We present a new imaging technique, swept-angle synthetic wavelength interferometry, for full-field micron-scale 3D sensing. As in conventional synthetic wavelength interferometry, our technique uses light consisting of two narrowly-separated optical wavelengths, resulting in per-pixel interferometric measurements whose phase encodes scene depth. Our technique additionally uses a new type of light source that, by emulating spatially-incoherent illumination, makes interferometric measurements insensitive to aberrations and (sub)surface scattering, effects that corrupt phase measurements. The resulting technique combines the robustness to such corruptions of scanning interferometric setups, with the speed of full-field interferometric setups. Overall, our technique can recover full-frame depth at a lateral and axial resolution of 5 microns, at frame rates of 5 Hz, even under strong ambient light. We build an experimental prototype, and use it to demonstrate these capabilities by scanning a variety of objects, including objects representative of applications in inspection and fabrication, and objects that contain challenging light scattering effects.

We present Reference-guided Super-Resolution Neural Radiance Field (RefSR-NeRF) that extends NeRF to super resolution and photorealistic novel view synthesis. Despite NeRF’s extraordinary success in the neural rendering field, it suffers from blur in high resolution rendering because its inherent multilayer perceptron struggles to learn high frequency details and incurs a computational explosion as resolution increases. Therefore, we propose RefSR-NeRF, an end-to-end framework that first learns a low resolution NeRF representation, and then reconstructs the high frequency details with the help of a high resolution reference image. We observe that simply introducing the pre-trained models from the literature tends to produce unsatisfied artifacts due to the divergence in the degradation model. To this end, we design a novel lightweight RefSR model to learn the inverse degradation process from NeRF renderings to target HR ones. Extensive experiments on multiple benchmarks demonstrate that our method exhibits an impressive trade-off among rendering quality, speed, and memory usage, outperforming or on par with NeRF and its variants while being 52× speedup with minor extra memory usage.

Novel view synthesis with sparse inputs is a challenging problem for neural radiance fields (NeRF). Recent efforts alleviate this challenge by introducing external supervision, such as pre-trained models and extra depth signals, or by using non-trivial patch-based rendering. In this paper, we present Frequency regularized NeRF (FreeNeRF), a surprisingly simple baseline that outperforms previous methods with minimal modifications to plain NeRF. We analyze the key challenges in few-shot neural rendering and find that frequency plays an important role in NeRF’s training. Based on this analysis, we propose two regularization terms: one to regularize the frequency range of NeRF’s inputs, and the other to penalize the near-camera density fields. Both techniques are “free lunches” that come at no additional computational cost. We demonstrate that even with just one line of code change, the original NeRF can achieve similar performance to other complicated methods in the few-shot setting. FreeNeRF achieves state-of-the-art performance across diverse datasets, including Blender, DTU, and LLFF. We hope that this simple baseline will motivate a rethinking of the fundamental role of frequency in NeRF’s training, under both the low-data regime and beyond. This project is released at https://jiawei-yang.github.io/FreeNeRF/.

Neural Radiance Fields (NeRF) have achieved photorealistic novel views synthesis; however, the requirement of accurate camera poses limits its application. Despite analysis-by-synthesis extensions for jointly learning neural 3D representations and registering camera frames exist, they are susceptible to suboptimal solutions if poorly initialized. We propose L2G-NeRF, a Local-to-Global registration method for bundle-adjusting Neural Radiance Fields: first, a pixel-wise flexible alignment, followed by a frame-wise constrained parametric alignment. Pixel-wise local alignment is learned in an unsupervised way via a deep network which optimizes photometric reconstruction errors. Frame-wise global alignment is performed using differentiable parameter estimation solvers on the pixel-wise correspondences to find a global transformation. Experiments on synthetic and real-world data show that our method outperforms the current state-of-the-art in terms of high-fidelity reconstruction and resolving large camera pose misalignment. Our module is an easy-to-use plugin that can be applied to NeRF variants and other neural field applications.

We address efficient and structure-aware 3D scene representation from images. Nerflets are our key contribution-- a set of local neural radiance fields that together represent a scene. Each nerflet maintains its own spatial position, orientation, and extent, within which it contributes to panoptic, density, and radiance reconstructions. By leveraging only photometric and inferred panoptic image supervision, we can directly and jointly optimize the parameters of a set of nerflets so as to form a decomposed representation of the scene, where each object instance is represented by a group of nerflets. During experiments with indoor and outdoor environments, we find that nerflets: (1) fit and approximate the scene more efficiently than traditional global NeRFs, (2) allow the extraction of panoptic and photometric renderings from arbitrary views, and (3) enable tasks rare for NeRFs, such as 3D panoptic segmentation and interactive editing.

Dynamic Neural Radiance Field (NeRF) is a powerful algorithm capable of rendering photo-realistic novel view images from a monocular RGB video of a dynamic scene. Although it warps moving points across frames from the observation spaces to a common canonical space for rendering, dynamic NeRF does not model the change of the reflected color during the warping. As a result, this approach often fails drastically on challenging specular objects in motion. We address this limitation by reformulating the neural radiance field function to be conditioned on surface position and orientation in the observation space. This allows the specular surface at different poses to keep the different reflected colors when mapped to the common canonical space. Additionally, we add the mask of moving objects to guide the deformation field. As the specular surface changes color during motion, the mask mitigates the problem of failure to find temporal correspondences with only RGB supervision. We evaluate our model based on the novel view synthesis quality with a self-collected dataset of different moving specular objects in realistic environments. The experimental results demonstrate that our method significantly improves the reconstruction quality of moving specular objects from monocular RGB videos compared to the existing NeRF models. Our code and data are available at the project website https://github.com/JokerYan/NeRF-DS.

Purely MLP-based neural radiance fields (NeRF-based methods) often suffer from underfitting with blurred renderings on large-scale scenes due to limited model capacity. Recent approaches propose to geographically divide the scene and adopt multiple sub-NeRFs to model each region individually, leading to linear scale-up in training costs and the number of sub-NeRFs as the scene expands. An alternative solution is to use a feature grid representation, which is computationally efficient and can naturally scale to a large scene with increased grid resolutions. However, the feature grid tends to be less constrained and often reaches suboptimal solutions, producing noisy artifacts in renderings, especially in regions with complex geometry and texture. In this work, we present a new framework that realizes high-fidelity rendering on large urban scenes while being computationally efficient. We propose to use a compact multi-resolution ground feature plane representation to coarsely capture the scene, and complement it with positional encoding inputs through another NeRF branch for rendering in a joint learning fashion. We show that such an integration can utilize the advantages of two alternative solutions: a light-weighted NeRF is sufficient, under the guidance of the feature grid representation, to render photorealistic novel views with fine details; and the jointly optimized ground feature planes, can meanwhile gain further refinements, forming a more accurate and compact feature space and output much more natural rendering results.

Neural radiance fields (NeRFs) enable novel view synthesis with unprecedented visual quality. However, to render photorealistic images, NeRFs require hundreds of deep multilayer perceptron (MLP) evaluations -- for each pixel. This is prohibitively expensive and makes real-time rendering infeasible, even on powerful modern GPUs. In this paper, we propose a novel approach to distill and bake NeRFs into highly efficient mesh-based neural representations that are fully compatible with the massively parallel graphics rendering pipeline. We represent scenes as neural radiance features encoded on a two-layer duplex mesh, which effectively overcomes the inherent inaccuracies in 3D surface reconstruction by learning the aggregated radiance information from a reliable interval of ray-surface intersections. To exploit local geometric relationships of nearby pixels, we leverage screen-space convolutions instead of the MLPs used in NeRFs to achieve high-quality appearance. Finally, the performance of the whole framework is further boosted by a novel multi-view distillation optimization strategy. We demonstrate the effectiveness and superiority of our approach via extensive experiments on a range of standard datasets.

Neural radiance fields (NeRF) achieve highly photo-realistic novel-view synthesis, but it’s a challenging problem to edit the scenes modeled by NeRF-based methods, especially for dynamic scenes. We propose editable neural radiance fields that enable end-users to easily edit dynamic scenes and even support topological changes. Input with an image sequence from a single camera, our network is trained fully automatically and models topologically varying dynamics using our picked-out surface key points. Then end-users can edit the scene by easily dragging the key points to desired new positions. To achieve this, we propose a scene analysis method to detect and initialize key points by considering the dynamics in the scene, and a weighted key points strategy to model topologically varying dynamics by joint key points and weights optimization. Our method supports intuitive multi-dimensional (up to 3D) editing and can generate novel scenes that are unseen in the input sequence. Experiments demonstrate that our method achieves high-quality editing on various dynamic scenes and outperforms the state-of-the-art. Our code and captured data are available at https://chengwei-zheng.github.io/EditableNeRF/.

Recent efforts in Neural Rendering Fields (NeRF) have shown impressive results on novel view synthesis by utilizing implicit neural representation to represent 3D scenes. Due to the process of volumetric rendering, the inference speed for NeRF is extremely slow, limiting the application scenarios of utilizing NeRF on resource-constrained hardware, such as mobile devices. Many works have been conducted to reduce the latency of running NeRF models. However, most of them still require high-end GPU for acceleration or extra storage memory, which is all unavailable on mobile devices. Another emerging direction utilizes the neural light field (NeLF) for speedup, as only one forward pass is performed on a ray to predict the pixel color. Nevertheless, to reach a similar rendering quality as NeRF, the network in NeLF is designed with intensive computation, which is not mobile-friendly. In this work, we propose an efficient network that runs in real-time on mobile devices for neural rendering. We follow the setting of NeLF to train our network. Unlike existing works, we introduce a novel network architecture that runs efficiently on mobile devices with low latency and small size, i.e., saving 15× ~ 24× storage compared with MobileNeRF. Our model achieves high-resolution generation while maintaining real-time inference for both synthetic and real-world scenes on mobile devices, e.g., 18.04ms (iPhone 13) for rendering one 1008×756 image of real 3D scenes. Additionally, we achieve similar image quality as NeRF and better quality than MobileNeRF (PSNR 26.15 vs. 25.91 on the real-world forward-facing dataset).

3D style transfer aims to render stylized novel views of a 3D scene with multi-view consistency. However, most existing work suffers from a three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to arbitrary new styles. We propose StyleRF (Style Radiance Fields), an innovative 3D style transfer technique that resolves the three-way dilemma by performing style transformation within the feature space of a radiance field. StyleRF employs an explicit grid of high-level features to represent 3D scenes, with which high-fidelity geometry can be reliably restored via volume rendering. In addition, it transforms the grid features according to the reference style which directly leads to high-quality zero-shot style transfer. StyleRF consists of two innovative designs. The first is sampling-invariant content transformation that makes the transformation invariant to the holistic statistics of the sampled 3D points and accordingly ensures multi-view consistency. The second is deferred style transformation of 2D feature maps which is equivalent to the transformation of 3D points but greatly reduces memory footprint without degrading multi-view consistency. Extensive experiments show that StyleRF achieves superior 3D stylization quality with precise geometry reconstruction and it can generalize to various new styles in a zero-shot manner. Project website: https://kunhao-liu.github.io/StyleRF/

Synthesizing photo-realistic images from a point cloud is challenging because of the sparsity of point cloud representation. Recent Neural Radiance Fields and extensions are proposed to synthesize realistic images from 2D input. In this paper, we present Point2Pix as a novel point renderer to link the 3D sparse point clouds with 2D dense image pixels. Taking advantage of the point cloud 3D prior and NeRF rendering pipeline, our method can synthesize high-quality images from colored point clouds, generally for novel indoor scenes. To improve the efficiency of ray sampling, we propose point-guided sampling, which focuses on valid samples. Also, we present Point Encoding to build Multi-scale Radiance Fields that provide discriminative 3D point features. Finally, we propose Fusion Encoding to efficiently synthesize high-quality images. Extensive experiments on the ScanNet and ArkitScenes datasets demonstrate the effectiveness and generalization.

We propose a novel method that renders point clouds as if they are surfaces. The proposed method is differentiable and requires no scene-specific optimization. This unique capability enables, out-of-the-box, surface normal estimation, rendering room-scale point clouds, inverse rendering, and ray tracing with global illumination. Unlike existing work that focuses on converting point clouds to other representations--e.g., surfaces or implicit functions--our key idea is to directly infer the intersection of a light ray with the underlying surface represented by the given point cloud. Specifically, we train a set transformer that, given a small number of local neighbor points along a light ray, provides the intersection point, the surface normal, and the material blending weights, which are used to render the outcome of this light ray. Localizing the problem into small neighborhoods enables us to train a model with only 48 meshes and apply it to unseen point clouds. Our model achieves higher estimation accuracy than state-of-the-art surface reconstruction and point-cloud rendering methods on three test sets. When applied to room-scale point clouds, without any scene-specific optimization, the model achieves competitive quality with the state-of-the-art novel-view rendering methods. Moreover, we demonstrate ability to render and manipulate Lidar-scanned point clouds such as lighting control and object insertion.

Reconstruction and intrinsic decomposition of scenes from captured imagery would enable many applications such as relighting and virtual object insertion. Recent NeRF based methods achieve impressive fidelity of 3D reconstruction, but bake the lighting and shadows into the radiance field, while mesh-based methods that facilitate intrinsic decomposition through differentiable rendering have not yet scaled to the complexity and scale of outdoor scenes. We present a novel inverse rendering framework for large urban scenes capable of jointly reconstructing the scene geometry, spatially-varying materials, and HDR lighting from a set of posed RGB images with optional depth. Specifically, we use a neural field to account for the primary rays, and use an explicit mesh (reconstructed from the underlying neural field) for modeling secondary rays that produce higher-order lighting effects such as cast shadows. By faithfully disentangling complex geometry and materials from lighting effects, our method enables photorealistic relighting with specular and shadow effects on several outdoor datasets. Moreover, it supports physics-based scene manipulations such as virtual object insertion with ray-traced shadow casting.

Uncalibrated photometric stereo (UPS) is challenging due to the inherent ambiguity brought by the unknown light. Although the ambiguity is alleviated on non-Lambertian objects, the problem is still difficult to solve for more general objects with complex shapes introducing irregular shadows and general materials with complex reflectance like anisotropic reflectance. To exploit cues from shadow and reflectance to solve UPS and improve performance on general materials, we propose DANI-Net, an inverse rendering framework with differentiable shadow handling and anisotropic reflectance modeling. Unlike most previous methods that use non-differentiable shadow maps and assume isotropic material, our network benefits from cues of shadow and anisotropic reflectance through two differentiable paths. Experiments on multiple real-world datasets demonstrate our superior and robust performance.

We propose a scene-level inverse rendering framework that uses multi-view images to decompose the scene into geometry, a SVBRDF, and 3D spatially-varying lighting. Because multi-view images provide a variety of information about the scene, multi-view images in object-level inverse rendering have been taken for granted. However, owing to the absence of multi-view HDR synthetic dataset, scene-level inverse rendering has mainly been studied using single-view image. We were able to successfully perform scene-level inverse rendering using multi-view images by expanding OpenRooms dataset and designing efficient pipelines to handle multi-view images, and splitting spatially-varying lighting. Our experiments show that the proposed method not only achieves better performance than single-view-based methods, but also achieves robust performance on unseen real-world scene. Also, our sophisticated 3D spatially-varying lighting volume allows for photorealistic object insertion in any 3D location.

