.

We tackle the Few-Shot Open-Set Recognition (FSOSR) problem, i.e. classifying instances among a set of classes for which we only have a few labeled samples, while simultaneously detecting instances that do not belong to any known class. We explore the popular transductive setting, which leverages the unlabelled query instances at inference. Motivated by the observation that existing transductive methods perform poorly in open-set scenarios, we propose a generalization of the maximum likelihood principle, in which latent scores down-weighing the influence of potential outliers are introduced alongside the usual parametric model. Our formulation embeds supervision constraints from the support set and additional penalties discouraging overconfident predictions on the query set. We proceed with a block-coordinate descent, with the latent scores and parametric model co-optimized alternately, thereby benefiting from each other. We call our resulting formulation Open-Set Likelihood Optimization (OSLO). OSLO is interpretable and fully modular; it can be applied on top of any pre-trained model seamlessly. Through extensive experiments, we show that our method surpasses existing inductive and transductive methods on both aspects of open-set recognition, namely inlier classification and outlier detection. Code is available at https://github.com/ebennequin/few-shot-open-set.

Recent developments of the application of Contrastive Learning in Semi-Supervised Learning (SSL) have demonstrated significant advancements, as a result of its exceptional ability to learn class-aware cluster representations and the full exploitation of massive unlabeled data. However, mismatched instance pairs caused by inaccurate pseudo labels would assign an unlabeled instance to the incorrect class in feature space, hence exacerbating SSL’s renowned confirmation bias. To address this issue, we introduced a novel SSL approach, HyperMatch, which is a plug-in to several SSL designs enabling noise-tolerant utilization of unlabeled data. In particular, confidence predictions are combined with semantic similarities to generate a more objective class distribution, followed by a Gaussian Mixture Model to divide pseudo labels into a ‘confident’ and a ‘less confident’ subset. Then, we introduce Relaxed Contrastive Loss by assigning the ‘less-confident’ samples to a hyper-class, i.e. the union of top-K nearest classes, which effectively regularizes the interference of incorrect pseudo labels and even increases the probability of pulling a ‘less confident’ sample close to its true class. Experiments and in-depth studies demonstrate that HyperMatch delivers remarkable state-of-the-art performance, outperforming FixMatch on CIFAR100 with 400 and 2500 labeled samples by 11.86% and 4.88%, respectively.

Learning with large-scale unlabeled data has become a powerful tool for pre-training Visual Transformers (VTs). However, prior works tend to overlook that, in real-world scenarios, the input data may be corrupted and unreliable. Pre-training VTs on such corrupted data can be challenging, especially when we pre-train via the masked autoencoding approach, where both the inputs and masked “ground truth” targets can potentially be unreliable in this case. To address this limitation, we introduce the Token Boosting Module (TBM) as a plug-and-play component for VTs that effectively allows the VT to learn to extract clean and robust features during masked autoencoding pre-training. We provide theoretical analysis to show how TBM improves model pre-training with more robust and generalizable representations, thus benefiting downstream tasks. We conduct extensive experiments to analyze TBM’s effectiveness, and results on four corrupted datasets demonstrate that TBM consistently improves performance on downstream tasks.

Contrastive learning is a self-supervised representation learning method that achieves milestone performance in various classification tasks. However, due to its unsupervised fashion, it suffers from the false negative sample problem: randomly drawn negative samples that are assumed to have a different label but actually have the same label as the anchor. This deteriorates the performance of contrastive learning as it contradicts the motivation of contrasting semantically similar and dissimilar pairs. This raised the attention and the importance of finding legitimate negative samples, which should be addressed by distinguishing between 1) true vs. false negatives; 2) easy vs. hard negatives. However, previous works were limited to the statistical approach to handle false negative and hard negative samples with hyperparameters tuning. In this paper, we go beyond the statistical approach and explore the connection between hard negative samples and data bias. We introduce a novel debiased contrastive learning method to explore hard negatives by relative difficulty referencing the bias-amplifying counterpart. We propose triplet loss for training a biased encoder that focuses more on easy negative samples. We theoretically show that the triplet loss amplifies the bias in self-supervised representation learning. Finally, we empirically show the proposed method improves downstream classification performance.

Despite advances in Visual Question Answering (VQA), the ability of models to assess their own correctness remains underexplored. Recent work has shown that VQA models, out-of-the-box, can have difficulties abstaining from answering when they are wrong. The option to abstain, also called Selective Prediction, is highly relevant when deploying systems to users who must trust the system’s output (e.g., VQA assistants for users with visual impairments). For such scenarios, abstention can be especially important as users may provide out-of-distribution (OOD) or adversarial inputs that make incorrect answers more likely. In this work, we explore Selective VQA in both in-distribution (ID) and OOD scenarios, where models are presented with mixtures of ID and OOD data. The goal is to maximize the number of questions answered while minimizing the risk of error on those questions. We propose a simple yet effective Learning from Your Peers (LYP) approach for training multimodal selection functions for making abstention decisions. Our approach uses predictions from models trained on distinct subsets of the training data as targets for optimizing a Selective VQA model. It does not require additional manual labels or held-out data and provides a signal for identifying examples that are easy/difficult to generalize to. In our extensive evaluations, we show this benefits a number of models across different architectures and scales. Overall, for ID, we reach 32.92% in the selective prediction metric coverage at 1% risk of error (C@1%) which doubles the previous best coverage of 15.79% on this task. For mixed ID/OOD, using models’ softmax confidences for abstention decisions performs very poorly, answering <5% of questions at 1% risk of error even when faced with only 10% OOD examples, but a learned selection function with LYP can increase that to 25.38% C@1%.

In many real scenarios, data are often divided into a handful of artificial super categories in terms of expert knowledge rather than the representations of images. Concretely, a superclass may contain massive and various raw categories, such as refuse sorting. Due to the lack of common semantic features, the existing classification techniques are intractable to recognize superclass without raw class labels, thus they suffer severe performance damage or require huge annotation costs. To narrow this gap, this paper proposes a superclass learning framework, called SuperClass Learning with Representation Enhancement(SCLRE), to recognize super categories by leveraging enhanced representation. Specifically, by exploiting the self-attention technique across the batch, SCLRE collapses the boundaries of those raw categories and enhances the representation of each superclass. On the enhanced representation space, a superclass-aware decision boundary is then reconstructed. Theoretically, we prove that by leveraging attention techniques the generalization error of SCLRE can be bounded under superclass scenarios. Experimentally, extensive results demonstrate that SCLRE outperforms the baseline and other contrastive-based methods on CIFAR-100 datasets and four high-resolution datasets.

Existing studies indicate that deep neural networks (DNNs) can eventually memorize the label noise. We observe that the memorization strength of DNNs towards each instance is different and can be represented by the confidence value, which becomes larger and larger during the training process. Based on this, we propose a Dynamic Instance-specific Selection and Correction method (DISC) for learning from noisy labels (LNL). We first use a two-view-based backbone for image classification, obtaining confidence for each image from two views. Then we propose a dynamic threshold strategy for each instance, based on the momentum of each instance’s memorization strength in previous epochs to select and correct noisy labeled data. Benefiting from the dynamic threshold strategy and two-view learning, we can effectively group each instance into one of the three subsets (i.e., clean, hard, and purified) based on the prediction consistency and discrepancy by two views at each epoch. Finally, we employ different regularization strategies to conquer subsets with different degrees of label noise, improving the whole network’s robustness. Comprehensive evaluations on three controllable and four real-world LNL benchmarks show that our method outperforms the state-of-the-art (SOTA) methods to leverage useful information in noisy data while alleviating the pollution of label noise.

Deep Neural Networks (DNNs) are rather restrictive in long-tailed data, since they commonly exhibit an under-representation for minority classes. Various remedies have been proposed to tackle this problem from different perspectives, but they ignore the impact of the density of Backbone Features (BFs) on this issue. Through representation learning, DNNs can map BFs into dense clusters in feature space, while the features of minority classes often show sparse clusters. In practical applications, these features are discretely mapped or even cross the decision boundary resulting in misclassification. Inspired by this observation, we propose a simple and generic method, namely Feature Clusters Compression (FCC), to increase the density of BFs by compressing backbone feature clusters. The proposed FCC can be easily achieved by only multiplying original BFs by a scaling factor in training phase, which establishes a linear compression relationship between the original and multiplied features, and forces DNNs to map the former into denser clusters. In test phase, we directly feed original features without multiplying the factor to the classifier, such that BFs of test samples are mapped closer together and do not easily cross the decision boundary. Meanwhile, FCC can be friendly combined with existing long-tailed methods and further boost them. We apply FCC to numerous state-of-the-art methods and evaluate them on widely used long-tailed benchmark datasets. Extensive experiments fully verify the effectiveness and generality of our method. Code is available at https://github.com/lijian16/FCC.

In this paper, we study the application of Test-time domain adaptation in semantic segmentation (TTDA-Seg) where both efficiency and effectiveness are crucial. Existing methods either have low efficiency (e.g., backward optimization) or ignore semantic adaptation (e.g., distribution alignment). Besides, they would suffer from the accumulated errors caused by unstable optimization and abnormal distributions. To solve these problems, we propose a novel backward-free approach for TTDA-Seg, called Dynamically Instance-Guided Adaptation (DIGA). Our principle is utilizing each instance to dynamically guide its own adaptation in a non-parametric way, which avoids the error accumulation issue and expensive optimizing cost. Specifically, DIGA is composed of a distribution adaptation module (DAM) and a semantic adaptation module (SAM), enabling us to jointly adapt the model in two indispensable aspects. DAM mixes the instance and source BN statistics to encourage the model to capture robust representation. SAM combines the historical prototypes with instance-level prototypes to adjust semantic predictions, which can be associated with the parametric classifier to mutually benefit the final results. Extensive experiments evaluated on five target domains demonstrate the effectiveness and efficiency of the proposed method. Our DIGA establishes new state-of-the-art performance in TTDA-Seg.

Semi-Supervised Domain Adaptation (SSDA) involves learning to classify unseen target data with a few labeled and lots of unlabeled target data, along with many labeled source data from a related domain. Current SSDA approaches usually aim at aligning the target data to the labeled source data with feature space mapping and pseudo-label assignments. Nevertheless, such a source-oriented model can sometimes align the target data to source data of the wrong classes, degrading the classification performance. This paper presents a novel source-adaptive paradigm that adapts the source data to match the target data. Our key idea is to view the source data as a noisily-labeled version of the ideal target data. Then, we propose an SSDA model that cleans up the label noise dynamically with the help of a robust cleaner component designed from the target perspective. Since the paradigm is very different from the core ideas behind existing SSDA approaches, our proposed model can be easily coupled with them to improve their performance. Empirical results on two state-of-the-art SSDA approaches demonstrate that the proposed model effectively cleans up the noise within the source labels and exhibits superior performance over those approaches across benchmark datasets. Our code is available at https://github.com/chu0802/SLA.

Open Set Domain Adaptation (OSDA) transfers the model from a label-rich domain to a label-free one containing novel-class samples. Existing OSDA works overlook abundant novel-class semantics hidden in the source domain, leading to a biased model learning and transfer. Although the causality has been studied to remove the semantic-level bias, the non-available novel-class samples result in the failure of existing causal solutions in OSDA. To break through this barrier, we propose a novel causality-driven solution with the unexplored front-door adjustment theory, and then implement it with a theoretically grounded framework, coined AdjustmeNt aNd Alignment (ANNA), to achieve an unbiased OSDA. In a nutshell, ANNA consists of Front-Door Adjustment (FDA) to correct the biased learning in the source domain and Decoupled Causal Alignment (DCA) to transfer the model unbiasedly. On the one hand, FDA delves into fine-grained visual blocks to discover novel-class regions hidden in the base-class image. Then, it corrects the biased model optimization by implementing causal debiasing. On the other hand, DCA disentangles the base-class and novel-class regions with orthogonal masks, and then adapts the decoupled distribution for an unbiased model transfer. Extensive experiments show that ANNA achieves state-of-the-art results. The code is available at https://github.com/CityU-AIM-Group/Anna.

Unsupervised domain adaptation (UDA) approaches focus on adapting models trained on a labeled source domain to an unlabeled target domain. In contrast to UDA, source-free domain adaptation (SFDA) is a more practical setup as access to source data is no longer required during adaptation. Recent state-of-the-art (SOTA) methods on SFDA mostly focus on pseudo-label refinement based self-training which generally suffers from two issues: i) inevitable occurrence of noisy pseudo-labels that could lead to early training time memorization, ii) refinement process requires maintaining a memory bank which creates a significant burden in resource constraint scenarios. To address these concerns, we propose C-SFDA, a curriculum learning aided self-training framework for SFDA that adapts efficiently and reliably to changes across domains based on selective pseudo-labeling. Specifically, we employ a curriculum learning scheme to promote learning from a restricted amount of pseudo labels selected based on their reliabilities. This simple yet effective step successfully prevents label noise propagation during different stages of adaptation and eliminates the need for costly memory-bank based label refinement. Our extensive experimental evaluations on both image recognition and semantic segmentation tasks confirm the effectiveness of our method. C-SFDA is also applicable to online test-time domain adaptation and outperforms previous SOTA methods in this task.

Domain generalization (DG) aims to learn a model that generalizes well to unseen target domains utilizing multiple source domains without re-training. Most existing DG works are based on convolutional neural networks (CNNs). However, the local operation of the convolution kernel makes the model focus too much on local representations (e.g., texture), which inherently causes the model more prone to overfit to the source domains and hampers its generalization ability. Recently, several MLP-based methods have achieved promising results in supervised learning tasks by learning global interactions among different patches of the image. Inspired by this, in this paper, we first analyze the difference between CNN and MLP methods in DG and find that MLP methods exhibit a better generalization ability because they can better capture the global representations (e.g., structure) than CNN methods. Then, based on a recent lightweight MLP method, we obtain a strong baseline that outperforms most start-of-the-art CNN-based methods. The baseline can learn global structure representations with a filter to suppress structure-irrelevant information in the frequency space. Moreover, we propose a dynAmic LOw-Frequency spectrum Transform (ALOFT) that can perturb local texture features while preserving global structure features, thus enabling the filter to remove structure-irrelevant information sufficiently. Extensive experiments on four benchmarks have demonstrated that our method can achieve great performance improvement with a small number of parameters compared to SOTA CNN-based DG methods. Our code is available at https://github.com/lingeringlight/ALOFT/.

Deep neural networks (DNNs) usually fail to generalize well to outside of distribution (OOD) data, especially in the extreme case of single domain generalization (single-DG) that transfers DNNs from single domain to multiple unseen domains. Existing single-DG techniques commonly devise various data-augmentation algorithms, and remould the multi-source domain generalization methodology to learn domain-generalized (semantic) features. Nevertheless, these methods are typically modality-specific, thereby being only applicable to one single modality (e.g., image). In contrast, we target a versatile Modality-Agnostic Debiasing (MAD) framework for single-DG, that enables generalization for different modalities. Technically, MAD introduces a novel two-branch classifier: a biased-branch encourages the classifier to identify the domain-specific (superficial) features, and a general-branch captures domain-generalized features based on the knowledge from biased-branch. Our MAD is appealing in view that it is pluggable to most single-DG models. We validate the superiority of our MAD in a variety of single-DG scenarios with different modalities, including recognition on 1D texts, 2D images, 3D point clouds, and semantic segmentation on 2D images. More remarkably, for recognition on 3D point clouds and semantic segmentation on 2D images, MAD improves DSU by 2.82% and 1.5% in accuracy and mIOU.

Test-Time-Training (TTT) is an approach to cope with out-of-distribution (OOD) data by adapting a trained model to distribution shifts occurring at test-time. We propose to perform this adaptation via Activation Matching (ActMAD): We analyze activations of the model and align activation statistics of the OOD test data to those of the training data. In contrast to existing methods, which model the distribution of entire channels in the ultimate layer of the feature extractor, we model the distribution of each feature in multiple layers across the network. This results in a more fine-grained supervision and makes ActMAD attain state of the art performance on CIFAR-100C and Imagenet-C. ActMAD is also architecture- and task-agnostic, which lets us go beyond image classification, and score 15.4% improvement over previous approaches when evaluating a KITTI-trained object detector on KITTI-Fog. Our experiments highlight that ActMAD can be applied to online adaptation in realistic scenarios, requiring little data to attain its full performance.

When deploying a machine learning model to a new environment, we often encounter the distribution shift problem -- meaning the target data distribution is different from the model’s training distribution. In this paper, we assume that labels are not provided for this new domain, and that we do not store the source data (e.g., for privacy reasons). It has been shown that even small shifts in the data distribution can affect the model’s performance severely. Test Time Adaptation offers a means to combat this problem, as it allows the model to adapt during test time to the new data distribution, using only unlabeled test data batches. To achieve this, the predominant approach is to optimize a surrogate loss on the test-time unlabeled target data. In particular, minimizing the prediction’s entropy on target samples has received much interest as it is task-agnostic and does not require altering the model’s training phase (e.g., does not require adding a self-supervised task during training on the source domain). However, as the target data’s batch size is often small in real-world scenarios (e.g., autonomous driving models process each few frames in real-time), we argue that this surrogate loss is not optimal since it often collapses with small batch sizes. To tackle this problem, in this paper, we propose to use an invariance regularizer as the surrogate loss during test-time adaptation, motivated by our theoretical results regarding the model’s performance under input transformations. The resulting method (TIPI -- Test tIme adaPtation with transformation Invariance) is validated with extensive experiments in various benchmarks (Cifar10-C, Cifar100-C, ImageNet-C, DIGITS, and VisDA17). Remarkably, TIPI is robust against small batch sizes (as small as 2 in our experiments), and consistently outperforms TENT in all settings. Our code is released at https://github.com/atuannguyen/TIPI.