We present a learning-based approach to relight a single image of Lambertian and low-frequency specular objects. Our method enables inserting objects from photographs into new scenes and relighting them under the new environment lighting, which is essential for AR applications. To relight the object, we solve both inverse rendering and re-rendering. To resolve the ill-posed inverse rendering, we propose a weakly-supervised method by a low-rank constraint. To facilitate the weakly-supervised training, we contribute Relit, a large-scale (750K images) dataset of videos with aligned objects under changing illuminations. For re-rendering, we propose a differentiable specular rendering layer to render low-frequency non-Lambertian materials under various illuminations of spherical harmonics. The whole pipeline is end-to-end and efficient, allowing for a mobile app implementation of AR object insertion. Extensive evaluations demonstrate that our method achieves state-of-the-art performance. Project page: https://renjiaoyi.github.io/relighting/.

We introduce light diffusion, a novel method to improve lighting in portraits, softening harsh shadows and specular highlights while preserving overall scene illumination. Inspired by professional photographers’ diffusers and scrims, our method softens lighting given only a single portrait photo. Previous portrait relighting approaches focus on changing the entire lighting environment, removing shadows (ignoring strong specular highlights), or removing shading entirely. In contrast, we propose a learning based method that allows us to control the amount of light diffusion and apply it on in-the-wild portraits. Additionally, we design a method to synthetically generate plausible external shadows with sub-surface scattering effects while conforming to the shape of the subject’s face. Finally, we show how our approach can increase the robustness of higher level vision applications, such as albedo estimation, geometry estimation and semantic segmentation.

We address the challenge of recovering an underlying scene geometry and colors from a sparse set of RGBD view observations. In this work, we present a new solution termed RGBD2 that sequentially generates novel RGBD views along a camera trajectory, and the scene geometry is simply the fusion result of these views. More specifically, we maintain an intermediate surface mesh used for rendering new RGBD views, which subsequently becomes complete by an inpainting network; each rendered RGBD view is later back-projected as a partial surface and is supplemented into the intermediate mesh. The use of intermediate mesh and camera projection helps solve the tough problem of multi-view inconsistency. We practically implement the RGBD inpainting network as a versatile RGBD diffusion model, which is previously used for 2D generative modeling; we make a modification to its reverse diffusion process to enable our use. We evaluate our approach on the task of 3D scene synthesis from sparse RGBD inputs; extensive experiments on the ScanNet dataset demonstrate the superiority of our approach over existing ones. Project page: https://jblei.site/proj/rgbd-diffusion.

Recent methods for 3D reconstruction and rendering increasingly benefit from end-to-end optimization of the entire image formation process. However, this approach is currently limited: effects of the optical hardware stack and in particular lenses are hard to model in a unified way. This limits the quality that can be achieved for camera calibration and the fidelity of the results of 3D reconstruction. In this paper, we propose NeuroLens, a neural lens model for distortion and vignetting that can be used for point projection and ray casting and can be optimized through both operations. This means that it can (optionally) be used to perform pre-capture calibration using classical calibration targets, and can later be used to perform calibration or refinement during 3D reconstruction, e.g., while optimizing a radiance field. To evaluate the performance of our proposed model, we create a comprehensive dataset assembled from the Lensfun database with a multitude of lenses. Using this and other real-world datasets, we show that the quality of our proposed lens model outperforms standard packages as well as recent approaches while being much easier to use and extend. The model generalizes across many lens types and is trivial to integrate into existing 3D reconstruction and rendering systems. Visit our project website at: https://neural-lens.github.io.

We consider the problem of reconstructing a full 360° photographic model of an object from a single image of it. We do so by fitting a neural radiance field to the image, but find this problem to be severely ill-posed. We thus take an off-the-self conditional image generator based on diffusion and engineer a prompt that encourages it to “dream up” novel views of the object. Using the recent DreamFusion method, we fuse the given input view, the conditional prior, and other regularizers in a final, consistent reconstruction. We demonstrate state-of-the-art reconstruction results on benchmark images when compared to prior methods for monocular 3D reconstruction of objects. Qualitatively, our reconstructions provide a faithful match of the input view and a plausible extrapolation of its appearance and 3D shape, including to the side of the object not visible in the image.

Neural surface reconstruction has been shown to be powerful for recovering dense 3D surfaces via image-based neural rendering. However, current methods struggle to recover detailed structures of real-world scenes. To address the issue, we present Neuralangelo, which combines the representation power of multi-resolution 3D hash grids with neural surface rendering. Two key ingredients enable our approach: (1) numerical gradients for computing higher-order derivatives as a smoothing operation and (2) coarse-to-fine optimization on the hash grids controlling different levels of details. Even without auxiliary inputs such as depth, Neuralangelo can effectively recover dense 3D surface structures from multi-view images with fidelity significantly surpassing previous methods, enabling detailed large-scale scene reconstruction from RGB video captures.

Neural radiance-density field methods have become increasingly popular for the task of novel-view rendering. Their recent extension to hash-based positional encoding ensures fast training and inference with visually pleasing results. However, density-based methods struggle with recovering accurate surface geometry. Hybrid methods alleviate this issue by optimizing the density based on an underlying SDF. However, current SDF methods are overly smooth and miss fine geometric details. In this work, we combine the strengths of these two lines of work in a novel hash-based implicit surface representation. We propose improvements to the two areas by replacing the voxel hash encoding with a permutohedral lattice which optimizes faster, especially for higher dimensions. We additionally propose a regularization scheme which is crucial for recovering high-frequency geometric detail. We evaluate our method on multiple datasets and show that we can recover geometric detail at the level of pores and wrinkles while using only RGB images for supervision. Furthermore, using sphere tracing we can render novel views at 30 fps on an RTX 3090. Code is publicly available at https://radualexandru.github.io/permuto_sdf

This paper studies implicit surface reconstruction leveraging differentiable ray casting. Previous works such as IDR and NeuS overlook the spatial context in 3D space when predicting and rendering the surface, thereby may fail to capture sharp local topologies such as small holes and structures. To mitigate the limitation, we propose a flexible neural implicit representation leveraging hierarchical voxel grids, namely Neural Deformable Anchor (NeuDA), for high-fidelity surface reconstruction. NeuDA maintains the hierarchical anchor grids where each vertex stores a 3d position (or anchor) instead of the direct embedding (or feature). We optimize the anchor grids such that different local geometry structures can be adaptively encoded. Besides, we dig into the frequency encoding strategies and introduce a simple hierarchical positional encoding method for the hierarchical anchor structure to flexibly exploited the properties of high-frequency and low-frequency geometry and appearance. Experiments on both the DTU and BlendedMVS datasets demonstrate that NeuDA can produce promising mesh surfaces.

We study the problem of reconstructing 3D feature curves of an object from a set of calibrated multi-view images. To do so, we learn a neural implicit field representing the density distribution of 3D edges which we refer to as Neural Edge Field (NEF). Inspired by NeRF, NEF is optimized with a view-based rendering loss where a 2D edge map is rendered at a given view and is compared to the ground-truth edge map extracted from the image of that view. The rendering-based differentiable optimization of NEF fully exploits 2D edge detection, without needing a supervision of 3D edges, a 3D geometric operator or cross-view edge correspondence. Several technical designs are devised to ensure learning a range-limited and view-independent NEF for robust edge extraction. The final parametric 3D curves are extracted from NEF with an iterative optimization method. On our benchmark with synthetic data, we demonstrate that NEF outperforms existing state-of-the-art methods on all metrics. Project page: https://yunfan1202.github.io/NEF/.

Automatically generating high-quality real world 3D scenes is of enormous interest for applications such as virtual reality and robotics simulation. Towards this goal, we introduce NeuralField-LDM, a generative model capable of synthesizing complex 3D environments. We leverage Latent Diffusion Models that have been successfully utilized for efficient high-quality 2D content creation. We first train a scene auto-encoder to express a set of image and pose pairs as a neural field, represented as density and feature voxel grids that can be projected to produce novel views of the scene. To further compress this representation, we train a latent-autoencoder that maps the voxel grids to a set of latent representations. A hierarchical diffusion model is then fit to the latents to complete the scene generation pipeline. We achieve a substantial improvement over existing state-of-the-art scene generation models. Additionally, we show how NeuralField-LDM can be used for a variety of 3D content creation applications, including conditional scene generation, scene inpainting and scene style manipulation.

Generative models have shown great promise in synthesizing photorealistic 3D objects, but they require large amounts of training data. We introduce SinGRAF, a 3D-aware generative model that is trained with a few input images of a single scene. Once trained, SinGRAF generates different realizations of this 3D scene that preserve the appearance of the input while varying scene layout. For this purpose, we build on recent progress in 3D GAN architectures and introduce a novel progressive-scale patch discrimination approach during training. With several experiments, we demonstrate that the results produced by SinGRAF outperform the closest related works in both quality and diversity by a large margin.

We introduce a novel approach that takes a single semantic mask as input to synthesize multi-view consistent color images of natural scenes, trained with a collection of single images from the Internet. Prior works on 3D-aware image synthesis either require multi-view supervision or learning category-level prior for specific classes of objects, which are inapplicable to natural scenes. Our key idea to solve this challenge is to use a semantic field as the intermediate representation, which is easier to reconstruct from an input semantic mask and then translated to a radiance field with the assistance of off-the-shelf semantic image synthesis models. Experiments show that our method outperforms baseline methods and produces photorealistic and multi-view consistent videos of a variety of natural scenes. The project website is https://zju3dv.github.io/paintingnature/.

While 3D-based GAN techniques have been successfully applied to render photo-realistic 3D images with a variety of attributes while preserving view consistency, there has been little research on how to fine-control 3D images without limiting to a specific category of objects of their properties. To fill such research gap, we propose a novel image manipulation model of 3D-based GAN representations for a fine-grained control of specific custom attributes. By extending the latest 3D-based GAN models (e.g., EG3D), our user-friendly quantitative manipulation model enables a fine yet normalized control of 3D manipulation of multi-attribute quantities while achieving view consistency. We validate the effectiveness of our proposed technique both qualitatively and quantitatively through various experiments.

Nerf-based Generative models have shown impressive capacity in generating high-quality images with consistent 3D geometry. Despite successful synthesis of fake identity images randomly sampled from latent space, adopting these models for generating face images of real subjects is still a challenging task due to its so-called inversion issue. In this paper, we propose a universal method to surgically fine-tune these NeRF-GAN models in order to achieve high-fidelity animation of real subjects only by a single image. Given the optimized latent code for an out-of-domain real image, we employ 2D loss functions on the rendered image to reduce the identity gap. Furthermore, our method leverages explicit and implicit 3D regularizations using the in-domain neighborhood samples around the optimized latent code to remove geometrical and visual artifacts. Our experiments confirm the effectiveness of our method in realistic, high-fidelity, and 3D consistent animation of real faces on multiple NeRF-GAN models across different datasets.

We study the 3D-aware image attribute editing problem in this paper, which has wide applications in practice. Recent methods solved the problem by training a shared encoder to map images into a 3D generator’s latent space or by per-image latent code optimization and then edited images in the latent space. Despite their promising results near the input view, they still suffer from the 3D inconsistency of produced images at large camera poses and imprecise image attribute editing, like affecting unspecified attributes during editing. For more efficient image inversion, we train a shared encoder for all images. To alleviate 3D inconsistency at large camera poses, we propose two novel methods, an alternating training scheme and a multi-view identity loss, to maintain 3D consistency and subject identity. As for imprecise image editing, we attribute the problem to the gap between the latent space of real images and that of generated images. We compare the latent space and inversion manifold of GAN models and demonstrate that editing in the inversion manifold can achieve better results in both quantitative and qualitative evaluations. Extensive experiments show that our method produces more 3D consistent images and achieves more precise image editing than previous work. Source code and pretrained models can be found on our project page: https://mybabyyh.github.io/Preim3D.

Representing a 3D shape with a set of primitives can aid perception of structure, improve robotic object manipulation, and enable editing, stylization, and compression of 3D shapes. Existing methods either use simple parametric primitives or learn a generative shape space of parts. Both have limitations: parametric primitives lead to coarse approximations, while learned parts offer too little control over the decomposition. We instead propose to decompose shapes using a library of 3D parts provided by the user, giving full control over the choice of parts. The library can contain parts with high-quality geometry that are suitable for a given category, resulting in meaningful decom- positions with clean geometry. The type of decomposition can also be controlled through the choice of parts in the library. Our method works via a unsupervised approach that iteratively retrieves parts from the library and refines their placements. We show that this approach gives higher reconstruction accuracy and more desirable decompositions than existing approaches. Additionally, we show how the decom- position can be controlled through the part library by using different part libraries to reconstruct the same shapes.

In this paper, we propose DiffSwap, a diffusion model based framework for high-fidelity and controllable face swapping. Unlike previous work that relies on carefully designed network architectures and loss functions to fuse the information from the source and target faces, we reformulate the face swapping as a conditional inpainting task, performed by a powerful diffusion model guided by the desired face attributes (e.g., identity and landmarks). An important issue that makes it nontrivial to apply diffusion models to face swapping is that we cannot perform the time-consuming multi-step sampling to obtain the generated image during training. To overcome this, we propose a midpoint estimation method to efficiently recover a reasonable diffusion result of the swapped face with only 2 steps, which enables us to introduce identity constraints to improve the face swapping quality. Our framework enjoys several favorable properties more appealing than prior arts: 1) Controllable. Our method is based on conditional masked diffusion on the latent space, where the mask and the conditions can be fully controlled and customized. 2) High-fidelity. The formulation of conditional inpainting can fully exploit the generative ability of diffusion models and can preserve the background of target images with minimal artifacts. 3) Shape-preserving. The controllability of our method enables us to use 3D-aware landmarks as the condition during generation to preserve the shape of the source face. Extensive experiments on both FF++ and FFHQ demonstrate that our method can achieve state-of-the-art face swapping results both qualitatively and quantitatively.

We present a novel paradigm for high-fidelity face swapping that faithfully preserves the desired subtle geometry and texture details. We rethink face swapping from the perspective of fine-grained face editing, i.e., editing for swapping (E4S), and propose a framework that is based on the explicit disentanglement of the shape and texture of facial components. Following the E4S principle, our framework enables both global and local swapping of facial features, as well as controlling the amount of partial swapping specified by the user. Furthermore, the E4S paradigm is inherently capable of handling facial occlusions by means of facial masks. At the core of our system lies a novel Regional GAN Inversion (RGI) method, which allows the explicit disentanglement of shape and texture. It also allows face swapping to be performed in the latent space of StyleGAN. Specifically, we design a multi-scale mask-guided encoder to project the texture of each facial component into regional style codes. We also design a mask-guided injection module to manipulate the feature maps with the style codes. Based on the disentanglement, face swapping is reformulated as a simplified problem of style and mask swapping. Extensive experiments and comparisons with current state-of-the-art methods demonstrate the superiority of our approach in preserving texture and shape details, as well as working with high resolution images. The project page is https://e4s2022.github.io

Face attribute research has so far used only simple binary attributes for facial hair; e.g., beard / no beard. We have created a new, more descriptive facial hair annotation scheme and applied it to create a new facial hair attribute dataset, FH37K. Face attribute research also so far has not dealt with logical consistency and completeness. For example, in prior research, an image might be classified as both having no beard and also having a goatee (a type of beard). We show that the test accuracy of previous classification methods on facial hair attribute classification drops significantly if logical consistency of classifications is enforced. We propose a logically consistent prediction loss, LCPLoss, to aid learning of logical consistency across attributes, and also a label compensation training strategy to eliminate the problem of no positive prediction across a set of related attributes. Using an attribute classifier trained on FH37K, we investigate how facial hair affects face recognition accuracy, including variation across demographics. Results show that similarity and difference in facial hairstyle have important effects on the impostor and genuine score distributions in face recognition. The code is at https:// github.com/ HaiyuWu/ facial hair logical.

Modeling non-Lambertian effects such as facial specularity leads to a more realistic 3D Morphable Face Model. Existing works build parametric models for diffuse and specular albedo using Light Stage data. However, only diffuse and specular albedo cannot determine the full BRDF. In addition, the requirement of Light Stage data is hard to fulfill for the research communities. This paper proposes the first 3D morphable face reflectance model with spatially varying BRDF using only low-cost publicly-available data. We apply linear shiness weighting into parametric modeling to represent spatially varying specular intensity and shiness. Then an inverse rendering algorithm is developed to reconstruct the reflectance parameters from non-Light Stage data, which are used to train an initial morphable reflectance model. To enhance the model’s generalization capability and expressive power, we further propose an update-by-reconstruction strategy to finetune it on an in-the-wild dataset. Experimental results show that our method obtains decent rendering results with plausible facial specularities. Our code is released at https://yxuhan.github.io/ReflectanceMM/index.html.

Our paper seeks to transfer the hairstyle of a reference image to an input photo for virtual hair try-on. We target a variety of challenges scenarios, such as transforming a long hairstyle with bangs to a pixie cut, which requires removing the existing hair and inferring how the forehead would look, or transferring partially visible hair from a hat-wearing person in a different pose. Past solutions leverage StyleGAN for hallucinating any missing parts and producing a seamless face-hair composite through so-called GAN inversion or projection. However, there remains a challenge in controlling the hallucinations to accurately transfer hairstyle and preserve the face shape and identity of the input. To overcome this, we propose a multi-view optimization framework that uses “two different views” of reference composites to semantically guide occluded or ambiguous regions. Our optimization shares information between two poses, which allows us to produce high fidelity and realistic results from incomplete references. Our framework produces high-quality results and outperforms prior work in a user study that consists of significantly more challenging hair transfer scenarios than previously studied. Project page: https://stylegan-salon.github.io/.

We propose a generative framework, FaceLit, capable of generating a 3D face that can be rendered at various user-defined lighting conditions and views, learned purely from 2D images in-the-wild without any manual annotation. Unlike existing works that require careful capture setup or human labor, we rely on off-the-shelf pose and illumination estimators. With these estimates, we incorporate the Phong reflectance model in the neural volume rendering framework. Our model learns to generate shape and material properties of a face such that, when rendered according to the natural statistics of pose and illumination, produces photorealistic face images with multiview 3D and illumination consistency. Our method enables photorealistic generation of faces with explicit illumination and view controls on multiple datasets -- FFHQ, MetFaces and CelebA-HQ. We show state-of-the-art photorealism among 3D aware GANs on FFHQ dataset achieving an FID score of 3.5.

In this paper, we introduce FitMe, a facial reflectance model and a differentiable rendering optimization pipeline, that can be used to acquire high-fidelity renderable human avatars from single or multiple images. The model consists of a multi-modal style-based generator, that captures facial appearance in terms of diffuse and specular reflectance, and a PCA-based shape model. We employ a fast differentiable rendering process that can be used in an optimization pipeline, while also achieving photorealistic facial shading. Our optimization process accurately captures both the facial reflectance and shape in high-detail, by exploiting the expressivity of the style-based latent representation and of our shape model. FitMe achieves state-of-the-art reflectance acquisition and identity preservation on single “in-the-wild” facial images, while it produces impressive scan-like results, when given multiple unconstrained facial images pertaining to the same identity. In contrast with recent implicit avatar reconstructions, FitMe requires only one minute and produces relightable mesh and texture-based avatars, that can be used by end-user applications.

The capture and animation of human hair are two of the major challenges in the creation of realistic avatars for the virtual reality. Both problems are highly challenging, because hair has complex geometry and appearance, as well as exhibits challenging motion. In this paper, we present a two-stage approach that models hair independently from the head to address these challenges in a data-driven manner. The first stage, state compression, learns a low-dimensional latent space of 3D hair states containing motion and appearance, via a novel autoencoder-as-a-tracker strategy. To better disentangle the hair and head in appearance learning, we employ multi-view hair segmentation masks in combination with a differentiable volumetric renderer. The second stage learns a novel hair dynamics model that performs temporal hair transfer based on the discovered latent codes. To enforce higher stability while driving our dynamics model, we employ the 3D point-cloud autoencoder from the compression stage for de-noising of the hair state. Our model outperforms the state of the art in novel view synthesis and is capable of creating novel hair animations without having to rely on hair observations as a driving signal

Generating talking head videos through a face image and a piece of speech audio still contains many challenges. i.e., unnatural head movement, distorted expression, and identity modification. We argue that these issues are mainly caused by learning from the coupled 2D motion fields. On the other hand, explicitly using 3D information also suffers problems of stiff expression and incoherent video. We present SadTalker, which generates 3D motion coefficients (head pose, expression) of the 3DMM from audio and implicitly modulates a novel 3D-aware face render for talking head generation. To learn the realistic motion coefficients, we explicitly model the connections between audio and different types of motion coefficients individually. Precisely, we present ExpNet to learn the accurate facial expression from audio by distilling both coefficients and 3D-rendered faces. As for the head pose, we design PoseVAE via a conditional VAE to synthesize head motion in different styles. Finally, the generated 3D motion coefficients are mapped to the unsupervised 3D keypoints space of the proposed face render to synthesize the final video. We conducted extensive experiments to show the superior of our method in terms of motion and video quality.

This paper presents a framework for efficient 3D clothed avatar reconstruction. By combining the advantages of the high accuracy of optimization-based methods and the efficiency of learning-based methods, we propose a coarse-to-fine way to realize a high-fidelity clothed avatar reconstruction (CAR) from a single image. At the first stage, we use an implicit model to learn the general shape in the canonical space of a person in a learning-based way, and at the second stage, we refine the surface detail by estimating the non-rigid deformation in the posed space in an optimization way. A hyper-network is utilized to generate a good initialization so that the convergence of the optimization process is greatly accelerated. Extensive experiments on various datasets show that the proposed CAR successfully produces high-fidelity avatars for arbitrarily clothed humans in real scenes. The codes will be released.

Music-driven choreography is a challenging problem with a wide variety of industrial applications. Recently, many methods have been proposed to synthesize dance motions from music for a single dancer. However, generating dance motion for a group remains an open problem. In this paper, we present AIOZ-GDANCE, a new largescale dataset for music-driven group dance generation. Unlike existing datasets that only support single dance, our new dataset contains group dance videos, hence supporting the study of group choreography. We propose a semiautonomous labeling method with humans in the loop to obtain the 3D ground truth for our dataset. The proposed dataset consists of 16.7 hours of paired music and 3D motion from in-the-wild videos, covering 7 dance styles and 16 music genres. We show that naively applying single dance generation technique to creating group dance motion may lead to unsatisfactory results, such as inconsistent movements and collisions between dancers. Based on our new dataset, we propose a new method that takes an input music sequence and a set of 3D positions of dancers to efficiently produce multiple group-coherent choreographies. We propose new evaluation metrics for measuring group dance quality and perform intensive experiments to demonstrate the effectiveness of our method. Our project facilitates future research on group dance generation and is available at https://aioz-ai.github.io/AIOZ-GDANCE/.

We present HandAvatar, a novel representation for hand animation and rendering, which can generate smoothly compositional geometry and self-occlusion-aware texture. Specifically, we first develop a MANO-HD model as a high-resolution mesh topology to fit personalized hand shapes. Sequentially, we decompose hand geometry into per-bone rigid parts, and then re-compose paired geometry encodings to derive an across-part consistent occupancy field. As for texture modeling, we propose a self-occlusion-aware shading field (SelF). In SelF, drivable anchors are paved on the MANO-HD surface to record albedo information under a wide variety of hand poses. Moreover, directed soft occupancy is designed to describe the ray-to-surface relation, which is leveraged to generate an illumination field for the disentanglement of pose-independent albedo and pose-dependent illumination. Trained from monocular video data, our HandAvatar can perform free-pose hand animation and rendering while at the same time achieving superior appearance fidelity. We also demonstrate that HandAvatar provides a route for hand appearance editing.

Existing AU detection algorithms are mainly based on appearance information extracted from 2D images, and well-established facial biomechanics that governs 3D facial skin deformation is rarely considered. In this paper, we propose a biomechanics-guided AU detection approach, where facial muscle activation forces are modelled, and are employed to predict AU activation. Specifically, our model consists of two branches: 3D physics branch and 2D image branch. In 3D physics branch, we first derive the Euler-Lagrange equation governing facial deformation. The Euler-Lagrange equation represented as an ordinary differential equation (ODE) is embedded into a differentiable ODE solver. Muscle activation forces together with other physics parameters are firstly regressed, and then are utilized to simulate 3D deformation by solving the ODE. By leveraging facial biomechanics, we obtain physically plausible facial muscle activation forces. 2D image branch compensates 3D physics branch by employing additional appearance information from 2D images. Both estimated forces and appearance features are employed for AU detection. The proposed approach achieves competitive AU detection performance on two benchmark datasets. Furthermore, by leveraging biomechanics, our approach achieves outstanding performance with reduced training data.

Transferring the pose of a reference avatar to stylized 3D characters of various shapes is a fundamental task in computer graphics. Existing methods either require the stylized characters to be rigged, or they use the stylized character in the desired pose as ground truth at training. We present a zero-shot approach that requires only the widely available deformed non-stylized avatars in training, and deforms stylized characters of significantly different shapes at inference. Classical methods achieve strong generalization by deforming the mesh at the triangle level, but this requires labelled correspondences. We leverage the power of local deformation, but without requiring explicit correspondence labels. We introduce a semi-supervised shape-understanding module to bypass the need for explicit correspondences at test time, and an implicit pose deformation module that deforms individual surface points to match the target pose. Furthermore, to encourage realistic and accurate deformation of stylized characters, we introduce an efficient volume-based test-time training procedure. Because it does not need rigging, nor the deformed stylized character at training time, our model generalizes to categories with scarce annotation, such as stylized quadrupeds. Extensive experiments demonstrate the effectiveness of the proposed method compared to the state-of-the-art approaches trained with comparable or more supervision. Our project page is available at https://jiashunwang.github.io/ZPT

Building animatable and editable models of clothed humans from raw 3D scans and poses is a challenging problem. Existing reposing methods suffer from the limited expressiveness of Linear Blend Skinning (LBS), require costly mesh extraction to generate each new pose, and typically do not preserve surface correspondences across different poses. In this work, we introduce Invertible Neural Skinning (INS) to address these shortcomings. To maintain correspondences, we propose a Pose-conditioned Invertible Network (PIN) architecture, which extends the LBS process by learning additional pose-varying deformations. Next, we combine PIN with a differentiable LBS module to build an expressive and end-to-end Invertible Neural Skinning (INS) pipeline. We demonstrate the strong performance of our method by outperforming the state-of-the-art reposing techniques on clothed humans and preserving surface correspondences, while being an order of magnitude faster. We also perform an ablation study, which shows the usefulness of our pose-conditioning formulation, and our qualitative results display that INS can rectify artefacts introduced by LBS well.

We show, for the first time, that neural networks trained only on synthetic data achieve state-of-the-art accuracy on the problem of 3D human pose and shape (HPS) estimation from real images. Previous synthetic datasets have been small, unrealistic, or lacked realistic clothing. Achieving sufficient realism is non-trivial and we show how to do this for full bodies in motion. Specifically, our BEDLAM dataset contains monocular RGB videos with ground-truth 3D bodies in SMPL-X format. It includes a diversity of body shapes, motions, skin tones, hair, and clothing. The clothing is realistically simulated on the moving bodies using commercial clothing physics simulation. We render varying numbers of people in realistic scenes with varied lighting and camera motions. We then train various HPS regressors using BEDLAM and achieve state-of-the-art accuracy on real-image benchmarks despite training with synthetic data. We use BEDLAM to gain insights into what model design choices are important for accuracy. With good synthetic training data, we find that a basic method like HMR approaches the accuracy of the current SOTA method (CLIFF). BEDLAM is useful for a variety of tasks and all images, ground truth bodies, 3D clothing, support code, and more are available for research purposes. Additionally, we provide detailed information about our synthetic data generation pipeline, enabling others to generate their own datasets. See the project page: https://bedlam.is.tue.mpg.de/.

Recently, implicit function (IF)-based methods for clothed human reconstruction using a single image have received a lot of attention. Most existing methods rely on a 3D embedding branch using volume such as the skinned multi-person linear (SMPL) model, to compensate for the lack of information in a single image. Beyond the SMPL, which provides skinned parametric human 3D information, in this paper, we propose a new IF-based method, DIFu, that utilizes a projected depth prior containing textured and non-parametric human 3D information. In particular, DIFu consists of a generator, an occupancy prediction network, and a texture prediction network. The generator takes an RGB image of the human front-side as input, and hallucinates the human back-side image. After that, depth maps for front/back images are estimated and projected into 3D volume space. Finally, the occupancy prediction network extracts a pixel-aligned feature and a voxel-aligned feature through a 2D encoder and a 3D encoder, respectively, and estimates occupancy using these features. Note that voxel-aligned features are obtained from the projected depth maps, thus it can contain detailed 3D information such as hair and cloths. Also, colors of each 3D point are also estimated with the texture inference branch. The effectiveness of DIFu is demonstrated by comparing to recent IF-based models quantitatively and qualitatively.

This paper presents a method to reconstruct a complete human geometry and texture from an image of a person with only partial body observed, e.g., a torso. The core challenge arises from the occlusion: there exists no pixel to reconstruct where many existing single-view human reconstruction methods are not designed to handle such invisible parts, leading to missing data in 3D. To address this challenge, we introduce a novel coarse-to-fine human reconstruction framework. For coarse reconstruction, explicit volumetric features are learned to generate a complete human geometry with 3D convolutional neural networks conditioned by a 3D body model and the style features from visible parts. An implicit network combines the learned 3D features with the high-quality surface normals enhanced from multiview to produce fine local details, e.g., high-frequency wrinkles. Finally, we perform progressive texture inpainting to reconstruct a complete appearance of the person in a view-consistent way, which is not possible without the reconstruction of a complete geometry. In experiments, we demonstrate that our method can reconstruct high-quality 3D humans, which is robust to occlusion.