The main challenge in domain generalization (DG) is to handle the distribution shift problem that lies between the training and test data. Recent studies suggest that test-time training (TTT), which adapts the learned model with test data, might be a promising solution to the problem. Generally, a TTT strategy hinges its performance on two main factors: selecting an appropriate auxiliary TTT task for updating and identifying reliable parameters to update during the test phase. Both previous arts and our experiments indicate that TTT may not improve but be detrimental to the learned model if those two factors are not properly considered. This work addresses those two factors by proposing an Improved Test-Time Adaptation (ITTA) method. First, instead of heuristically defining an auxiliary objective, we propose a learnable consistency loss for the TTT task, which contains learnable parameters that can be adjusted toward better alignment between our TTT task and the main prediction task. Second, we introduce additional adaptive parameters for the trained model, and we suggest only updating the adaptive parameters during the test phase. Through extensive experiments, we show that the proposed two strategies are beneficial for the learned model (see Figure 1), and ITTA could achieve superior performance to the current state-of-the-arts on several DG benchmarks.

Few-shot class-incremental learning (FSCIL) aims at learning to classify new classes continually from limited samples without forgetting the old classes. The mainstream framework tackling FSCIL is first to adopt the cross-entropy (CE) loss for training at the base session, then freeze the feature extractor to adapt to new classes. However, in this work, we find that the CE loss is not ideal for the base session training as it suffers poor class separation in terms of representations, which further degrades generalization to novel classes. One tempting method to mitigate this problem is to apply an additional naive supervised contrastive learning (SCL) in the base session. Unfortunately, we find that although SCL can create a slightly better representation separation among different base classes, it still struggles to separate base classes and new classes. Inspired by the observations made, we propose Semantic-Aware Virtual Contrastive model (SAVC), a novel method that facilitates separation between new classes and base classes by introducing virtual classes to SCL. These virtual classes, which are generated via pre-defined transformations, not only act as placeholders for unseen classes in the representation space but also provide diverse semantic information. By learning to recognize and contrast in the fantasy space fostered by virtual classes, our SAVC significantly boosts base class separation and novel class generalization, achieving new state-of-the-art performance on the three widely-used FSCIL benchmark datasets. Code is available at: https://github.com/zysong0113/SAVC.

Privacy and memory are two recurring themes in a broad conversation about the societal impact of AI. These concerns arise from the need for huge amounts of data to train deep neural networks. A promise of Generalized Few-shot Object Detection (G-FSOD), a learning paradigm in AI, is to alleviate the need for collecting abundant training samples of novel classes we wish to detect by leveraging prior knowledge from old classes (i.e., base classes). G-FSOD strives to learn these novel classes while alleviating catastrophic forgetting of the base classes. However, existing approaches assume that the base images are accessible, an assumption that does not hold when sharing and storing data is problematic. In this work, we propose the first data-free knowledge distillation (DFKD) approach for G-FSOD that leverages the statistics of the region of interest (RoI) features from the base model to forge instance-level features without accessing the base images. Our contribution is three-fold: (1) we design a standalone lightweight generator with (2) class-wise heads (3) to generate and replay diverse instance-level base features to the RoI head while finetuning on the novel data. This stands in contrast to standard DFKD approaches in image classification, which invert the entire network to generate base images. Moreover, we make careful design choices in the novel finetuning pipeline to regularize the model. We show that our approach can dramatically reduce the base memory requirements, all while setting a new standard for G-FSOD on the challenging MS-COCO and PASCAL-VOC benchmarks.

Recently, finetuning pretrained vision-language models (VLMs) has been a prevailing paradigm for achieving state-of-the-art performance in VQA. However, as VLMs scale, it becomes computationally expensive, storage inefficient, and prone to overfitting when tuning full model parameters for a specific task in low-resource settings. Although current parameter-efficient tuning methods dramatically reduce the number of tunable parameters, there still exists a significant performance gap with full finetuning. In this paper, we propose MixPHM, a redundancy-aware parameter-efficient tuning method that outperforms full finetuning in low-resource VQA. Specifically, MixPHM is a lightweight module implemented by multiple PHM-experts in a mixture-of-experts manner. To reduce parameter redundancy, we reparameterize expert weights in a low-rank subspace and share part of the weights inside and across MixPHM. Moreover, based on our quantitative analysis of representation redundancy, we propose Redundancy Regularization, which facilitates MixPHM to reduce task-irrelevant redundancy while promoting task-relevant correlation. Experiments conducted on VQA v2, GQA, and OK-VQA with different low-resource settings show that our MixPHM outperforms state-of-the-art parameter-efficient methods and is the only one consistently surpassing full finetuning.

Modern machine learning pipelines are limited due to data availability, storage quotas, privacy regulations, and expensive annotation processes. These constraints make it difficult or impossible to train and update large-scale models on such dynamic annotated sets. Continual learning directly approaches this problem, with the ultimate goal of devising methods where a deep neural network effectively learns relevant patterns for new (unseen) classes, without significantly altering its performance on previously learned ones. In this paper, we address the problem of continual learning for video data. We introduce PIVOT, a novel method that leverages extensive knowledge in pre-trained models from the image domain, thereby reducing the number of trainable parameters and the associated forgetting. Unlike previous methods, ours is the first approach that effectively uses prompting mechanisms for continual learning without any in-domain pre-training. Our experiments show that PIVOT improves state-of-the-art methods by a significant 27% on the 20-task ActivityNet setup.

With the surge of large-scale pre-trained models (PTMs), fine-tuning these models to numerous downstream tasks becomes a crucial problem. Consequently, parameter efficient transfer learning (PETL) of large models has grasped huge attention. While recent PETL methods showcase impressive performance, they rely on optimistic assumptions: 1) the entire parameter set of a PTM is available, and 2) a sufficiently large memory capacity for the fine-tuning is equipped. However, in most real-world applications, PTMs are served as a black-box API or proprietary software without explicit parameter accessibility. Besides, it is hard to meet a large memory requirement for modern PTMs. In this work, we propose black-box visual prompting (BlackVIP), which efficiently adapts the PTMs without knowledge about model architectures and parameters. BlackVIP has two components; 1) Coordinator and 2) simultaneous perturbation stochastic approximation with gradient correction (SPSA-GC). The Coordinator designs input-dependent image-shaped visual prompts, which improves few-shot adaptation and robustness on distribution/location shift. SPSA-GC efficiently estimates the gradient of a target model to update Coordinator. Extensive experiments on 16 datasets demonstrate that BlackVIP enables robust adaptation to diverse domains without accessing PTMs’ parameters, with minimal memory requirements. Code: https://github.com/changdaeoh/BlackVIP

Deep neural networks suffer from catastrophic forgetting in class incremental learning, where the classification accuracy of old classes drastically deteriorates when the networks learn the knowledge of new classes. Many works have been proposed to solve the class incremental learning problem. However, most of them either suffer from serious catastrophic forgetting and stability-plasticity dilemma or need too many extra parameters and computations. To meet the challenge, we propose a novel framework, Diverse Knowledge Transfer Transformer~(DKT). which contains two novel knowledge transfers based on the attention mechanism to transfer the task-general knowledge and task-specific knowledge to the current task to alleviate catastrophic forgetting. Besides, we propose a duplex classifier to address the stability-plasticity dilemma, and a novel loss function to cluster the same categories in feature space and discriminate the features between old and new tasks to force the task specific knowledge to be more diverse. Our method needs only a few extra parameters, which are negligible, to tackle the increasing number of tasks. We conduct comprehensive experimental results on CIFAR100, ImageNet100/1000 datasets. The experiment results show that our method outperforms other competitive methods and achieves state-of-the-art performance.

Online class-incremental continual learning is a specific task of continual learning. It aims to continuously learn new classes from data stream and the samples of data stream are seen only once, which suffers from the catastrophic forgetting issue, i.e., forgetting historical knowledge of old classes. Existing replay-based methods effectively alleviate this issue by saving and replaying part of old data in a proxy-based or contrastive-based replay manner. Although these two replay manners are effective, the former would incline to new classes due to class imbalance issues, and the latter is unstable and hard to converge because of the limited number of samples. In this paper, we conduct a comprehensive analysis of these two replay manners and find that they can be complementary. Inspired by this finding, we propose a novel replay-based method called proxy-based contrastive replay (PCR). The key operation is to replace the contrastive samples of anchors with corresponding proxies in the contrastive-based way. It alleviates the phenomenon of catastrophic forgetting by effectively addressing the imbalance issue, as well as keeps a faster convergence of the model. We conduct extensive experiments on three real-world benchmark datasets, and empirical results consistently demonstrate the superiority of PCR over various state-of-the-art methods.