This paper addresses the challenge of efficiently reconstructing volumetric videos of dynamic humans from sparse multi-view videos. Some recent works represent a dynamic human as a canonical neural radiance field (NeRF) and a motion field, which are learned from input videos through differentiable rendering. But the per-scene optimization generally requires hours. Other generalizable NeRF models leverage learned prior from datasets to reduce the optimization time by only finetuning on new scenes at the cost of visual fidelity. In this paper, we propose a novel method for learning neural volumetric representations of dynamic humans in minutes with competitive visual quality. Specifically, we define a novel part-based voxelized human representation to better distribute the representational power of the network to different human parts. Furthermore, we propose a novel 2D motion parameterization scheme to increase the convergence rate of deformation field learning. Experiments demonstrate that our model can be learned 100 times faster than previous per-scene optimization methods while being competitive in the rendering quality. Training our model on a 512x512 video with 100 frames typically takes about 5 minutes on a single RTX 3090 GPU. The code is available on our project page: https://zju3dv.github.io/instant_nvr

Representing complex objects with basic geometric primitives has long been a topic in computer vision. Primitive-based representations have the merits of compactness and computational efficiency in higher-level tasks such as physics simulation, collision checking, and robotic manipulation. Unlike previous works which extract polygonal meshes from a signed distance function (SDF), in this paper, we present a novel method, named Marching-Primitives, to obtain a primitive-based abstraction directly from an SDF. Our method grows geometric primitives (such as superquadrics) iteratively by analyzing the connectivity of voxels while marching at different levels of signed distance. For each valid connected volume of interest, we march on the scope of voxels from which a primitive is able to be extracted in a probabilistic sense and simultaneously solve for the parameters of the primitive to capture the underlying local geometry. We evaluate the performance of our method on both synthetic and real-world datasets. The results show that the proposed method outperforms the state-of-the-art in terms of accuracy, and is directly generalizable among different categories and scales. The code is open-sourced at https://github.com/ChirikjianLab/Marching-Primitives.git.

Despite various probabilistic methods for modeling the uncertainty and ambiguity in human mesh recovery, their overall precision is limited because existing formulations for joint rotations are either not constrained to SO(3) or difficult to learn for neural networks. To address such an issue, we derive a novel analytical formulation for learning posterior probability distributions of human joint rotations conditioned on bone directions in a Bayesian manner, and based on this, we propose a new posterior-guided framework for human mesh recovery. We demonstrate that our framework is not only superior to existing SOTA baselines on multiple benchmarks but also flexible enough to seamlessly incorporate with additional sensors due to its Bayesian nature. The code is available at https://github.com/NetEase-GameAI/ProPose.

We consider the problem of predicting the 3D shape, articulation, viewpoint, texture, and lighting of an articulated animal like a horse given a single test image as input. We present a new method, dubbed MagicPony, that learns this predictor purely from in-the-wild single-view images of the object category, with minimal assumptions about the topology of deformation. At its core is an implicit-explicit representation of articulated shape and appearance, combining the strengths of neural fields and meshes. In order to help the model understand an object’s shape and pose, we distil the knowledge captured by an off-the-shelf self-supervised vision transformer and fuse it into the 3D model. To overcome local optima in viewpoint estimation, we further introduce a new viewpoint sampling scheme that comes at no additional training cost. MagicPony outperforms prior work on this challenging task and demonstrates excellent generalisation in reconstructing art, despite the fact that it is only trained on real images. The code can be found on the project page at https://3dmagicpony.github.io/.

Tactile sensing is one of the modalities human rely on heavily to perceive the world. Working with vision, this modality refines local geometry structure, measures deformation at contact area, and indicates hand-object contact state. With the availability of open-source tactile sensors such as DIGIT, research on visual-tactile learning is becoming more accessible and reproducible. Leveraging this tactile sensor, we propose a novel visual-tactile in-hand object reconstruction framework VTacO, and extend it to VTacOH for hand-object reconstruction. Since our method can support both rigid and deformable object reconstruction, and no existing benchmark are proper for the goal. We propose a simulation environment, VT-Sim, which supports to generate hand-object interaction for both rigid and deformable objects. With VT-Sim, we generate a large-scale training dataset, and evaluate our method on it. Extensive experiments demonstrate that our proposed method can outperform the previous baseline methods qualitatively and quantitatively. Finally, we directly apply our model trained in simulation to various real-world test cases, which display qualitative results. Codes, models, simulation environment, datasets will be publicly available.

We present Cart, a new approach towards articulated-object manipulations by human commands. Beyond the existing work that focuses on inferring articulation structures, we further support manipulating articulated shapes to align them subject to simple command templates. The key of Cart is to utilize the prediction of object structures to connect visual observations with user commands for effective manipulations. It is achieved by encoding command messages for motion prediction and a test-time adaptation to adjust the amount of movement from only command supervision. For a rich variety of object categories, Cart can accurately manipulate object shapes and outperform the state-of-the-art approaches in understanding the inherent articulation structures. Also, it can well generalize to unseen object categories and real-world objects. We hope Cart could open new directions for instructing machines to operate articulated objects.

Reactive grasping, which enables the robot to successfully grasp dynamic moving objects, is of great interest in robotics. Current methods mainly focus on the temporal smoothness of the predicted grasp poses but few consider their semantic consistency. Consequently, the predicted grasps are not guaranteed to fall on the same part of the same object, especially in cluttered scenes. In this paper, we propose to solve reactive grasping in a target-referenced setting by tracking through generated grasp spaces. Given a targeted grasp pose on an object and detected grasp poses in a new observation, our method is composed of two stages: 1) discovering grasp pose correspondences through an attentional graph neural network and selecting the one with the highest similarity with respect to the target pose; 2) refining the selected grasp poses based on target and historical information. We evaluate our method on a large-scale benchmark GraspNet-1Billion. We also collect 30 scenes of dynamic objects for testing. The results suggest that our method outperforms other representative methods. Furthermore, our real robot experiments achieve an average success rate of over 80 percent.

Humans constantly interact with objects in daily life tasks. Capturing such processes and subsequently conducting visual inferences from a fixed viewpoint suffers from occlusions, shape and texture ambiguities, motions, etc. To mitigate the problem, it is essential to build a training dataset that captures free-viewpoint interactions. We construct a dense multi-view dome to acquire a complex human object interaction dataset, named HODome, that consists of ~71M frames on 10 subjects interacting with 23 objects. To process the HODome dataset, we develop NeuralDome, a layer-wise neural processing pipeline tailored for multi-view video inputs to conduct accurate tracking, geometry reconstruction and free-view rendering, for both human subjects and objects. Extensive experiments on the HODome dataset demonstrate the effectiveness of NeuralDome on a variety of inference, modeling, and rendering tasks. Both the dataset and the NeuralDome tools will be disseminated to the community for further development, which can be found at https://juzezhang.github.io/NeuralDome

3D interacting hand pose estimation from a single RGB image is a challenging task, due to serious self-occlusion and inter-occlusion towards hands, confusing similar appearance patterns between 2 hands, ill-posed joint position mapping from 2D to 3D, etc.. To address these, we propose to extend A2J-the state-of-the-art depth-based 3D single hand pose estimation method-to RGB domain under interacting hand condition. Our key idea is to equip A2J with strong local-global aware ability to well capture interacting hands’ local fine details and global articulated clues among joints jointly. To this end, A2J is evolved under Transformer’s non-local encoding-decoding framework to build A2J-Transformer. It holds 3 main advantages over A2J. First, self-attention across local anchor points is built to make them global spatial context aware to better capture joints’ articulation clues for resisting occlusion. Secondly, each anchor point is regarded as learnable query with adaptive feature learning for facilitating pattern fitting capacity, instead of having the same local representation with the others. Last but not least, anchor point locates in 3D space instead of 2D as in A2J, to leverage 3D pose prediction. Experiments on challenging InterHand 2.6M demonstrate that, A2J-Transformer can achieve state-of-the-art model-free performance (3.38mm MPJPE advancement in 2-hand case) and can also be applied to depth domain with strong generalization.

Although the estimation of 3D human pose and shape (HPS) is rapidly progressing, current methods still cannot reliably estimate moving humans in global coordinates, which is critical for many applications. This is particularly challenging when the camera is also moving, entangling human and camera motion. To address these issues, we adopt a novel 5D representation (space, time, and identity) that enables end-to-end reasoning about people in scenes. Our method, called TRACE, introduces several novel architectural components. Most importantly, it uses two new “maps” to reason about the 3D trajectory of people over time in camera, and world, coordinates. An additional memory unit enables persistent tracking of people even during long occlusions. TRACE is the first one-stage method to jointly recover and track 3D humans in global coordinates from dynamic cameras. By training it end-to-end, and using full image information, TRACE achieves state-of-the-art performance on tracking and HPS benchmarks. The code and dataset are released for research purposes.

We address the problem of inferring the 3D shape and pose of dogs from images. Given the lack of 3D training data, this problem is challenging, and the best methods lag behind those designed to estimate human shape and pose. To make progress, we attack the problem from multiple sides at once. First, we need a good 3D shape prior, like those available for humans. To that end, we learn a dog-specific 3D parametric model, called D-SMAL. Second, existing methods focus on dogs in standing poses because when they sit or lie down, their legs are self occluded and their bodies deform. Without access to a good pose prior or 3D data, we need an alternative approach. To that end, we exploit contact with the ground as a form of side information. We consider an existing large dataset of dog images and label any 3D contact of the dog with the ground. We exploit body-ground contact in estimating dog pose and find that it significantly improves results. Third, we develop a novel neural network architecture to infer and exploit this contact information. Fourth, to make progress, we have to be able to measure it. Current evaluation metrics are based on 2D features like keypoints and silhouettes, which do not directly correlate with 3D errors. To address this, we create a synthetic dataset containing rendered images of scanned 3D dogs. With these advances, our method recovers significantly better dog shape and pose than the state of the art, and we evaluate this improvement in 3D. Our code, model and test dataset are publicly available for research purposes at https://bite.is.tue.mpg.de.

Recently, transformer-based methods have gained significant success in sequential 2D-to-3D lifting human pose estimation. As a pioneering work, PoseFormer captures spatial relations of human joints in each video frame and human dynamics across frames with cascaded transformer layers and has achieved impressive performance. However, in real scenarios, the performance of PoseFormer and its follow-ups is limited by two factors: (a) The length of the input joint sequence; (b) The quality of 2D joint detection. Existing methods typically apply self-attention to all frames of the input sequence, causing a huge computational burden when the frame number is increased to obtain advanced estimation accuracy, and they are not robust to noise naturally brought by the limited capability of 2D joint detectors. In this paper, we propose PoseFormerV2, which exploits a compact representation of lengthy skeleton sequences in the frequency domain to efficiently scale up the receptive field and boost robustness to noisy 2D joint detection. With minimum modifications to PoseFormer, the proposed method effectively fuses features both in the time domain and frequency domain, enjoying a better speed-accuracy trade-off than its precursor. Extensive experiments on two benchmark datasets (i.e., Human3.6M and MPI-INF-3DHP) demonstrate that the proposed approach significantly outperforms the original PoseFormer and other transformer-based variants. Code is released at https://github.com/QitaoZhao/PoseFormerV2.

Video-based 3D human pose and shape estimations are evaluated by intra-frame accuracy and inter-frame smoothness. Although these two metrics are responsible for different ranges of temporal consistency, existing state-of-the-art methods treat them as a unified problem and use monotonous modeling structures (e.g., RNN or attention-based block) to design their networks. However, using a single kind of modeling structure is difficult to balance the learning of short-term and long-term temporal correlations, and may bias the network to one of them, leading to undesirable predictions like global location shift, temporal inconsistency, and insufficient local details. To solve these problems, we propose to structurally decouple the modeling of long-term and short-term correlations in an end-to-end framework, Global-to-Local Transformer (GLoT). First, a global transformer is introduced with a Masked Pose and Shape Estimation strategy for long-term modeling. The strategy stimulates the global transformer to learn more inter-frame correlations by randomly masking the features of several frames. Second, a local transformer is responsible for exploiting local details on the human mesh and interacting with the global transformer by leveraging cross-attention. Moreover, a Hierarchical Spatial Correlation Regressor is further introduced to refine intra-frame estimations by decoupled global-local representation and implicit kinematic constraints. Our GLoT surpasses previous state-of-the-art methods with the lowest model parameters on popular benchmarks, i.e., 3DPW, MPI-INF-3DHP, and Human3.6M. Codes are available at https://github.com/sxl142/GLoT.

Head pose estimation (HPE) has been widely used in the fields of human machine interaction, self-driving, and attention estimation. However, existing methods cannot deal with extreme head pose randomness and serious occlusions. To address these challenges, we identify three cues from head images, namely, neighborhood similarities, significant facial changes, and critical minority relationships. To leverage the observed findings, we propose a novel critical minority relationship-aware method based on the Transformer architecture in which the facial part relationships can be learned. Specifically, we design several orientation tokens to explicitly encode the basic orientation regions. Meanwhile, a novel token guide multi-loss function is designed to guide the orientation tokens as they learn the desired regional similarities and relationships. We evaluate the proposed method on three challenging benchmark HPE datasets. Experiments show that our method achieves better performance compared with state-of-the-art methods. Our code is publicly available at https://github.com/zc2023/TokenHPE.

Gaze-following is a kind of research that requires locating where the person in the scene is looking automatically under the topic of gaze estimation. It is an important clue for understanding human intention, such as identifying objects or regions of interest to humans. However, a survey of datasets used for gaze-following tasks reveals defects in the way they collect gaze point labels. Manual labeling may introduce subjective bias and is labor-intensive, while automatic labeling with an eye-tracking device would alter the person’s appearance. In this work, we introduce GFIE, a novel dataset recorded by a gaze data collection system we developed. The system is constructed with two devices, an Azure Kinect and a laser rangefinder, which generate the laser spot to steer the subject’s attention as they perform in front of the camera. And an algorithm is developed to locate laser spots in images for annotating 2D/3D gaze targets and removing ground truth introduced by the spots. The whole procedure of collecting gaze behavior allows us to obtain unbiased labels in unconstrained environments semi-automatically. We also propose a baseline method with stereo field-of-view (FoV) perception for establishing a 2D/3D gaze-following benchmark on the GFIE dataset. Project page: https://sites.google.com/view/gfie.

In this work, we tackle the problem of online camera-to-robot pose estimation from single-view successive frames of an image sequence, a crucial task for robots to interact with the world. The primary obstacles of this task are the robot’s self-occlusions and the ambiguity of single-view images. This work demonstrates, for the first time, the effectiveness of temporal information and the robot structure prior in addressing these challenges. Given the successive frames and the robot joint configuration, our method learns to accurately regress the 2D coordinates of the predefined robot’s keypoints (e.g., joints). With the camera intrinsic and robotic joints status known, we get the camera-to-robot pose using a Perspective-n-point (PnP) solver. We further improve the camera-to-robot pose iteratively using the robot structure prior. To train the whole pipeline, we build a large-scale synthetic dataset generated with domain randomisation to bridge the sim-to-real gap. The extensive experiments on synthetic and real-world datasets and the downstream robotic grasping task demonstrate that our method achieves new state-of-the-art performances and outperforms traditional hand-eye calibration algorithms in real-time (36 FPS). Code and data are available at the project page: https://sites.google.com/view/sgtapose.

Most recent 6D object pose estimation methods first use object detection to obtain 2D bounding boxes before actually regressing the pose. However, the general object detection methods they use are ill-suited to handle cluttered scenes, thus producing poor initialization to the subsequent pose network. To address this, we propose a rigidity-aware detection method exploiting the fact that, in 6D pose estimation, the target objects are rigid. This lets us introduce an approach to sampling positive object regions from the entire visible object area during training, instead of naively drawing samples from the bounding box center where the object might be occluded. As such, every visible object part can contribute to the final bounding box prediction, yielding better detection robustness. Key to the success of our approach is a visibility map, which we propose to build using a minimum barrier distance between every pixel in the bounding box and the box boundary. Our results on seven challenging 6D pose estimation datasets evidence that our method outperforms general detection frameworks by a large margin. Furthermore, combined with a pose regression network, we obtain state-of-the-art pose estimation results on the challenging BOP benchmark.

Image-based multi-person reconstruction in wide-field large scenes is critical for crowd analysis and security alert. However, existing methods cannot deal with large scenes containing hundreds of people, which encounter the challenges of large number of people, large variations in human scale, and complex spatial distribution. In this paper, we propose Crowd3D, the first framework to reconstruct the 3D poses, shapes and locations of hundreds of people with global consistency from a single large-scene image. The core of our approach is to convert the problem of complex crowd localization into pixel localization with the help of our newly defined concept, Human-scene Virtual Interaction Point (HVIP). To reconstruct the crowd with global consistency, we propose a progressive reconstruction network based on HVIP by pre-estimating a scene-level camera and a ground plane. To deal with a large number of persons and various human sizes, we also design an adaptive human-centric cropping scheme. Besides, we contribute a benchmark dataset, LargeCrowd, for crowd reconstruction in a large scene. Experimental results demonstrate the effectiveness of the proposed method. The code and the dataset are available at http://cic.tju.edu.cn/faculty/likun/projects/Crowd3D.

The two-stage object pose estimation paradigm first detects semantic keypoints on the image and then estimates the 6D pose by minimizing reprojection errors. Despite performing well on standard benchmarks, existing techniques offer no provable guarantees on the quality and uncertainty of the estimation. In this paper, we inject two fundamental changes, namely conformal keypoint detection and geometric uncertainty propagation, into the two-stage paradigm and propose the first pose estimator that endows an estimation with provable and computable worst-case error bounds. On one hand, conformal keypoint detection applies the statistical machinery of inductive conformal prediction to convert heuristic keypoint detections into circular or elliptical prediction sets that cover the groundtruth keypoints with a user-specified marginal probability (e.g., 90%). Geometric uncertainty propagation, on the other, propagates the geometric constraints on the keypoints to the 6D object pose, leading to a Pose UnceRtainty SEt (PURSE) that guarantees coverage of the groundtruth pose with the same probability. The PURSE, however, is a nonconvex set that does not directly lead to estimated poses and uncertainties. Therefore, we develop RANdom SAmple averaGing (RANSAG) to compute an average pose and apply semidefinite relaxation to upper bound the worst-case errors between the average pose and the groundtruth. On the LineMOD Occlusion dataset we demonstrate: (i) the PURSE covers the groundtruth with valid probabilities; (ii) the worst-case error bounds provide correct uncertainty quantification; and (iii) the average pose achieves better or similar accuracy as representative methods based on sparse keypoints.

Bundle adjustment is a key component in practically all available Structure from Motion systems. While it is crucial for achieving accurate reconstruction, convergence to the right solution hinges on good initialization. The recently introduced factorization-based pOSE methods formulate a surrogate for the bundle adjustment error without reliance on good initialization. In this paper, we show that pOSE has an undesirable penalization of large depths. To address this we propose expOSE which has an exponential regularization that is negligible for positive depths. To achieve efficient inference we use a quadratic approximation that allows an iterative solution with VarPro. Furthermore, we extend the method with radial distortion robustness by decomposing the Object Space Error into radial and tangential components. Experimental results confirm that the proposed method is robust to initialization and improves reconstruction quality compared to state-of-the-art methods even without bundle adjustment refinement.

We present a unified framework for camera-space 3D hand pose estimation from a single RGB image based on 3D implicit representation. As opposed to recent works, most of which first adopt holistic or pixel-level dense regression to obtain relative 3D hand pose and then follow with complex second-stage operations for 3D global root or scale recovery, we propose a novel unified 3D dense regression scheme to estimate camera-space 3D hand pose via dense 3D point-wise voting in camera frustum. Through direct dense modeling in 3D domain inspired by Pixel-aligned Implicit Functions for 3D detailed reconstruction, our proposed Neural Voting Field (NVF) fully models 3D dense local evidence and hand global geometry, helping to alleviate common 2D-to-3D ambiguities. Specifically, for a 3D query point in camera frustum and its pixel-aligned image feature, NVF, represented by a Multi-Layer Perceptron, regresses: (i) its signed distance to the hand surface; (ii) a set of 4D offset vectors (1D voting weight and 3D directional vector to each hand joint). Following a vote-casting scheme, 4D offset vectors from near-surface points are selected to calculate the 3D hand joint coordinates by a weighted average. Experiments demonstrate that NVF outperforms existing state-of-the-art algorithms on FreiHAND dataset for camera-space 3D hand pose estimation. We also adapt NVF to the classic task of root-relative 3D hand pose estimation, for which NVF also obtains state-of-the-art results on HO3D dataset.

Camera pose estimation for two-view geometry traditionally relies on RANSAC. Normally, a multitude of image correspondences leads to a pool of proposed hypotheses, which are then scored to find a winning model. The inlier count is generally regarded as a reliable indicator of “consensus”. We examine this scoring heuristic, and find that it favors disappointing models under certain circumstances. As a remedy, we propose the Fundamental Scoring Network (FSNet), which infers a score for a pair of overlapping images and any proposed fundamental matrix. It does not rely on sparse correspondences, but rather embodies a two-view geometry model through an epipolar attention mechanism that predicts the pose error of the two images. FSNet can be incorporated into traditional RANSAC loops. We evaluate FSNet on fundamental and essential matrix estimation on indoor and outdoor datasets, and establish that FSNet can successfully identify good poses for pairs of images with few or unreliable correspondences. Besides, we show that naively combining FSNet with MAGSAC++ scoring approach achieves state of the art results.

We present novel solutions to previously unsolved problems of relative pose estimation from images whose calibration parameters, namely focal lengths and radial distortion, are unknown. Our approach enables metric reconstruction without modeling these parameters. The minimal case for reconstruction requires 13 points in 4 views for both the calibrated and uncalibrated cameras. We describe and implement the first solution to these minimal problems. In the calibrated case, this may be modeled as a polynomial system of equations with 3584 solutions. Despite the apparent intractability, the problem decomposes spectacularly. Each solution falls into a Euclidean symmetry class of size 16, and we can estimate 224 class representatives by solving a sequence of three subproblems with 28, 2, and 4 solutions. We highlight the relationship between internal constraints on the radial quadrifocal tensor and the relations among the principal minors of a 4×4 matrix. We also address the case of 4 upright cameras, where 7 points are minimal. Finally, we evaluate our approach on simulated and real data and benchmark against previous calibration-free solutions, and show that our method provides an efficient startup for an SfM pipeline with radial cameras.

Quantifying motion in 3D is important for studying the behavior of humans and other animals, but manual pose annotations are expensive and time-consuming to obtain. Self-supervised keypoint discovery is a promising strategy for estimating 3D poses without annotations. However, current keypoint discovery approaches commonly process single 2D views and do not operate in the 3D space. We propose a new method to perform self-supervised keypoint discovery in 3D from multi-view videos of behaving agents, without any keypoint or bounding box supervision in 2D or 3D. Our method, BKinD-3D, uses an encoder-decoder architecture with a 3D volumetric heatmap, trained to reconstruct spatiotemporal differences across multiple views, in addition to joint length constraints on a learned 3D skeleton of the subject. In this way, we discover keypoints without requiring manual supervision in videos of humans and rats, demonstrating the potential of 3D keypoint discovery for studying behavior.

3D traffic scene comprises various 3D information about car objects, including their pose and shape. However, most recent studies pay relatively less attention to reconstructing detailed shapes. Furthermore, most of them treat each 3D object as an independent one, resulting in losses of relative context inter-objects and scene context reflecting road circumstances. A novel monocular 3D pose and shape reconstruction algorithm, based on bi-contextual attention and attention-guided modeling (BAAM), is proposed in this work. First, given 2D primitives, we reconstruct 3D object shape based on attention-guided modeling that considers the relevance between detected objects and vehicle shape priors. Next, we estimate 3D object pose through bi-contextual attention, which leverages relation-context inter objects and scene-context between an object and road environment. Finally, we propose a 3D non maximum suppression algorithm to eliminate spurious objects based on their Bird-Eye-View distance. Extensive experiments demonstrate that the proposed BAAM yields state-of-the-art performance on ApolloCar3D. Also, they show that the proposed BAAM can be plugged into any mature monocular 3D object detector on KITTI and significantly boost their performance.

Learned visual dynamics models have proven effective for robotic manipulation tasks. Yet, it remains unclear how best to represent scenes involving multi-object interactions. Current methods decompose a scene into discrete objects, yet they struggle with precise modeling and manipulation amid challenging lighting conditions since they only encode appearance tied with specific illuminations. In this work, we propose using object-centric neural scattering functions (OSFs) as object representations in a model-predictive control framework. OSFs model per-object light transport, enabling compositional scene re-rendering under object rearrangement and varying lighting conditions. By combining this approach with inverse parameter estimation and graph-based neural dynamics models, we demonstrate improved model-predictive control performance and generalization in compositional multi-object environments, even in previously unseen scenarios and harsh lighting conditions.

3D scene understanding has seen significant advances in recent years, but has largely focused on object understanding in 3D scenes with independent per-object predictions. We thus propose to learn Neural Part Priors (NPPs), parametric spaces of objects and their parts, that enable optimizing to fit to a new input 3D scan geometry with global scene consistency constraints. The rich structure of our NPPs enables accurate, holistic scene reconstruction across similar objects in the scene. Both objects and their part geometries are characterized by coordinate field MLPs, facilitating optimization at test time to fit to input geometric observations as well as similar objects in the input scan. This enables more accurate reconstructions than independent per-object predictions as a single forward pass, while establishing global consistency within a scene. Experiments on the ScanNet dataset demonstrate that NPPs significantly outperforms the state-of-the-art in part decomposition and object completion in real-world scenes.

We propose Panoptic Lifting, a novel approach for learning panoptic 3D volumetric representations from images of in-the-wild scenes. Once trained, our model can render color images together with 3D-consistent panoptic segmentation from novel viewpoints. Unlike existing approaches which use 3D input directly or indirectly, our method requires only machine-generated 2D panoptic segmentation masks inferred from a pre-trained network. Our core contribution is a panoptic lifting scheme based on a neural field representation that generates a unified and multi-view consistent, 3D panoptic representation of the scene. To account for inconsistencies of 2D instance identifiers across views, we solve a linear assignment with a cost based on the model’s current predictions and the machine-generated segmentation masks, thus enabling us to lift 2D instances to 3D in a consistent way. We further propose and ablate contributions that make our method more robust to noisy, machine-generated labels, including test-time augmentations for confidence estimates, segment consistency loss, bounded segmentation fields, and gradient stopping. Experimental results validate our approach on the challenging Hypersim, Replica, and ScanNet datasets, improving by 8.4, 13.8, and 10.6% in scene-level PQ over state of the art.

For augmented reality (AR), it is important that virtual assets appear to ‘sit among’ real world objects. The virtual element should variously occlude and be occluded by real matter, based on a plausible depth ordering. This occlusion should be consistent over time as the viewer’s camera moves. Unfortunately, small mistakes in the estimated scene depth can ruin the downstream occlusion mask, and thereby the AR illusion. Especially in real-time settings, depths inferred near boundaries or across time can be inconsistent. In this paper, we challenge the need for depth-regression as an intermediate step. We instead propose an implicit model for depth and use that to predict the occlusion mask directly. The inputs to our network are one or more color images, plus the known depths of any virtual geometry. We show how our occlusion predictions are more accurate and more temporally stable than predictions derived from traditional depth-estimation models. We obtain state-of-the-art occlusion results on the challenging ScanNetv2 dataset and superior qualitative results on real scenes.

A central goal of visual recognition is to understand objects and scenes from a single image. 2D recognition has witnessed tremendous progress thanks to large-scale learning and general-purpose representations. But, 3D poses new challenges stemming from occlusions not depicted in the image. Prior works try to overcome these by inferring from multiple views or rely on scarce CAD models and category-specific priors which hinder scaling to novel settings. In this work, we explore single-view 3D reconstruction by learning generalizable representations inspired by advances in self-supervised learning. We introduce a simple framework that operates on 3D points of single objects or whole scenes coupled with category-agnostic large-scale training from diverse RGB-D videos. Our model, Multiview Compressive Coding (MCC), learns to compress the input appearance and geometry to predict the 3D structure by querying a 3D-aware decoder. MCC’s generality and efficiency allow it to learn from large-scale and diverse data sources with strong generalization to novel objects imagined by DALL·E 2 or captured in-the-wild with an iPhone.

Inferring a meaningful geometric scene representation from a single image is a fundamental problem in computer vision. Approaches based on traditional depth map prediction can only reason about areas that are visible in the image. Currently, neural radiance fields (NeRFs) can capture true 3D including color, but are too complex to be generated from a single image. As an alternative, we propose to predict an implicit density field from a single image. It maps every location in the frustum of the image to volumetric density. By directly sampling color from the available views instead of storing color in the density field, our scene representation becomes significantly less complex compared to NeRFs, and a neural network can predict it in a single forward pass. The network is trained through self-supervision from only video data. Our formulation allows volume rendering to perform both depth prediction and novel view synthesis. Through experiments, we show that our method is able to predict meaningful geometry for regions that are occluded in the input image. Additionally, we demonstrate the potential of our approach on three datasets for depth prediction and novel-view synthesis.

Humans can easily imagine the complete 3D geometry of occluded objects and scenes. This appealing ability is vital for recognition and understanding. To enable such capability in AI systems, we propose VoxFormer, a Transformer-based semantic scene completion framework that can output complete 3D volumetric semantics from only 2D images. Our framework adopts a two-stage design where we start from a sparse set of visible and occupied voxel queries from depth estimation, followed by a densification stage that generates dense 3D voxels from the sparse ones. A key idea of this design is that the visual features on 2D images correspond only to the visible scene structures rather than the occluded or empty spaces. Therefore, starting with the featurization and prediction of the visible structures is more reliable. Once we obtain the set of sparse queries, we apply a masked autoencoder design to propagate the information to all the voxels by self-attention. Experiments on SemanticKITTI show that VoxFormer outperforms the state of the art with a relative improvement of 20.0% in geometry and 18.1% in semantics and reduces GPU memory during training to less than 16GB. Our code is available on https://github.com/NVlabs/VoxFormer.