This paper studies the potential of distilling knowledge from pre-trained models, especially Masked Autoencoders. Our approach is simple: in addition to optimizing the pixel reconstruction loss on masked inputs, we minimize the distance between the intermediate feature map of the teacher model and that of the student model. This design leads to a computationally efficient knowledge distillation framework, given 1) only a small visible subset of patches is used, and 2) the (cumbersome) teacher model only needs to be partially executed, i.e., forward propagate inputs through the first few layers, for obtaining intermediate feature maps. Compared to directly distilling fine-tuned models, distilling pre-trained models substantially improves downstream performance. For example, by distilling the knowledge from an MAE pre-trained ViT-L into a ViT-B, our method achieves 84.0% ImageNet top-1 accuracy, outperforming the baseline of directly distilling a fine-tuned ViT-L by 1.2%. More intriguingly, our method can robustly distill knowledge from teacher models even with extremely high masking ratios: e.g., with 95% masking ratio where merely TEN patches are visible during distillation, our ViT-B competitively attains a top-1 ImageNet accuracy of 83.6%; surprisingly, it can still secure 82.4% top-1 ImageNet accuracy by aggressively training with just FOUR visible patches (98% masking ratio). The code will be made publicly available.

Despite the tremendous progress on data-free knowledge distillation (DFKD) based on synthetic data generation, there are still limitations in diverse and efficient data synthesis. It is naive to expect that a simple combination of generative network-based data synthesis and data augmentation will solve these issues. Therefore, this paper proposes a novel data-free knowledge distillation method (SpaceshipNet) based on channel-wise feature exchange (CFE) and multi-scale spatial activation region consistency (mSARC) constraint. Specifically, CFE allows our generative network to better sample from the feature space and efficiently synthesize diverse images for learning the student network. However, using CFE alone can severely amplify the unwanted noises in the synthesized images, which may result in failure to improve distillation learning and even have negative effects. Therefore, we propose mSARC to assure the student network can imitate not only the logit output but also the spatial activation region of the teacher network in order to alleviate the influence of unwanted noises in diverse synthetic images on distillation learning. Extensive experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet, Imagenette, and ImageNet100 show that our method can work well with different backbone networks, and outperform the state-of-the-art DFKD methods. Code will be available at: https://github.com/skgyu/SpaceshipNet.

Knowledge Distillation (KD) aims at distilling the knowledge from the large teacher model to a lightweight student model. Mainstream KD methods can be divided into two categories, logit distillation, and feature distillation. The former is easy to implement, but inferior in performance, while the latter is not applicable to some practical circumstances due to concerns such as privacy and safety. Towards this dilemma, in this paper, we explore a stronger logit distillation method via making better utilization of logit outputs. Concretely, we propose a simple yet effective approach to logit distillation via multi-level prediction alignment. Through this framework, the prediction alignment is not only conducted at the instance level, but also at the batch and class level, through which the student model learns instance prediction, input correlation, and category correlation simultaneously. In addition, a prediction augmentation mechanism based on model calibration further boosts the performance. Extensive experiment results validate that our method enjoys consistently higher performance than previous logit distillation methods, and even reaches competitive performance with mainstream feature distillation methods. We promise to release our code and models to ensure reproducibility.

Catastrophic forgetting has been a major challenge in continual learning, where the model needs to learn new tasks with limited or no access to data from previously seen tasks. To tackle this challenge, methods based on knowledge distillation in feature space have been proposed and shown to reduce forgetting. However, most feature distillation methods directly constrain the new features to match the old ones, overlooking the need for plasticity. To achieve a better stability-plasticity trade-off, we propose Backward Feature Projection (BFP), a method for continual learning that allows the new features to change up to a learnable linear transformation of the old features. BFP preserves the linear separability of the old classes while allowing the emergence of new feature directions to accommodate new classes. BFP can be integrated with existing experience replay methods and boost performance by a significant margin. We also demonstrate that BFP helps learn a better representation space, in which linear separability is well preserved during continual learning and linear probing achieves high classification accuracy.

We show that the ability of a neural network to integrate information from diverse sources hinges critically on being exposed to properly correlated signals during the early phases of training. Interfering with the learning process during this initial stage can permanently impair the development of a skill, both in artificial and biological systems where the phenomenon is known as a critical learning period. We show that critical periods arise from the complex and unstable early transient dynamics, which are decisive of final performance of the trained system and their learned representations. This evidence challenges the view, engendered by analysis of wide and shallow networks, that early learning dynamics of neural networks are simple, akin to those of a linear model. Indeed, we show that even deep linear networks exhibit critical learning periods for multi-source integration, while shallow networks do not. To better understand how the internal representations change according to disturbances or sensory deficits, we introduce a new measure of source sensitivity, which allows us to track the inhibition and integration of sources during training. Our analysis of inhibition suggests cross-source reconstruction as a natural auxiliary training objective, and indeed we show that architectures trained with cross-sensor reconstruction objectives are remarkably more resilient to critical periods. Our findings suggest that the recent success in self-supervised multi-modal training compared to previous supervised efforts may be in part due to more robust learning dynamics and not solely due to better architectures and/or more data.

Data augmentation is known to improve the generalization capabilities of neural networks, provided that the set of transformations is chosen with care, a selection often performed manually. Automatic data augmentation aims at automating this process. However, most recent approaches still rely on some prior information; they start from a small pool of manually-selected default transformations that are either used to pretrain the network or forced to be part of the policy learned by the automatic data augmentation algorithm. In this paper, we propose to directly learn the augmentation policy without leveraging such prior knowledge. The resulting bilevel optimization problem becomes more challenging due to the larger search space and the inherent instability of bilevel optimization algorithms. To mitigate these issues (i) we follow a successive cold-start strategy with a Kullback-Leibler regularization, and (ii) we parameterize magnitudes as continuous distributions. Our approach leads to competitive results on standard benchmarks despite a more challenging setting, and generalizes beyond natural images.

Domain generalization (DG) tends to alleviate the poor generalization capability of deep neural networks by learning model with multiple source domains. A classical solution to DG is domain augmentation, the common belief of which is that diversifying source domains will be conducive to the out-of-distribution generalization. However, these claims are understood intuitively, rather than mathematically. Our explorations empirically reveal that the correlation between model generalization and the diversity of domains may be not strictly positive, which limits the effectiveness of domain augmentation. This work therefore aim to guarantee and further enhance the validity of this strand. To this end, we propose a new perspective on DG that recasts it as a convex game between domains. We first encourage each diversified domain to enhance model generalization by elaborately designing a regularization term based on supermodularity. Meanwhile, a sample filter is constructed to eliminate low-quality samples, thereby avoiding the impact of potentially harmful information. Our framework presents a new avenue for the formal analysis of DG, heuristic analysis and extensive experiments demonstrate the rationality and effectiveness.

Continual learning aims to efficiently learn from a non-stationary stream of data while avoiding forgetting the knowledge of old data. In many practical applications, data complies with non-Euclidean geometry. As such, the commonly used Euclidean space cannot gracefully capture non-Euclidean geometric structures of data, leading to inferior results. In this paper, we study continual learning from a novel perspective by exploring data geometry for the non-stationary stream of data. Our method dynamically expands the geometry of the underlying space to match growing geometric structures induced by new data, and prevents forgetting by keeping geometric structures of old data into account. In doing so, we make use of the mixed-curvature space and propose an incremental search scheme, through which the growing geometric structures are encoded. Then, we introduce an angular-regularization loss and a neighbor-robustness loss to train the model, capable of penalizing the change of global geometric structures and local geometric structures. Experiments show that our method achieves better performance than baseline methods designed in Euclidean space.

Generative modeling with ordinary differential equations (ODEs) has achieved fantastic results on a variety of applications. Yet, few works have focused on controlling the generated content of a pre-trained ODE-based generative model. In this paper, we propose to optimize the output of ODE models according to a guidance function to achieve controllable generation. We point out that, the gradients can be efficiently back-propagated from the output to any intermediate time steps on the ODE trajectory, by decomposing the back-propagation and computing vector-Jacobian products. To further accelerate the computation of the back-propagation, we propose to use a non-uniform discretization to approximate the ODE trajectory, where we measure how straight the trajectory is and gather the straight parts into one discretization step. This allows us to save ~90% of the back-propagation time with ignorable error. Our framework, named FlowGrad, outperforms the state-of-the-art baselines on text-guided image manipulation. Moreover, FlowGrad enables us to find global semantic directions in frozen ODE-based generative models that can be used to manipulate new images without extra optimization.