We propose a novel type of map for visual navigation, a renderable neural radiance map (RNR-Map), which is designed to contain the overall visual information of a 3D environment. The RNR-Map has a grid form and consists of latent codes at each pixel. These latent codes are embedded from image observations, and can be converted to the neural radiance field which enables image rendering given a camera pose. The recorded latent codes implicitly contain visual information about the environment, which makes the RNR-Map visually descriptive. This visual information in RNR-Map can be a useful guideline for visual localization and navigation. We develop localization and navigation frameworks that can effectively utilize the RNR-Map. We evaluate the proposed frameworks on camera tracking, visual localization, and image-goal navigation. Experimental results show that the RNR-Map-based localization framework can find the target location based on a single query image with fast speed and competitive accuracy compared to other baselines. Also, this localization framework is robust to environmental changes, and even finds the most visually similar places when a query image from a different environment is given. The proposed navigation framework outperforms the existing image-goal navigation methods in difficult scenarios, under odometry and actuation noises. The navigation framework shows 65.7% success rate in curved scenarios of the NRNS dataset, which is an improvement of 18.6% over the current state-of-the-art. Project page: https://rllab-snu.github.io/projects/RNR-Map/

Detecting mirrors from static images has received significant research interest recently. However, detecting mirrors over dynamic scenes is still under-explored due to the lack of a high-quality dataset and an effective method for video mirror detection (VMD). To the best of our knowledge, this is the first work to address the VMD problem from a deep-learning-based perspective. Our observation is that there are often correspondences between the contents inside (reflected) and outside (real) of a mirror, but such correspondences may not always appear in every frame, e.g., due to the change of camera pose. This inspires us to propose a video mirror detection method, named VMD-Net, that can tolerate spatially missing correspondences by considering the mirror correspondences at both the intra-frame level as well as inter-frame level via a dual correspondence module that looks over multiple frames spatially and temporally for correlating correspondences. We further propose a first large-scale dataset for VMD (named VMD-D), which contains 14,987 image frames from 269 videos with corresponding manually annotated masks. Experimental results show that the proposed method outperforms SOTA methods from relevant fields. To enable real-time VMD, our method efficiently utilizes the backbone features by removing the redundant multi-level module design and gets rid of post-processing of the output maps commonly used in existing methods, making it very efficient and practical for real-time video-based applications. Code, dataset, and models are available at https://jiaying.link/cvpr2023-vmd/

Depth estimation is an important step in many computer vision problems such as 3D reconstruction, novel view synthesis, and computational photography. Most existing work focuses on depth estimation from single frames. When applied to videos, the result lacks temporal consistency, showing flickering and swimming artifacts. In this paper we aim to estimate temporally consistent depth maps of video streams in an online setting. This is a difficult problem as future frames are not available and the method must choose between enforcing consistency and correcting errors from previous estimations. The presence of dynamic objects further complicates the problem. We propose to address these challenges by using a global point cloud that is dynamically updated each frame, along with a learned fusion approach in image space. Our approach encourages consistency while simultaneously allowing updates to handle errors and dynamic objects. Qualitative and quantitative results show that our method achieves state-of-the-art quality for consistent video depth estimation.

The asymmetric dual-lens configuration is commonly available on mobile devices nowadays, which naturally stores a pair of wide-angle and telephoto images of the same scene to support realistic super-resolution (SR). Even on the same device, however, the degradation for modeling realistic SR is image-specific due to the unknown acquisition process (e.g., tiny camera motion). In this paper, we propose a zero-shot solution for dual-lens SR (ZeDuSR), where only the dual-lens pair at test time is used to learn an image-specific SR model. As such, ZeDuSR adapts itself to the current scene without using external training data, and thus gets rid of generalization difficulty. However, there are two major challenges to achieving this goal: 1) dual-lens alignment while keeping the realistic degradation, and 2) effective usage of highly limited training data. To overcome these two challenges, we propose a degradation-invariant alignment method and a degradation-aware training strategy to fully exploit the information within a single dual-lens pair. Extensive experiments validate the superiority of ZeDuSR over existing solutions on both synthesized and real-world dual-lens datasets.

Depth-from-defocus (DFD), modeling the relationship between depth and defocus pattern in images, has demonstrated promising performance in depth estimation. Recently, several self-supervised works try to overcome the difficulties in acquiring accurate depth ground-truth. However, they depend on the all-in-focus (AIF) images, which cannot be captured in real-world scenarios. Such limitation discourages the applications of DFD methods. To tackle this issue, we propose a completely self-supervised framework that estimates depth purely from a sparse focal stack. We show that our framework circumvents the needs for the depth and AIF image ground-truth, and receives superior predictions, thus closing the gap between the theoretical success of DFD works and their applications in the real world. In particular, we propose (i) a more realistic setting for DFD tasks, where no depth or AIF image ground-truth is available; (ii) a novel self-supervision framework that provides reliable predictions of depth and AIF image under the the challenging setting. The proposed framework uses a neural model to predict the depth and AIF image, and utilizes an optical model to validate and refine the prediction. We verify our framework on three benchmark datasets with rendered focal stacks and real focal stacks. Qualitative and quantitative evaluations show that our method provides a strong baseline for self-supervised DFD tasks. The source code is publicly available at https://github.com/Ehzoahis/DEReD.

Being data-driven is one of the most iconic properties of deep learning algorithms. The birth of ImageNet drives a remarkable trend of “learning from large-scale data” in computer vision. Pretraining on ImageNet to obtain rich universal representations has been manifested to benefit various 2D visual tasks, and becomes a standard in 2D vision. However, due to the laborious collection of real-world 3D data, there is yet no generic dataset serving as a counterpart of ImageNet in 3D vision, thus how such a dataset can impact the 3D community is unraveled. To remedy this defect, we introduce MVImgNet, a large-scale dataset of multi-view images, which is highly convenient to gain by shooting videos of real-world objects in human daily life. It contains 6.5 million frames from 219,188 videos crossing objects from 238 classes, with rich annotations of object masks, camera parameters, and point clouds. The multi-view attribute endows our dataset with 3D-aware signals, making it a soft bridge between 2D and 3D vision. We conduct pilot studies for probing the potential of MVImgNet on a variety of 3D and 2D visual tasks, including radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising performance, remaining lots of possibilities for future explorations. Besides, via dense reconstruction on MVImgNet, a 3D object point cloud dataset is derived, called MVPNet, covering 87,200 samples from 150 categories, with the class label on each point cloud. Experiments show that MVPNet can benefit the real-world 3D object classification while posing new challenges to point cloud understanding. MVImgNet and MVPNet will be publicly available, hoping to inspire the broader vision community.

Current stack-based inverse tone mapping (ITM) methods can recover high dynamic range (HDR) radiance by predicting a set of multi-exposure images from a single low dynamic range image. However, there are still some limitations. On the one hand, these methods estimate a fixed number of images (e.g., three exposure-up and three exposure-down), which may introduce unnecessary computational cost or reconstruct incorrect results. On the other hand, they neglect the connections between the up-exposure and down-exposure models and thus fail to fully excavate effective features. In this paper, we revisit the stack-based ITM approaches and propose a novel method to reconstruct HDR radiance from a single image, which only needs to estimate two exposure images. At first, we design the exposure adaptive block that can adaptively adjust the exposure based on the luminance distribution of the input image. Secondly, we devise the cross-model attention block to connect the exposure adjustment models. Thirdly, we propose an end-to-end ITM pipeline by incorporating the multi-exposure fusion model. Furthermore, we propose and open a multi-exposure dataset that indicates the optimal exposure-up/down levels. Experimental results show that the proposed method outperforms some state-of-the-art methods.

Since light field simultaneously records spatial information and angular information of light rays, it is considered to be beneficial for many potential applications, and semantic segmentation is one of them. The regular variation of image information across views facilitates a comprehensive scene understanding. However, in the case of limited memory, the high-dimensional property of light field makes the problem more intractable than generic semantic segmentation, manifested in the difficulty of fully exploiting the relationships among views while maintaining contextual information in single view. In this paper, we propose a novel network called LF-IENet for light field semantic segmentation. It contains two different manners to mine complementary information from surrounding views to segment central view. One is implicit feature integration that leverages attention mechanism to compute inter-view and intra-view similarity to modulate features of central view. The other is explicit feature propagation that directly warps features of other views to central view under the guidance of disparity. They complement each other and jointly realize complementary information fusion across views in light field. The proposed method achieves outperforming performance on both real-world and synthetic light field datasets, demonstrating the effectiveness of this new architecture.

In-depth understanding of a 3D scene not only involves locating/recognizing individual objects, but also requires to infer the relationships and interactions among them. However, since 3D scenes contain partially scanned objects with physical connections, dense placement, changing sizes, and a wide variety of challenging relationships, existing methods perform quite poorly with limited training samples. In this work, we find that the inherently hierarchical structures of physical space in 3D scenes aid in the automatic association of semantic and spatial arrangements, specifying clear patterns and leading to less ambiguous predictions. Thus, they well meet the challenges due to the rich variations within scene categories. To achieve this, we explicitly unify these structural cues of 3D physical spaces into deep neural networks to facilitate scene graph prediction. Specifically, we exploit an external knowledge base as a baseline to accumulate both contextualized visual content and textual facts to form a 3D spatial multimodal knowledge graph. Moreover, we propose a knowledge-enabled scene graph prediction module benefiting from the 3D spatial knowledge to effectively regularize semantic space of relationships. Extensive experiments demonstrate the superiority of the proposed method over current state-of-the-art competitors. Our code is available at https://github.com/HHrEtvP/SMKA.

The goal of non-line-of-sight (NLOS) imaging is to image objects occluded from the camera’s field of view using multiply scattered light. Recent works have demonstrated the feasibility of two-bounce (2B) NLOS imaging by scanning a laser and measuring cast shadows of occluded objects in scenes with two relay surfaces. In this work, we study the role of time-of-flight (ToF) measurements, i.e. transients, in 2B-NLOS under multiplexed illumination. Specifically, we study how ToF information can reduce the number of measurements and spatial resolution needed for shape reconstruction. We present our findings with respect to tradeoffs in (1) temporal resolution, (2) spatial resolution, and (3) number of image captures by studying SNR and recoverability as functions of system parameters. This leads to a formal definition of the mathematical constraints for 2B lidar. We believe that our work lays an analytical groundwork for design of future NLOS imaging systems, especially as ToF sensors become increasingly ubiquitous.

Humans are able to accurately reason in 3D by gathering multi-view observations of the surrounding world. Inspired by this insight, we introduce a new large-scale benchmark for 3D multi-view visual question answering (3DMV-VQA). This dataset is collected by an embodied agent actively moving and capturing RGB images in an environment using the Habitat simulator. In total, it consists of approximately 5k scenes, 600k images, paired with 50k questions. We evaluate various state-of-the-art models for visual reasoning on our benchmark and find that they all perform poorly. We suggest that a principled approach for 3D reasoning from multi-view images should be to infer a compact 3D representation of the world from the multi-view images, which is further grounded on open-vocabulary semantic concepts, and then to execute reasoning on these 3D representations. As the first step towards this approach, we propose a novel 3D concept learning and reasoning (3D-CLR) framework that seamlessly combines these components via neural fields, 2D pre-trained vision-language models, and neural reasoning operators. Experimental results suggest that our framework outperforms baseline models by a large margin, but the challenge remains largely unsolved. We further perform an in-depth analysis of the challenges and highlight potential future directions.

3D object detection from visual sensors is a cornerstone capability of robotic systems. State-of-the-art methods focus on reasoning and decoding object bounding boxes from multi-view camera input. In this work we gain intuition from the integral role of multi-view consistency in 3D scene understanding and geometric learning. To this end, we introduce VEDet, a novel 3D object detection framework that exploits 3D multi-view geometry to improve localization through viewpoint awareness and equivariance. VEDet leverages a query-based transformer architecture and encodes the 3D scene by augmenting image features with positional encodings from their 3D perspective geometry. We design view-conditioned queries at the output level, which enables the generation of multiple virtual frames during training to learn viewpoint equivariance by enforcing multi-view consistency. The multi-view geometry injected at the input level as positional encodings and regularized at the loss level provides rich geometric cues for 3D object detection, leading to state-of-the-art performance on the nuScenes benchmark. The code and model are made available at https://github.com/TRI-ML/VEDet.

Modern methods for vision-centric autonomous driving perception widely adopt the bird’s-eye-view (BEV) representation to describe a 3D scene. Despite its better efficiency than voxel representation, it has difficulty describing the fine-grained 3D structure of a scene with a single plane. To address this, we propose a tri-perspective view (TPV) representation which accompanies BEV with two additional perpendicular planes. We model each point in the 3D space by summing its projected features on the three planes. To lift image features to the 3D TPV space, we further propose a transformer-based TPV encoder (TPVFormer) to obtain the TPV features effectively. We employ the attention mechanism to aggregate the image features corresponding to each query in each TPV plane. Experiments show that our model trained with sparse supervision effectively predicts the semantic occupancy for all voxels. We demonstrate for the first time that using only camera inputs can achieve comparable performance with LiDAR-based methods on the LiDAR segmentation task on nuScenes. Code: https://github.com/wzzheng/TPVFormer.

Depth completion plays a crucial role in autonomous driving, in which cameras and LiDARs are two complementary sensors. Recent approaches attempt to exploit spatial geometric constraints hidden in LiDARs to enhance image-guided depth completion. However, only low efficiency and poor generalization can be achieved. In this paper, we propose BEV@DC, a more efficient and powerful multi-modal training scheme, to boost the performance of image-guided depth completion. In practice, the proposed BEV@DC model comprehensively takes advantage of LiDARs with rich geometric details in training, employing an enhanced depth completion manner in inference, which takes only images (RGB and depth) as input. Specifically, the geometric-aware LiDAR features are projected onto a unified BEV space, combining with RGB features to perform BEV completion. By equipping a newly proposed point-voxel spatial propagation network (PV-SPN), this auxiliary branch introduces strong guidance to the original image branches via 3D dense supervision and feature consistency. As a result, our baseline model demonstrates significant improvements with the sole image inputs. Concretely, it achieves state-of-the-art on several benchmarks, e.g., ranking Top-1 on the challenging KITTI depth completion benchmark.

Camera-only 3D detection provides an economical solution with a simple configuration for localizing objects in 3D space compared to LiDAR-based detection systems. However, a major challenge lies in precise depth estimation due to the lack of direct 3D measurements in the input. Many previous methods attempt to improve depth estimation through network designs, e.g., deformable layers and larger receptive fields. This work proposes an orthogonal direction, improving the camera-only 3D detection by introducing multi-agent collaborations. Our proposed collaborative camera-only 3D detection (CoCa3D) enables agents to share complementary information with each other through communication. Meanwhile, we optimize communication efficiency by selecting the most informative cues. The shared messages from multiple viewpoints disambiguate the single-agent estimated depth and complement the occluded and long-range regions in the single-agent view. We evaluate CoCa3D in one real-world dataset and two new simulation datasets. Results show that CoCa3D improves previous SOTA performances by 44.21% on DAIR-V2X, 30.60% on OPV2V+, 12.59% on CoPerception-UAVs+ for AP@70. Our preliminary results show a potential that with sufficient collaboration, the camera might overtake LiDAR in some practical scenarios. We released the dataset and code at https://siheng-chen.github.io/dataset/CoPerception+ and https://github.com/MediaBrain-SJTU/CoCa3D.

Current 3D object detection models follow a single dataset-specific training and testing paradigm, which often faces a serious detection accuracy drop when they are directly deployed in another dataset. In this paper, we study the task of training a unified 3D detector from multiple datasets. We observe that this appears to be a challenging task, which is mainly due to that these datasets present substantial data-level differences and taxonomy-level variations caused by different LiDAR types and data acquisition standards. Inspired by such observation, we present a Uni3D which leverages a simple data-level correction operation and a designed semantic-level coupling-and-recoupling module to alleviate the unavoidable data-level and taxonomy-level differences, respectively. Our method is simple and easily combined with many 3D object detection baselines such as PV-RCNN and Voxel-RCNN, enabling them to effectively learn from multiple off-the-shelf 3D datasets to obtain more discriminative and generalizable representations. Experiments are conducted on many dataset consolidation settings. Their results demonstrate that Uni3D exceeds a series of individual detectors trained on a single dataset, with a 1.04× parameter increase over a selected baseline detector. We expect this work will inspire the research of 3D generalization since it will push the limits of perceptual performance. Our code is available at: https://github.com/PJLab-ADG/3DTrans

The current approach for testing the robustness of object detectors suffers from serious deficiencies such as improper methods of performing out-of-distribution detection and using calibration metrics which do not consider both localisation and classification quality. In this work, we address these issues, and introduce the Self Aware Object Detection (SAOD) task, a unified testing framework which respects and adheres to the challenges that object detectors face in safety-critical environments such as autonomous driving. Specifically, the SAOD task requires an object detector to be: robust to domain shift; obtain reliable uncertainty estimates for the entire scene; and provide calibrated confidence scores for the detections. We extensively use our framework, which introduces novel metrics and large scale test datasets, to test numerous object detectors in two different use-cases, allowing us to highlight critical insights into their robustness performance. Finally, we introduce a simple baseline for the SAOD task, enabling researchers to benchmark future proposed methods and move towards robust object detectors which are fit for purpose. Code is available at: https://github.com/fiveai/saod

We present a method for inferring dense depth from a camera image and a sparse noisy radar point cloud. We first describe the mechanics behind mmWave radar point cloud formation and the challenges that it poses, i.e. ambiguous elevation and noisy depth and azimuth components that yields incorrect positions when projected onto the image, and how existing works have overlooked these nuances in camera-radar fusion. Our approach is motivated by these mechanics, leading to the design of a network that maps each radar point to the possible surfaces that it may project onto in the image plane. Unlike existing works, we do not process the raw radar point cloud as an erroneous depth map, but query each raw point independently to associate it with likely pixels in the image -- yielding a semi-dense radar depth map. To fuse radar depth with an image, we propose a gated fusion scheme that accounts for the confidence scores of the correspondence so that we selectively combine radar and camera embeddings to yield a dense depth map. We test our method on the NuScenes benchmark and show a 10.3% improvement in mean absolute error and a 9.1% improvement in root-mean-square error over the best method.

LiDAR-based absolute pose regression estimates the global pose through a deep network in an end-to-end manner, achieving impressive results in learning-based localization. However, the accuracy of existing methods still has room to improve due to the difficulty of effectively encoding the scene geometry and the unsatisfactory quality of the data. In this work, we propose a novel LiDAR localization framework, SGLoc, which decouples the pose estimation to point cloud correspondence regression and pose estimation via this correspondence. This decoupling effectively encodes the scene geometry because the decoupled correspondence regression step greatly preserves the scene geometry, leading to significant performance improvement. Apart from this decoupling, we also design a tri-scale spatial feature aggregation module and inter-geometric consistency constraint loss to effectively capture scene geometry. Moreover, we empirically find that the ground truth might be noisy due to GPS/INS measuring errors, greatly reducing the pose estimation performance. Thus, we propose a pose quality evaluation and enhancement method to measure and correct the ground truth pose. Extensive experiments on the Oxford Radar RobotCar and NCLT datasets demonstrate the effectiveness of SGLoc, which outperforms state-of-the-art regression-based localization methods by 68.5% and 67.6% on position accuracy, respectively.

Although DETR-based 3D detectors simplify the detection pipeline and achieve direct sparse predictions, their performance still lags behind dense detectors with post-processing for 3D object detection from point clouds. DETRs usually adopt a larger number of queries than GTs (e.g., 300 queries v.s. ~40 objects in Waymo) in a scene, which inevitably incur many false positives during inference. In this paper, we propose a simple yet effective sparse 3D detector, named Query Contrast Voxel-DETR (ConQueR), to eliminate the challenging false positives, and achieve more accurate and sparser predictions. We observe that most false positives are highly overlapping in local regions, caused by the lack of explicit supervision to discriminate locally similar queries. We thus propose a Query Contrast mechanism to explicitly enhance queries towards their best-matched GTs over all unmatched query predictions. This is achieved by the construction of positive and negative GT-query pairs for each GT, and a contrastive loss to enhance positive GT-query pairs against negative ones based on feature similarities. ConQueR closes the gap of sparse and dense 3D detectors, and reduces ~60% false positives. Our single-frame ConQueR achieves 71.6 mAPH/L2 on the challenging Waymo Open Dataset validation set, outperforming previous sota methods by over 2.0 mAPH/L2. Code: https://github.com/poodarchu/EFG.

LiDAR mapping is important yet challenging in self-driving and mobile robotics. To tackle such a global point cloud registration problem, DeepMapping converts the complex map estimation into a self-supervised training of simple deep networks. Despite its broad convergence range on small datasets, DeepMapping still cannot produce satisfactory results on large-scale datasets with thousands of frames. This is due to the lack of loop closures and exact cross-frame point correspondences, and the slow convergence of its global localization network. We propose DeepMapping2 by adding two novel techniques to address these issues: (1) organization of training batch based on map topology from loop closing, and (2) self-supervised local-to-global point consistency loss leveraging pairwise registration. Our experiments and ablation studies on public datasets such as KITTI, NCLT, and Nebula, demonstrate the effectiveness of our method.

In this paper, we study the problem of unsupervised object detection from 3D point clouds in self-driving scenes. We present a simple yet effective method that exploits (i) point clustering in near-range areas where the point clouds are dense, (ii) temporal consistency to filter out noisy unsupervised detections, (iii) translation equivariance of CNNs to extend the auto-labels to long range, and (iv) self-supervision for improving on its own. Our approach, OYSTER (Object Discovery via Spatio-Temporal Refinement), does not impose constraints on data collection (such as repeated traversals of the same location), is able to detect objects in a zero-shot manner without supervised finetuning (even in sparse, distant regions), and continues to self-improve given more rounds of iterative self-training. To better measure model performance in self-driving scenarios, we propose a new planning-centric perception metric based on distance-to-collision. We demonstrate that our unsupervised object detector significantly outperforms unsupervised baselines on PandaSet and Argoverse 2 Sensor dataset, showing promise that self-supervision combined with object priors can enable object discovery in the wild. For more information, visit the project website: https://waabi.ai/research/oyster.

Occluded and long-range objects are ubiquitous and challenging for 3D object detection. Point cloud sequence data provide unique opportunities to improve such cases, as an occluded or distant object can be observed from different viewpoints or gets better visibility over time. However, the efficiency and effectiveness in encoding long-term sequence data can still be improved. In this work, we propose MoDAR, using motion forecasting outputs as a type of virtual modality, to augment LiDAR point clouds. The MoDAR modality propagates object information from temporal contexts to a target frame, represented as a set of virtual points, one for each object from a waypoint on a forecasted trajectory. A fused point cloud of both raw sensor points and the virtual points can then be fed to any off-the-shelf point-cloud based 3D object detector. Evaluated on the Waymo Open Dataset, our method significantly improves prior art detectors by using motion forecasting from extra-long sequences (e.g. 18 seconds), achieving new state of the arts, while not adding much computation overhead.

This work proposes a novel approach to 4D radar-based scene flow estimation via cross-modal learning. Our approach is motivated by the co-located sensing redundancy in modern autonomous vehicles. Such redundancy implicitly provides various forms of supervision cues to the radar scene flow estimation. Specifically, we introduce a multi-task model architecture for the identified cross-modal learning problem and propose loss functions to opportunistically engage scene flow estimation using multiple cross-modal constraints for effective model training. Extensive experiments show the state-of-the-art performance of our method and demonstrate the effectiveness of cross-modal supervised learning to infer more accurate 4D radar scene flow. We also show its usefulness to two subtasks - motion segmentation and ego-motion estimation. Our source code will be available on https://github.com/Toytiny/CMFlow.

Despite the increasing popularity of LiDAR sensors, perception algorithms using 3D LiDAR data struggle with the ‘sensor-bias problem’. Specifically, the performance of perception algorithms significantly drops when an unseen specification of LiDAR sensor is applied at test time due to the domain discrepancy. This paper presents a fast and flexible LiDAR augmentation method for the semantic segmentation task, called ‘LiDomAug’. It aggregates raw LiDAR scans and creates a LiDAR scan of any configurations with the consideration of dynamic distortion and occlusion, resulting in instant domain augmentation. Our on-demand augmentation module runs at 330 FPS, so it can be seamlessly integrated into the data loader in the learning framework. In our experiments, learning-based approaches aided with the proposed LiDomAug are less affected by the sensor-bias issue and achieve new state-of-the-art domain adaptation performances on SemanticKITTI and nuScenes dataset without the use of the target domain data. We also present a sensor-agnostic model that faithfully works on the various LiDAR configurations.

Whilst the availability of 3D LiDAR point cloud data has significantly grown in recent years, annotation remains expensive and time-consuming, leading to a demand for semi-supervised semantic segmentation methods with application domains such as autonomous driving. Existing work very often employs relatively large segmentation backbone networks to improve segmentation accuracy, at the expense of computational costs. In addition, many use uniform sampling to reduce ground truth data requirements for learning needed, often resulting in sub-optimal performance. To address these issues, we propose a new pipeline that employs a smaller architecture, requiring fewer ground-truth annotations to achieve superior segmentation accuracy compared to contemporary approaches. This is facilitated via a novel Sparse Depthwise Separable Convolution module that significantly reduces the network parameter count while retaining overall task performance. To effectively sub-sample our training data, we propose a new Spatio-Temporal Redundant Frame Downsampling (ST-RFD) method that leverages knowledge of sensor motion within the environment to extract a more diverse subset of training data frame samples. To leverage the use of limited annotated data samples, we further propose a soft pseudo-label method informed by LiDAR reflectivity. Our method outperforms contemporary semi-supervised work in terms of mIoU, using less labeled data, on the SemanticKITTI (59.5@5%) and ScribbleKITTI (58.1@5%) benchmark datasets, based on a 2.3× reduction in model parameters and 641× fewer multiply-add operations whilst also demonstrating significant performance improvement on limited training data (i.e., Less is More).

3D semantic segmentation on multi-scan large-scale point clouds plays an important role in autonomous systems. Unlike the single-scan-based semantic segmentation task, this task requires distinguishing the motion states of points in addition to their semantic categories. However, methods designed for single-scan-based segmentation tasks perform poorly on the multi-scan task due to the lacking of an effective way to integrate temporal information. We propose MarS3D, a plug-and-play motion-aware model for semantic segmentation on multi-scan 3D point clouds. This module can be flexibly combined with single-scan models to allow them to have multi-scan perception abilities. The model encompasses two key designs: the Cross-Frame Feature Embedding module for enriching representation learning and the Motion-Aware Feature Learning module for enhancing motion awareness. Extensive experiments show that MarS3D can improve the performance of the baseline model by a large margin. The code is available at https://github.com/CVMI-Lab/MarS3D.

Robust point cloud parsing under all-weather conditions is crucial to level-5 autonomy in autonomous driving. However, how to learn a universal 3D semantic segmentation (3DSS) model is largely neglected as most existing benchmarks are dominated by point clouds captured under normal weather. We introduce SemanticSTF, an adverse-weather point cloud dataset that provides dense point-level annotations and allows to study 3DSS under various adverse weather conditions. We investigate universal 3DSS modeling with two tasks: 1) domain adaptive 3DSS that adapts from normal-weather data to adverse-weather data; 2) domain generalized 3DSS that learns a generalizable model from normal-weather data. Our studies reveal the challenge while existing 3DSS methods encounter adverse-weather data, showing the great value of SemanticSTF in steering the future endeavor along this very meaningful research direction. In addition, we design a domain randomization technique that alternatively randomizes the geometry styles of point clouds and aggregates their encoded embeddings, ultimately leading to a generalizable model that effectively improves 3DSS under various adverse weather. The SemanticSTF and related codes are available at https://github.com/xiaoaoran/SemanticSTF.

Novel class discovery (NCD) for semantic segmentation is the task of learning a model that can segment unlabelled (novel) classes using only the supervision from labelled (base) classes. This problem has recently been pioneered for 2D image data, but no work exists for 3D point cloud data. In fact, the assumptions made for 2D are loosely applicable to 3D in this case. This paper is presented to advance the state of the art on point cloud data analysis in four directions. Firstly, we address the new problem of NCD for point cloud semantic segmentation. Secondly, we show that the transposition of the only existing NCD method for 2D semantic segmentation to 3D data is suboptimal. Thirdly, we present a new method for NCD based on online clustering that exploits uncertainty quantification to produce prototypes for pseudo-labelling the points of the novel classes. Lastly, we introduce a new evaluation protocol to assess the performance of NCD for point cloud semantic segmentation. We thoroughly evaluate our method on SemanticKITTI and SemanticPOSS datasets, showing that it can significantly outperform the baseline. Project page: https://github.com/LuigiRiz/NOPS.

Despite the tremendous progress of Masked Autoencoders (MAE) in developing vision tasks such as image and video, exploring MAE in large-scale 3D point clouds remains challenging due to the inherent irregularity. In contrast to previous 3D MAE frameworks, which either design a complex decoder to infer masked information from maintained regions or adopt sophisticated masking strategies, we instead propose a much simpler paradigm. The core idea is to apply a Generative Decoder for MAE (GD-MAE) to automatically merges the surrounding context to restore the masked geometric knowledge in a hierarchical fusion manner. In doing so, our approach is free from introducing the heuristic design of decoders and enjoys the flexibility of exploring various masking strategies. The corresponding part costs less than 12% latency compared with conventional methods, while achieving better performance. We demonstrate the efficacy of the proposed method on several large-scale benchmarks: Waymo, KITTI, and ONCE. Consistent improvement on downstream detection tasks illustrates strong robustness and generalization capability. Not only our method reveals state-of-the-art results, but remarkably, we achieve comparable accuracy even with 20% of the labeled data on the Waymo dataset. Code will be released.

As a pioneering work, PointContrast conducts unsupervised 3D representation learning via leveraging contrastive learning over raw RGB-D frames and proves its effectiveness on various downstream tasks. However, the trend of large-scale unsupervised learning in 3D has yet to emerge due to two stumbling blocks: the inefficiency of matching RGB-D frames as contrastive views and the annoying mode collapse phenomenon mentioned in previous works. Turning the two stumbling blocks into empirical stepping stones, we first propose an efficient and effective contrastive learning framework, which generates contrastive views directly on scene-level point clouds by a well-curated data augmentation pipeline and a practical view mixing strategy. Second, we introduce reconstructive learning on the contrastive learning framework with an exquisite design of contrastive cross masks, which targets the reconstruction of point color and surfel normal. Our Masked Scene Contrast (MSC) framework is capable of extracting comprehensive 3D representations more efficiently and effectively. It accelerates the pre-training procedure by at least 3x and still achieves an uncompromised performance compared with previous work. Besides, MSC also enables large-scale 3D pre-training across multiple datasets, which further boosts the performance and achieves state-of-the-art fine-tuning results on several downstream tasks, e.g., 75.5% mIoU on ScanNet semantic segmentation validation set.