In this paper, we explore a novel model reusing task tailored for graph neural networks (GNNs), termed as “deep graph reprogramming”. We strive to reprogram a pre-trained GNN, without amending raw node features nor model parameters, to handle a bunch of cross-level downstream tasks in various domains. To this end, we propose an innovative Data Reprogramming paradigm alongside a Model Reprogramming paradigm. The former one aims to address the challenge of diversified graph feature dimensions for various tasks on the input side, while the latter alleviates the dilemma of fixed per-task-per-model behavior on the model side. For data reprogramming, we specifically devise an elaborated Meta-FeatPadding method to deal with heterogeneous input dimensions, and also develop a transductive Edge-Slimming as well as an inductive Meta-GraPadding approach for diverse homogenous samples. Meanwhile, for model reprogramming, we propose a novel task-adaptive Reprogrammable-Aggregator, to endow the frozen model with larger expressive capacities in handling cross-domain tasks. Experiments on fourteen datasets across node/graph classification/regression, 3D object recognition, and distributed action recognition, demonstrate that the proposed methods yield gratifying results, on par with those by re-training from scratch.

Recently vision transformer models have become prominent models for a range of tasks. These models, however, usually suffer from intensive computational costs and heavy memory requirements, making them impractical for deployment on edge platforms. Recent studies have proposed to prune transformers in an unexplainable manner, which overlook the relationship between internal units of the model and the target class, thereby leading to inferior performance. To alleviate this problem, we propose a novel explainable pruning framework dubbed X-Pruner, which is designed by considering the explainability of the pruning criterion. Specifically, to measure each prunable unit’s contribution to predicting each target class, a novel explainability-aware mask is proposed and learned in an end-to-end manner. Then, to preserve the most informative units and learn the layer-wise pruning rate, we adaptively search the layer-wise threshold that differentiates between unpruned and pruned units based on their explainability-aware mask values. To verify and evaluate our method, we apply the X-Pruner on representative transformer models including the DeiT and Swin Transformer. Comprehensive simulation results demonstrate that the proposed X-Pruner outperforms the state-of-the-art black-box methods with significantly reduced computational costs and slight performance degradation.

Pruning - that is, setting a significant subset of the parameters of a neural network to zero - is one of the most popular methods of model compression. Yet, several recent works have raised the issue that pruning may induce or exacerbate bias in the output of the compressed model. Despite existing evidence for this phenomenon, the relationship between neural network pruning and induced bias is not well-understood. In this work, we systematically investigate and characterize this phenomenon in Convolutional Neural Networks for computer vision. First, we show that it is in fact possible to obtain highly-sparse models, e.g. with less than 10% remaining weights, which do not decrease in accuracy nor substantially increase in bias when compared to dense models. At the same time, we also find that, at higher sparsities, pruned models exhibit higher uncertainty in their outputs, as well as increased correlations, which we directly link to increased bias. We propose easy-to-use criteria which, based only on the uncompressed model, establish whether bias will increase with pruning, and identify the samples most susceptible to biased predictions post-compression.

Binary Neural Network (BNN) represents convolution weights with 1-bit values, which enhances the efficiency of storage and computation. This paper is motivated by a previously revealed phenomenon that the binary kernels in successful BNNs are nearly power-law distributed: their values are mostly clustered into a small number of codewords. This phenomenon encourages us to compact typical BNNs and obtain further close performance through learning non-repetitive kernels within a binary kernel subspace. Specifically, we regard the binarization process as kernel grouping in terms of a binary codebook, and our task lies in learning to select a smaller subset of codewords from the full codebook. We then leverage the Gumbel-Sinkhorn technique to approximate the codeword selection process, and develop the Permutation Straight-Through Estimator (PSTE) that is able to not only optimize the selection process end-to-end but also maintain the non-repetitive occupancy of selected codewords. Experiments verify that our method reduces both the model size and bit-wise computational costs, and achieves accuracy improvements compared with state-of-the-art BNNs under comparable budgets.

Reliable uncertainty from deterministic single-forward pass models is sought after because conventional methods of uncertainty quantification are computationally expensive. We take two complex single-forward-pass uncertainty approaches, DUQ and SNGP, and examine whether they mainly rely on a well-regularized feature space. Crucially, without using their more complex methods for estimating uncertainty, we find that a single softmax neural net with such a regularized feature-space, achieved via residual connections and spectral normalization, outperforms DUQ and SNGP’s epistemic uncertainty predictions using simple Gaussian Discriminant Analysis post-training as a separate feature-space density estimator---without fine-tuning on OoD data, feature ensembling, or input pre-procressing. Our conceptually simple Deep Deterministic Uncertainty (DDU) baseline can also be used to disentangle aleatoric and epistemic uncertainty and performs as well as Deep Ensembles, the state-of-the art for uncertainty prediction, on several OoD benchmarks (CIFAR-10/100 vs SVHN/Tiny-ImageNet, ImageNet vs ImageNet-O), active learning settings across different model architectures, as well as in large scale vision tasks like semantic segmentation, while being computationally cheaper.

Understanding how well a deep generative model captures a distribution of high-dimensional data remains an important open challenge. It is especially difficult for certain model classes, such as Generative Adversarial Networks and Diffusion Models, whose models do not admit exact likelihoods. In this work, we demonstrate that generalized empirical likelihood (GEL) methods offer a family of diagnostic tools that can identify many deficiencies of deep generative models (DGMs). We show, with appropriate specification of moment conditions, that the proposed method can identify which modes have been dropped, the degree to which DGMs are mode imbalanced, and whether DGMs sufficiently capture intra-class diversity. We show how to combine techniques from Maximum Mean Discrepancy and Generalized Empirical Likelihood to create not only distribution tests that retain per-sample interpretability, but also metrics that include label information. We find that such tests predict the degree of mode dropping and mode imbalance up to 60% better than metrics such as improved precision/recall.

Recent studies suggest that computer vision models come at the risk of compromising fairness. There are extensive works to alleviate unfairness in computer vision using pre-processing, in-processing, and post-processing methods. In this paper, we lead a novel fairness-aware learning paradigm for in-processing methods through the lens of the lottery ticket hypothesis (LTH) in the context of computer vision fairness. We randomly initialize a dense neural network and find appropriate binary masks for the weights to obtain fair sparse subnetworks without any weight training. Interestingly, to the best of our knowledge, we are the first to discover that such sparse subnetworks with inborn fairness exist in randomly initialized networks, achieving an accuracy-fairness trade-off comparable to that of dense neural networks trained with existing fairness-aware in-processing approaches. We term these fair subnetworks as Fair Scratch Tickets (FSTs). We also theoretically provide fairness and accuracy guarantees for them. In our experiments, we investigate the existence of FSTs on various datasets, target attributes, random initialization methods, sparsity patterns, and fairness surrogates. We also find that FSTs can transfer across datasets and investigate other properties of FSTs.

Zero-shot quantization (ZSQ) is promising for compressing and accelerating deep neural networks when the data for training full-precision models are inaccessible. In ZSQ, network quantization is performed using synthetic samples, thus, the performance of quantized models depends heavily on the quality of synthetic samples. Nonetheless, we find that the synthetic samples constructed in existing ZSQ methods can be easily fitted by models. Accordingly, quantized models obtained by these methods suffer from significant performance degradation on hard samples. To address this issue, we propose HArd sample Synthesizing and Training (HAST). Specifically, HAST pays more attention to hard samples when synthesizing samples and makes synthetic samples hard to fit when training quantized models. HAST aligns features extracted by full-precision and quantized models to ensure the similarity between features extracted by these two models. Extensive experiments show that HAST significantly outperforms existing ZSQ methods, achieving performance comparable to models that are quantized with real data.

Post-training quantization (PTQ) is a neural network compression technique that converts a full-precision model into a quantized model using lower-precision data types. Although it can help reduce the size and computational cost of deep neural networks, it can also introduce quantization noise and reduce prediction accuracy, especially in extremely low-bit settings. How to determine the appropriate quantization parameters (e.g., scaling factors and rounding of weights) is the main problem facing now. Existing methods attempt to determine these parameters by minimize the distance between features before and after quantization, but such an approach only considers local information and may not result in the most optimal quantization parameters. We analyze this issue and propose PD-Quant, a method that addresses this limitation by considering global information. It determines the quantization parameters by using the information of differences between network prediction before and after quantization. In addition, PD-Quant can alleviate the overfitting problem in PTQ caused by the small number of calibration sets by adjusting the distribution of activations. Experiments show that PD-Quant leads to better quantization parameters and improves the prediction accuracy of quantized models, especially in low-bit settings. For example, PD-Quant pushes the accuracy of ResNet-18 up to 53.14% and RegNetX-600MF up to 40.67% in weight 2-bit activation 2-bit. The code is released at https://github.com/hustvl/PD-Quant.

Recently, the robustness of deep neural networks has drawn extensive attention due to the potential distribution shift between training and testing data (e.g., deep models trained on high-quality images are sensitive to corruption during testing). Many researchers attempt to make the model learn invariant representations from multiple corrupted data through data augmentation or image-pair-based feature distillation to improve the robustness. Inspired by sparse representation in image restoration, we opt to address this issue by learning image-quality-independent feature representation in a simple plug-and-play manner, that is, to introduce discrete vector quantization (VQ) to remove redundancy in recognition models. Specifically, we first add a codebook module to the network to quantize deep features. Then we concatenate them and design a self-attention module to enhance the representation. During training, we enforce the quantization of features from clean and corrupted images in the same discrete embedding space so that an invariant quality-independent feature representation can be learned to improve the recognition robustness of low-quality images. Qualitative and quantitative experimental results show that our method achieved this goal effectively, leading to a new state-of-the-art result of 43.1% mCE on ImageNet-C with ResNet50 as the backbone. On other robustness benchmark datasets, such as ImageNet-R, our method also has an accuracy improvement of almost 2%.

Masked Auto-Encoder (MAE) pretraining methods randomly mask image patches and then train a vision Transformer to reconstruct the original pixels based on the unmasked patches. While they demonstrates impressive performance for downstream vision tasks, it generally requires a large amount of training resource. In this paper, we introduce a novel Generative Adversarial Networks alike framework, referred to as GAN-MAE, where a generator is used to generate the masked patches according to the remaining visible patches, and a discriminator is employed to predict whether the patch is synthesized by the generator. We believe this capacity of distinguishing whether the image patch is predicted or original is benefit to representation learning. Another key point lies in that the parameters of the vision Transformer backbone in the generator and discriminator are shared. Extensive experiments demonstrate that adversarial training of GAN-MAE framework is more efficient and accordingly outperforms the standard MAE given the same model size, training data, and computation resource. The gains are substantially robust for different model sizes and datasets, in particular, a ViT-B model trained with GAN-MAE for 200 epochs outperforms the MAE with 1600 epochs on fine-tuning top-1 accuracy of ImageNet-1k with much less FLOPs. Besides, our approach also works well at transferring downstream tasks.

Modern Generative Adversarial Networks (GANs) generate realistic images remarkably well. Previous work has demonstrated the feasibility of “GAN-classifiers” that are distinct from the co-trained discriminator, and operate on images generated from a frozen GAN. That such classifiers work at all affirms the existence of “knowledge gaps” (out-of-distribution artifacts across samples) present in GAN training. We iteratively train GAN-classifiers and train GANs that “fool” the classifiers (in an attempt to fill the knowledge gaps), and examine the effect on GAN training dynamics, output quality, and GAN-classifier generalization. We investigate two settings, a small DCGAN architecture trained on low dimensional images (MNIST), and StyleGAN2, a SOTA GAN architecture trained on high dimensional images (FFHQ). We find that the DCGAN is unable to effectively fool a held-out GAN-classifier without compromising the output quality. However, StyleGAN2 can fool held-out classifiers with no change in output quality, and this effect persists over multiple rounds of GAN/classifier training which appears to reveal an ordering over optima in the generator parameter space. Finally, we study different classifier architectures and show that the architecture of the GAN-classifier has a strong influence on the set of its learned artifacts.

Recent work has shown that deep vision models tend to be overly dependent on low-level or “texture” features, leading to poor generalization. Various data augmentation strategies have been proposed to overcome this so-called texture bias in DNNs. We propose a simple, lightweight adversarial augmentation technique that explicitly incentivizes the network to learn holistic shapes for accurate prediction in an object classification setting. Our augmentations superpose edgemaps from one image onto another image with shuffled patches, using a randomly determined mixing proportion, with the image label of the edgemap image. To classify these augmented images, the model needs to not only detect and focus on edges but distinguish between relevant and spurious edges. We show that our augmentations significantly improve classification accuracy and robustness measures on a range of datasets and neural architectures. As an example, for ViT-S, We obtain absolute gains on classification accuracy gains up to 6%. We also obtain gains of up to 28% and 8.5% on natural adversarial and out-of-distribution datasets like ImageNet-A (for ViTB) and ImageNet-R (for ViT-S), respectively. Analysis using a range of probe datasets shows substantially increased shape sensitivity in our trained models, explaining the observed improvement in robustness and classification accuracy.

With generative models proliferating at a rapid rate, there is a growing need for general purpose fake image detectors. In this work, we first show that the existing paradigm, which consists of training a deep network for real-vs-fake classification, fails to detect fake images from newer breeds of generative models when trained to detect GAN fake images. Upon analysis, we find that the resulting classifier is asymmetrically tuned to detect patterns that make an image fake. The real class becomes a ‘sink’ class holding anything that is not fake, including generated images from models not accessible during training. Building upon this discovery, we propose to perform real-vs-fake classification without learning; i.e., using a feature space not explicitly trained to distinguish real from fake images. We use nearest neighbor and linear probing as instantiations of this idea. When given access to the feature space of a large pretrained vision-language model, the very simple baseline of nearest neighbor classification has surprisingly good generalization ability in detecting fake images from a wide variety of generative models; e.g., it improves upon the SoTA by +15.07 mAP and +25.90% acc when tested on unseen diffusion and autoregressive models.

Most anomaly detection (AD) models are learned using only normal samples in an unsupervised way, which may result in ambiguous decision boundary and insufficient discriminability. In fact, a few anomaly samples are often available in real-world applications, the valuable knowledge of known anomalies should also be effectively exploited. However, utilizing a few known anomalies during training may cause another issue that the model may be biased by those known anomalies and fail to generalize to unseen anomalies. In this paper, we tackle supervised anomaly detection, i.e., we learn AD models using a few available anomalies with the objective to detect both the seen and unseen anomalies. We propose a novel explicit boundary guided semi-push-pull contrastive learning mechanism, which can enhance model’s discriminability while mitigating the bias issue. Our approach is based on two core designs: First, we find an explicit and compact separating boundary as the guidance for further feature learning. As the boundary only relies on the normal feature distribution, the bias problem caused by a few known anomalies can be alleviated. Second, a boundary guided semi-push-pull loss is developed to only pull the normal features together while pushing the abnormal features apart from the separating boundary beyond a certain margin region. In this way, our model can form a more explicit and discriminative decision boundary to distinguish known and also unseen anomalies from normal samples more effectively. Code will be available at https://github.com/xcyao00/BGAD.

Anomaly detection in surveillance videos is a challenging computer vision task where only normal videos are available during training. Recent work released the first virtual anomaly detection dataset to assist real-world detection. However, an anomaly gap exists because the anomalies are bounded in the virtual dataset but unbounded in the real world, so it reduces the generalization ability of the virtual dataset. There also exists a scene gap between virtual and real scenarios, including scene-specific anomalies (events that are abnormal in one scene but normal in another) and scene-specific attributes, such as the viewpoint of the surveillance camera. In this paper, we aim to solve the problem of the anomaly gap and scene gap by proposing a prompt-based feature mapping framework (PFMF). The PFMF contains a mapping network guided by an anomaly prompt to generate unseen anomalies with unbounded types in the real scenario, and a mapping adaptation branch to narrow the scene gap by applying domain classifier and anomaly classifier. The proposed framework outperforms the state-of-the-art on three benchmark datasets. Extensive ablation experiments also show the effectiveness of our framework design.

Anomaly detection is an important application in large-scale industrial manufacturing. Recent methods for this task have demonstrated excellent accuracy but come with a latency trade-off. Memory based approaches with dominant performances like PatchCore or Coupled-hypersphere-based Feature Adaptation (CFA) require an external memory bank, which significantly lengthens the execution time. Another approach that employs Reversed Distillation (RD) can perform well while maintaining low latency. In this paper, we revisit this idea to improve its performance, establishing a new state-of-the-art benchmark on the challenging MVTec dataset for both anomaly detection and localization. The proposed method, called RD++, runs six times faster than PatchCore, and two times faster than CFA but introduces a negligible latency compared to RD. We also experiment on the BTAD and Retinal OCT datasets to demonstrate our method’s generalizability and conduct important ablation experiments to provide insights into its configurations. Source code will be available at https://github.com/tientrandinh/Revisiting-Reverse-Distillation.

Continual Learning (CL) has achieved rapid progress in recent years. However, it is still largely unknown how to determine whether a CL model is trustworthy and how to foster its trustworthiness. This work focuses on evaluating and improving the robustness to corruptions of existing CL models. Our empirical evaluation results show that existing state-of-the-art (SOTA) CL models are particularly vulnerable to various data corruptions during testing. To make them trustworthy and robust to corruptions deployed in safety-critical scenarios, we propose a meta-learning framework of self-adaptive data augmentation to tackle the corruption robustness in CL. The proposed framework, MetaMix, learns to augment and mix data, automatically transforming the new task data or memory data. It directly optimizes the generalization performance against data corruptions during training. To evaluate the corruption robustness of our proposed approach, we construct several CL corruption datasets with different levels of severity. We perform comprehensive experiments on both task- and class-continual learning. Extensive experiments demonstrate the effectiveness of our proposed method compared to SOTA baselines.

Federated learning (FL) is an attractive distributed learning paradigm supporting real-time continuous learning and client privacy by default. In most FL approaches, all edge clients are assumed to have sufficient computation capabilities to participate in the learning of a deep neural network (DNN) model. However, in real-life applications, some clients may have severely limited resources and can only train a much smaller local model. This paper presents ScaleFL, a novel FL approach with two distinctive mechanisms to handle resource heterogeneity and provide an equitable FL framework for all clients. First, ScaleFL adaptively scales down the DNN model along width and depth dimensions by leveraging early exits to find the best-fit models for resource-aware local training on distributed clients. In this way, ScaleFL provides an efficient balance of preserving basic and complex features in local model splits with various sizes for joint training while enabling fast inference for model deployment. Second, ScaleFL utilizes self-distillation among exit predictions during training to improve aggregation through knowledge transfer among subnetworks. We conduct extensive experiments on benchmark CV (CIFAR-10/100, ImageNet) and NLP datasets (SST-2, AgNews). We demonstrate that ScaleFL outperforms existing representative heterogeneous FL approaches in terms of global/local model performance and provides inference efficiency, with up to 2x latency and 4x model size reduction with negligible performance drop below 2%.

Federated Learning (FL) is a distributed learning scheme to train a shared model across clients. One common and fundamental challenge in FL is that the sets of data across clients could be non-identically distributed and have different sizes. Personalized Federated Learning (PFL) attempts to solve this challenge via locally adapted models. In this work, we present a novel framework for PFL based on hierarchical Bayesian modeling and variational inference. A global model is introduced as a latent variable to augment the joint distribution of clients’ parameters and capture the common trends of different clients, optimization is derived based on the principle of maximizing the marginal likelihood and conducted using variational expectation maximization. Our algorithm gives rise to a closed-form estimation of a confidence value which comprises the uncertainty of clients’ parameters and local model deviations from the global model. The confidence value is used to weigh clients’ parameters in the aggregation stage and adjust the regularization effect of the global model. We evaluate our method through extensive empirical studies on multiple datasets. Experimental results show that our approach obtains competitive results under mild heterogeneous circumstances while significantly outperforming state-of-the-art PFL frameworks in highly heterogeneous settings.

To defend the inference attacks and mitigate the sensitive information leakages in Federated Learning (FL), client-level Differentially Private FL (DPFL) is the de-facto standard for privacy protection by clipping local updates and adding random noise. However, existing DPFL methods tend to make a sharper loss landscape and have poorer weight perturbation robustness, resulting in severe performance degradation. To alleviate these issues, we propose a novel DPFL algorithm named DP-FedSAM, which leverages gradient perturbation to mitigate the negative impact of DP. Specifically, DP-FedSAM integrates Sharpness Aware Minimization (SAM) optimizer to generate local flatness models with better stability and weight perturbation robustness, which results in the small norm of local updates and robustness to DP noise, thereby improving the performance. From the theoretical perspective, we analyze in detail how DP-FedSAM mitigates the performance degradation induced by DP. Meanwhile, we give rigorous privacy guarantees with Rényi DP and present the sensitivity analysis of local updates. At last, we empirically confirm that our algorithm achieves state-of-the-art (SOTA) performance compared with existing SOTA baselines in DPFL.

This work studies the generalization issue of face anti-spoofing (FAS) models on domain gaps, such as image resolution, blurriness and sensor variations. Most prior works regard domain-specific signals as a negative impact, and apply metric learning or adversarial losses to remove it from feature representation. Though learning a domain-invariant feature space is viable for the training data, we show that the feature shift still exists in an unseen test domain, which backfires on the generalizability of the classifier. In this work, instead of constructing a domain-invariant feature space, we encourage domain separability while aligning the live-to-spoof transition (i.e., the trajectory from live to spoof) to be the same for all domains. We formulate this FAS strategy of separability and alignment (SA-FAS) as a problem of invariant risk minimization (IRM), and learn domain-variant feature representation but domain-invariant classifier. We demonstrate the effectiveness of SA-FAS on challenging cross-domain FAS datasets and establish state-of-the-art performance.

Cross-Domain Few-Shot Learning (CD-FSL) is a recently emerging task that tackles few-shot learning across different domains. It aims at transferring prior knowledge learned on the source dataset to novel target datasets. The CD-FSL task is especially challenged by the huge domain gap between different datasets. Critically, such a domain gap actually comes from the changes of visual styles, and wave-SAN empirically shows that spanning the style distribution of the source data helps alleviate this issue. However, wave-SAN simply swaps styles of two images. Such a vanilla operation makes the generated styles “real” and “easy”, which still fall into the original set of the source styles. Thus, inspired by vanilla adversarial learning, a novel model-agnostic meta Style Adversarial training (StyleAdv) method together with a novel style adversarial attack method is proposed for CD-FSL. Particularly, our style attack method synthesizes both “virtual” and “hard” adversarial styles for model training. This is achieved by perturbing the original style with the signed style gradients. By continually attacking styles and forcing the model to recognize these challenging adversarial styles, our model is gradually robust to the visual styles, thus boosting the generalization ability for novel target datasets. Besides the typical CNN-based backbone, we also employ our StyleAdv method on large-scale pretrained vision transformer. Extensive experiments conducted on eight various target datasets show the effectiveness of our method. Whether built upon ResNet or ViT, we achieve the new state of the art for CD-FSL. Code is available at https://github.com/lovelyqian/StyleAdv-CDFSL.

Recent advancements in deploying deep neural networks (DNNs) on resource-constrained devices have generated interest in input-adaptive dynamic neural networks (DyNNs). DyNNs offer more efficient inferences and enable the deployment of DNNs on devices with limited resources, such as mobile devices. However, we have discovered a new vulnerability in DyNNs that could potentially compromise their efficiency. Specifically, we investigate whether adversaries can manipulate DyNNs’ computational costs to create a false sense of efficiency. To address this question, we propose EfficFrog, an adversarial attack that injects universal efficiency backdoors in DyNNs. To inject a backdoor trigger into DyNNs, EfficFrog poisons only a minimal percentage of the DyNNs’ training data. During the inference phase, EfficFrog can slow down the backdoored DyNNs and abuse the computational resources of systems running DyNNs by adding the trigger to any input. To evaluate EfficFrog, we tested it on three DNN backbone architectures (based on VGG16, MobileNet, and ResNet56) using two popular datasets (CIFAR-10 and Tiny ImageNet). Our results demonstrate that EfficFrog reduces the efficiency of DyNNs on triggered input samples while keeping the efficiency of clean samples almost the same.

Machine learning is vulnerable to adversarial manipulation. Previous literature has demonstrated that at the training stage attackers can manipulate data (Gu et al.) and data sampling procedures (Shumailov et al.) to control model behaviour. A common attack goal is to plant backdoors i.e. force the victim model to learn to recognise a trigger known only by the adversary. In this paper, we introduce a new class of backdoor attacks that hide inside model architectures i.e. in the inductive bias of the functions used to train. These backdoors are simple to implement, for instance by publishing open-source code for a backdoored model architecture that others will reuse unknowingly. We demonstrate that model architectural backdoors represent a real threat and, unlike other approaches, can survive a complete re-training from scratch. We formalise the main construction principles behind architectural backdoors, such as a connection between the input and the output, and describe some possible protections against them. We evaluate our attacks on computer vision benchmarks of different scales and demonstrate the underlying vulnerability is pervasive in a variety of common training settings.

Vision Transformers (ViTs), which made a splash in the field of computer vision (CV), have shaken the dominance of convolutional neural networks (CNNs). However, in the process of industrializing ViTs, backdoor attacks have brought severe challenges to security. The success of ViTs benefits from the self-attention mechanism. However, compared with CNNs, we find that this mechanism of capturing global information within patches makes ViTs more sensitive to patch-wise triggers. Under such observations, we delicately design a novel backdoor attack framework for ViTs, dubbed BadViT, which utilizes a universal patch-wise trigger to catch the model’s attention from patches beneficial for classification to those with triggers, thereby manipulating the mechanism on which ViTs survive to confuse itself. Furthermore, we propose invisible variants of BadViT to increase the stealth of the attack by limiting the strength of the trigger perturbation. Through a large number of experiments, it is proved that BadViT is an efficient backdoor attack method against ViTs, which is less dependent on the number of poisons, with satisfactory convergence, and is transferable for downstream tasks. Furthermore, the risks inside of ViTs to backdoor attacks are also explored from the perspective of existing advanced defense schemes.

Model attribution is a critical component of deep neural networks (DNNs) for its interpretability to complex models. Recent studies bring up attention to the security of attribution methods as they are vulnerable to attribution attacks that generate similar images with dramatically different attributions. Existing works have been investigating empirically improving the robustness of DNNs against those attacks; however, none of them explicitly quantifies the actual deviations of attributions. In this work, for the first time, a constrained optimization problem is formulated to derive an upper bound that measures the largest dissimilarity of attributions after the samples are perturbed by any noises within a certain region while the classification results remain the same. Based on the formulation, different practical approaches are introduced to bound the attributions above using Euclidean distance and cosine similarity under both L2 and Linf-norm perturbations constraints. The bounds developed by our theoretical study are validated on various datasets and two different types of attacks (PGD attack and IFIA attribution attack). Over 10 million attacks in the experiments indicate that the proposed upper bounds effectively quantify the robustness of models based on the worst-case attribution dissimilarities.

A hard challenge in developing practical face recognition (FR) attacks is due to the black-box nature of the target FR model, i.e., inaccessible gradient and parameter information to attackers. While recent research took an important step towards attacking black-box FR models through leveraging transferability, their performance is still limited, especially against online commercial FR systems that can be pessimistic (e.g., a less than 50% ASR--attack success rate on average). Motivated by this, we present Sibling-Attack, a new FR attack technique for the first time explores a novel multi-task perspective (i.e., leveraging extra information from multi-correlated tasks to boost attacking transferability). Intuitively, Sibling-Attack selects a set of tasks correlated with FR and picks the Attribute Recognition (AR) task as the task used in Sibling-Attack based on theoretical and quantitative analysis. Sibling-Attack then develops an optimization framework that fuses adversarial gradient information through (1) constraining the cross-task features to be under the same space, (2) a joint-task meta optimization framework that enhances the gradient compatibility among tasks, and (3) a cross-task gradient stabilization method which mitigates the oscillation effect during attacking. Extensive experiments demonstrate that Sibling-Attack outperforms state-of-the-art FR attack techniques by a non-trivial margin, boosting ASR by 12.61% and 55.77% on average on state-of-the-art pre-trained FR models and two well-known, widely used commercial FR systems.

Deep learning models have shown extreme vulnerability to simple perturbations and spatial transformations. In this work, we explore whether we can adopt the characteristics of adversarial attack methods to help improve perturbation robustness for object detection. We study a class of realistic object detection settings wherein the target objects have control over their appearance. To this end, we propose a reversed Fast Gradient Sign Method (FGSM) to obtain these angelic patches}that significantly increase the detection probability, even without pre-knowledge of the perturbations. In detail, we apply the patch to each object instance simultaneously, strengthen not only classification but also bounding box accuracy. Experiments demonstrate the efficacy of the partial-covering patch in solving the complex bounding box problem. More importantly, the performance is also transferable to different detection models even under severe affine transformations and deformable shapes. To our knowledge, we are the first (object detection) patch that achieves both cross-model and multiple-patch efficacy. We observed average accuracy improvements of 30% in the real-world experiments, which brings large social value. Our code is available at: https://github.com/averysi224/angelic_patches.

Deep neural networks are widely known to be susceptible to adversarial examples, which can cause incorrect predictions through subtle input modifications. These adversarial examples tend to be transferable between models, but targeted attacks still have lower attack success rates due to significant variations in decision boundaries. To enhance the transferability of targeted adversarial examples, we propose introducing competition into the optimization process. Our idea is to craft adversarial perturbations in the presence of two new types of competitor noises: adversarial perturbations towards different target classes and friendly perturbations towards the correct class. With these competitors, even if an adversarial example deceives a network to extract specific features leading to the target class, this disturbance can be suppressed by other competitors. Therefore, within this competition, adversarial examples should take different attack strategies by leveraging more diverse features to overwhelm their interference, leading to improving their transferability to different models. Considering the computational complexity, we efficiently simulate various interference from these two types of competitors in feature space by randomly mixing up stored clean features in the model inference and named this method Clean Feature Mixup (CFM). Our extensive experimental results on the ImageNet-Compatible and CIFAR-10 datasets show that the proposed method outperforms the existing baselines with a clear margin. Our code is available at https://github.com/dreamflake/CFM.

Model robustness against adversarial examples of single perturbation type such as the Lp-norm has been widely studied, yet its generalization to more realistic scenarios involving multiple semantic perturbations and their composition remains largely unexplored. In this paper, we first propose a novel method for generating composite adversarial examples. Our method can find the optimal attack composition by utilizing component-wise projected gradient descent and automatic attack-order scheduling. We then propose generalized adversarial training (GAT) to extend model robustness from Lp-ball to composite semantic perturbations, such as the combination of Hue, Saturation, Brightness, Contrast, and Rotation. Results obtained using ImageNet and CIFAR-10 datasets indicate that GAT can be robust not only to all the tested types of a single attack, but also to any combination of such attacks. GAT also outperforms baseline L-infinity-norm bounded adversarial training approaches by a significant margin.

Distilled student models in teacher-student architectures are widely considered for computational-effective deployment in real-time applications and edge devices. However, there is a higher risk of student models to encounter adversarial attacks at the edge. Popular enhancing schemes such as adversarial training have limited performance on compressed networks. Thus, recent studies concern about adversarial distillation (AD) that aims to inherit not only prediction accuracy but also adversarial robustness of a robust teacher model under the paradigm of robust optimization. In the min-max framework of AD, existing AD methods generally use fixed supervision information from the teacher model to guide the inner optimization for knowledge distillation which often leads to an overcorrection towards model smoothness. In this paper, we propose an adaptive adversarial distillation (AdaAD) that involves the teacher model in the knowledge optimization process in a way interacting with the student model to adaptively search for the inner results. Comparing with state-of-the-art methods, the proposed AdaAD can significantly boost both the prediction accuracy and adversarial robustness of student models in most scenarios. In particular, the ResNet-18 model trained by AdaAD achieves top-rank performance (54.23% robust accuracy) on RobustBench under AutoAttack.

Although current deep learning techniques have yielded superior performance on various computer vision tasks, yet they are still vulnerable to adversarial examples. Adversarial training and its variants have been shown to be the most effective approaches to defend against adversarial examples. A particular class of these methods regularize the difference between output probabilities for an adversarial and its corresponding natural example. However, it may have a negative impact if a natural example is misclassified. To circumvent this issue, we propose a novel adversarial training scheme that encourages the model to produce similar output probabilities for an adversarial example and its “inverse adversarial” counterpart. Particularly, the counterpart is generated by maximizing the likelihood in the neighborhood of the natural example. Extensive experiments on various vision datasets and architectures demonstrate that our training method achieves state-of-the-art robustness as well as natural accuracy among robust models. Furthermore, using a universal version of inverse adversarial examples, we improve the performance of single-step adversarial training techniques at a low computational cost.

This paper addresses the problem of robust deep single-image reflection removal (SIRR) against adversarial attacks. Current deep learning based SIRR methods have shown significant performance degradation due to unnoticeable distortions and perturbations on input images. For a comprehensive robustness study, we first conduct diverse adversarial attacks specifically for the SIRR problem, i.e. towards different attacking targets and regions. Then we propose a robust SIRR model, which integrates the cross-scale attention module, the multi-scale fusion module, and the adversarial image discriminator. By exploiting the multi-scale mechanism, the model narrows the gap between features from clean and adversarial images. The image discriminator adaptively distinguishes clean or noisy inputs, and thus further gains reliable robustness. Extensive experiments on Nature, SIR^2, and Real datasets demonstrate that our model remarkably improves the robustness of SIRR across disparate scenes.

The success rate of current adversarial attacks remains low on real-world 3D face recognition tasks because the 3D-printing attacks need to meet the requirement that the generated points should be adjacent to the surface, which limits the adversarial example’ searching space. Additionally, they have not considered unpredictable head movements or the non-homogeneous nature of skin reflectance in the real world. To address the real-world challenges, we propose a novel structured-light attack against structured-light-based 3D face recognition. We incorporate the 3D reconstruction process and skin’s reflectance in the optimization process to get the end-to-end attack and present 3D transform invariant loss and sensitivity maps to improve robustness. Our attack enables adversarial points to be placed in any position and is resilient to random head movements while maintaining the perturbation unnoticeable. Experiments show that our new method can attack point-cloud-based and depth-image-based 3D face recognition systems with a high success rate, using fewer perturbations than previous physical 3D adversarial attacks.

Face forgery detection becomes increasingly crucial due to the serious security issues caused by face manipulation techniques. Recent studies in deepfake detection have yielded promising results when the training and testing face forgeries are from the same domain. However, the problem remains challenging when one tries to generalize the detector to forgeries created by unseen methods during training. Observing that face manipulation may alter the relation between different facial action units (AU), we propose the Action Units Relation Learning framework to improve the generality of forgery detection. In specific, it consists of the Action Units Relation Transformer (ART) and the Tampered AU Prediction (TAP). The ART constructs the relation between different AUs with AU-agnostic Branch and AU-specific Branch, which complement each other and work together to exploit forgery clues. In the Tampered AU Prediction, we tamper AU-related regions at the image level and develop challenging pseudo samples at the feature level. The model is then trained to predict the tampered AU regions with the generated location-specific supervision. Experimental results demonstrate that our method can achieve state-of-the-art performance in both the in-dataset and cross-dataset evaluations.