Recently, point cloud semantic segmentation has attracted much attention in computer vision. Most of the existing works in literature assume that the training and testing point clouds have the same object classes, but they are generally invalid in many real-world scenarios for identifying the 3D objects whose classes are not seen in the training set. To address this problem, we propose an Adversarial Prototype Framework (APF) for handling the open-set 3D semantic segmentation task, which aims to identify 3D unseen-class points while maintaining the segmentation performance on seen-class points. The proposed APF consists of a feature extraction module for extracting point features, a prototypical constraint module, and a feature adversarial module. The prototypical constraint module is designed to learn prototypes for each seen class from point features. The feature adversarial module utilizes generative adversarial networks to estimate the distribution of unseen-class features implicitly, and the synthetic unseen-class features are utilized to prompt the model to learn more effective point features and prototypes for discriminating unseen-class samples from the seen-class ones. Experimental results on two public datasets demonstrate that the proposed APF outperforms the comparative methods by a large margin in most cases.

Point cloud completion addresses filling in the missing parts of a partial point cloud obtained from depth sensors and generating a complete point cloud. Although there has been steep progress in the supervised methods on the synthetic point cloud completion task, it is hardly applicable in real-world scenarios due to the domain gap between the synthetic and real-world datasets or the requirement of prior information. To overcome these limitations, we propose a novel self-supervised framework ACL-SPC for point cloud completion to train and test on the same data. ACL-SPC takes a single partial input and attempts to output the complete point cloud using an adaptive closed-loop (ACL) system that enforces the output same for the variation of an input. We evaluate our ACL-SPC on various datasets to prove that it can successfully learn to complete a partial point cloud as the first self-supervised scheme. Results show that our method is comparable with unsupervised methods and achieves superior performance on the real-world dataset compared to the supervised methods trained on the synthetic dataset. Extensive experiments justify the necessity of self-supervised learning and the effectiveness of our proposed method for the real-world point cloud completion task. The code is publicly available from this link.

Diffusion models have emerged as a powerful tool for point cloud generation. A key component that drives the impressive performance for generating high-quality samples from noise is iteratively denoise for thousands of steps. While beneficial, the complexity of learning steps has limited its applications to many 3D real-world. To address this limitation, we propose Point Straight Flow (PSF), a model that exhibits impressive performance using one step. Our idea is based on the reformulation of the standard diffusion model, which optimizes the curvy learning trajectory into a straight path. Further, we develop a distillation strategy to shorten the straight path into one step without a performance loss, enabling applications to 3D real-world with latency constraints. We perform evaluations on multiple 3D tasks and find that our PSF performs comparably to the standard diffusion model, outperforming other efficient 3D point cloud generation methods. On real-world applications such as point cloud completion and training-free text-guided generation in a low-latency setup, PSF performs favorably.

In point cloud analysis, point-based methods have rapidly developed in recent years. These methods have recently focused on concise MLP structures, such as PointNeXt, which have demonstrated competitiveness with Convolutional and Transformer structures. However, standard MLPs are limited in their ability to extract local features effectively. To address this limitation, we propose a Vector-oriented Point Set Abstraction that can aggregate neighboring features through higher-dimensional vectors. To facilitate network optimization, we construct a transformation from scalar to vector using independent angles based on 3D vector rotations. Finally, we develop a PointVector model that follows the structure of PointNeXt. Our experimental results demonstrate that PointVector achieves state-of-the-art performance 72.3% mIOU on the S3DIS Area 5 and 78.4% mIOU on the S3DIS (6-fold cross-validation) with only 58% model parameters of PointNeXt. We hope our work will help the exploration of concise and effective feature representations. The code will be released soon.

Problems such as equipment defects or limited viewpoints will lead the captured point clouds to be incomplete. Therefore, recovering the complete point clouds from the partial ones plays an vital role in many practical tasks, and one of the keys lies in the prediction of the missing part. In this paper, we propose a novel point cloud completion approach namely ProxyFormer that divides point clouds into existing (input) and missing (to be predicted) parts and each part communicates information through its proxies. Specifically, we fuse information into point proxy via feature and position extractor, and generate features for missing point proxies from the features of existing point proxies. Then, in order to better perceive the position of missing points, we design a missing part sensitive transformer, which converts random normal distribution into reasonable position information, and uses proxy alignment to refine the missing proxies. It makes the predicted point proxies more sensitive to the features and positions of the missing part, and thus makes these proxies more suitable for subsequent coarse-to-fine processes. Experimental results show that our method outperforms state-of-the-art completion networks on several benchmark datasets and has the fastest inference speed.

Contrastive learning has recently demonstrated great potential for unsupervised pre-training in 3D scene understanding tasks. However, most existing work randomly selects point features as anchors while building contrast, leading to a clear bias toward background points that often dominate in 3D scenes. Also, object awareness and foreground-to-background discrimination are neglected, making contrastive learning less effective. To tackle these issues, we propose a general foreground-aware feature contrast (FAC) framework to learn more effective point cloud representations in pre-training. FAC consists of two novel contrast designs to construct more effective and informative contrast pairs. The first is building positive pairs within the same foreground segment where points tend to have the same semantics. The second is that we prevent over-discrimination between 3D segments/objects and encourage foreground-to-background distinctions at the segment level with adaptive feature learning in a Siamese correspondence network, which adaptively learns feature correlations within and across point cloud views effectively. Visualization with point activation maps shows that our contrast pairs capture clear correspondences among foreground regions during pre-training. Quantitative experiments also show that FAC achieves superior knowledge transfer and data efficiency in various downstream 3D semantic segmentation and object detection tasks. All codes, data, and models are available at:https://github.com/KangchengLiu/FACForegroundAware_Contrast.

Most existing approaches for point cloud normal estimation aim to locally fit a geometric surface and calculate the normal from the fitted surface. Recently, learning-based methods have adopted a routine of predicting point-wise weights to solve the weighted least-squares surface fitting problem. Despite achieving remarkable progress, these methods overlook the approximation error of the fitting problem, resulting in a less accurate fitted surface. In this paper, we first carry out in-depth analysis of the approximation error in the surface fitting problem. Then, in order to bridge the gap between estimated and precise surface normals, we present two basic design principles: 1) applies the Z-direction Transform to rotate local patches for a better surface fitting with a lower approximation error; 2) models the error of the normal estimation as a learnable term. We implement these two principles using deep neural networks, and integrate them with the state-of-the-art (SOTA) normal estimation methods in a plug-and-play manner. Extensive experiments verify our approaches bring benefits to point cloud normal estimation and push the frontier of state-of-the-art performance on both synthetic and real-world datasets. The code is available at https://github.com/hikvision-research/3DVision.

Point cloud classification is an essential component in many security-critical applications such as autonomous driving and augmented reality. However, point cloud classifiers are vulnerable to adversarially perturbed point clouds. Existing certified defenses against adversarial point clouds suffer from a key limitation: their certified robustness guarantees are probabilistic, i.e., they produce an incorrect certified robustness guarantee with some probability. In this work, we propose a general framework, namely PointCert, that can transform an arbitrary point cloud classifier to be certifiably robust against adversarial point clouds with deterministic guarantees. PointCert certifiably predicts the same label for a point cloud when the number of arbitrarily added, deleted, and/or modified points is less than a threshold. Moreover, we propose multiple methods to optimize the certified robustness guarantees of PointCert in three application scenarios. We systematically evaluate PointCert on ModelNet and ScanObjectNN benchmark datasets. Our results show that PointCert substantially outperforms state-of-the-art certified defenses even though their robustness guarantees are probabilistic.

In this paper, we present a new method for the multiview registration of point cloud. Previous multiview registration methods rely on exhaustive pairwise registration to construct a densely-connected pose graph and apply Iteratively Reweighted Least Square (IRLS) on the pose graph to compute the scan poses. However, constructing a densely-connected graph is time-consuming and contains lots of outlier edges, which makes the subsequent IRLS struggle to find correct poses. To address the above problems, we first propose to use a neural network to estimate the overlap between scan pairs, which enables us to construct a sparse but reliable pose graph. Then, we design a novel history reweighting function in the IRLS scheme, which has strong robustness to outlier edges on the graph. In comparison with existing multiview registration methods, our method achieves 11% higher registration recall on the 3DMatch dataset and ~13% lower registration errors on the ScanNet dataset while reducing ~70% required pairwise registrations. Comprehensive ablation studies are conducted to demonstrate the effectiveness of our designs. The source code is available at https://github.com/WHU-USI3DV/SGHR.

Visible-modal object tracking gives rise to a series of downstream multi-modal tracking tributaries. To inherit the powerful representations of the foundation model, a natural modus operandi for multi-modal tracking is full fine-tuning on the RGB-based parameters. Albeit effective, this manner is not optimal due to the scarcity of downstream data and poor transferability, etc. In this paper, inspired by the recent success of the prompt learning in language models, we develop Visual Prompt multi-modal Tracking (ViPT), which learns the modal-relevant prompts to adapt the frozen pre-trained foundation model to various downstream multimodal tracking tasks. ViPT finds a better way to stimulate the knowledge of the RGB-based model that is pre-trained at scale, meanwhile only introducing a few trainable parameters (less than 1% of model parameters). ViPT outperforms the full fine-tuning paradigm on multiple downstream tracking tasks including RGB+Depth, RGB+Thermal, and RGB+Event tracking. Extensive experiments show the potential of visual prompt learning for multi-modal tracking, and ViPT can achieve state-of-the-art performance while satisfying parameter efficiency. Code and models are available at https://github.com/jiawen-zhu/ViPT.

The goal of correspondence pruning is to recognize correct correspondences (inliers) from initial ones, with applications to various feature matching based tasks. Seeking neighbors in the coordinate and feature spaces is a common strategy in many previous methods. However, it is difficult to ensure that these neighbors are always consistent, since the distribution of false correspondences is extremely irregular. For addressing this problem, we propose a novel global-graph space to search for consistent neighbors based on a weighted global graph that can explicitly explore long-range dependencies among correspondences. On top of that, we progressively construct three neighbor embeddings according to different neighbor search spaces, and design a Neighbor Consistency block to extract neighbor context and explore their interactions sequentially. In the end, we develop a Neighbor Consistency Mining Network (NCMNet) for accurately recovering camera poses and identifying inliers. Experimental results indicate that our NCMNet achieves a significant performance advantage over state-of-the-art competitors on challenging outdoor and indoor matching scenes. The source code can be found at https://github.com/xinliu29/NCMNet.

Learning inter-image similarity is crucial for 3D medical images self-supervised pre-training, due to their sharing of numerous same semantic regions. However, the lack of the semantic prior in metrics and the semantic-independent variation in 3D medical images make it challenging to get a reliable measurement for the inter-image similarity, hindering the learning of consistent representation for same semantics. We investigate the challenging problem of this task, i.e., learning a consistent representation between images for a clustering effect of same semantic features. We propose a novel visual similarity learning paradigm, Geometric Visual Similarity Learning, which embeds the prior of topological invariance into the measurement of the inter-image similarity for consistent representation of semantic regions. To drive this paradigm, we further construct a novel geometric matching head, the Z-matching head, to collaboratively learn the global and local similarity of semantic regions, guiding the efficient representation learning for different scale-level inter-image semantic features. Our experiments demonstrate that the pre-training with our learning of inter-image similarity yields more powerful inner-scene, inter-scene, and global-local transferring ability on four challenging 3D medical image tasks. Our codes and pre-trained models will be publicly available in https://github.com/YutingHe-list/GVSL.

Unsupervised visible-infrared person re-identification is a challenging task due to the large modality gap and the unavailability of cross-modality correspondences. Cross-modality correspondences are very crucial to bridge the modality gap. Some existing works try to mine cross-modality correspondences, but they focus only on local information. They do not fully exploit the global relationship across identities, thus limiting the quality of the mined correspondences. Worse still, the number of clusters of the two modalities is often inconsistent, exacerbating the unreliability of the generated correspondences. In response, we devise a Progressive Graph Matching method to globally mine cross-modality correspondences under cluster imbalance scenarios. PGM formulates correspondences mining as a graph matching process and considers the global information by minimizing the global matching cost, where the matching cost measures the dissimilarity of clusters. Besides, PGM adopts a progressive strategy to address the imbalance issue with multiple dynamic matching processes. Based on PGM, we design an Alternate Cross Contrastive Learning (ACCL) module to reduce the modality gap with the mined cross-modality correspondences, while mitigating the effect of noise in correspondences through an alternate scheme. Extensive experiments demonstrate the reliability of the generated correspondences and the effectiveness of our method.

Recently, deep Stereo Matching (SM) networks have shown impressive performance and attracted increasing attention in computer vision. However, existing deep SM networks are prone to learn dataset-dependent shortcuts, which fail to generalize well on unseen realistic datasets. This paper takes a step towards training robust models for the domain generalized SM task, which mainly focuses on learning shortcut-invariant representation from synthetic data to alleviate the domain shifts. Specifically, we propose a Hierarchical Visual Transformation (HVT) network to 1) first transform the training sample hierarchically into new domains with diverse distributions from three levels: Global, Local, and Pixel, 2) then maximize the visual discrepancy between the source domain and new domains, and minimize the cross-domain feature inconsistency to capture domain-invariant features. In this way, we can prevent the model from exploiting the artifacts of synthetic stereo images as shortcut features, thereby estimating the disparity maps more effectively based on the learned robust and shortcut-invariant representation. We integrate our proposed HVT network with SOTA SM networks and evaluate its effectiveness on several public SM benchmark datasets. Extensive experiments clearly show that the HVT network can substantially enhance the performance of existing SM networks in synthetic-to-realistic domain generalization.

Optical flow has achieved great success under clean scenes, but suffers from restricted performance under foggy scenes. To bridge the clean-to-foggy domain gap, the existing methods typically adopt the domain adaptation to transfer the motion knowledge from clean to synthetic foggy domain. However, these methods unexpectedly neglect the synthetic-to-real domain gap, and thus are erroneous when applied to real-world scenes. To handle the practical optical flow under real foggy scenes, in this work, we propose a novel unsupervised cumulative domain adaptation optical flow (UCDA-Flow) framework: depth-association motion adaptation and correlation-alignment motion adaptation. Specifically, we discover that depth is a key ingredient to influence the optical flow: the deeper depth, the inferior optical flow, which motivates us to design a depth-association motion adaptation module to bridge the clean-to-foggy domain gap. Moreover, we figure out that the cost volume correlation shares similar distribution of the synthetic and real foggy images, which enlightens us to devise a correlation-alignment motion adaptation module to distill motion knowledge of the synthetic foggy domain to the real foggy domain. Note that synthetic fog is designed as the intermediate domain. Under this unified framework, the proposed cumulative adaptation progressively transfers knowledge from clean scenes to real foggy scenes. Extensive experiments have been performed to verify the superiority of the proposed method.

We present PVO, a novel panoptic visual odometry framework to achieve more comprehensive modeling of the scene motion, geometry, and panoptic segmentation information. Our PVO models visual odometry (VO) and video panoptic segmentation (VPS) in a unified view, which makes the two tasks mutually beneficial. Specifically, we introduce a panoptic update module into the VO Module with the guidance of image panoptic segmentation. This Panoptic-Enhanced VO Module can alleviate the impact of dynamic objects in the camera pose estimation with a panoptic-aware dynamic mask. On the other hand, the VO-Enhanced VPS Module also improves the segmentation accuracy by fusing the panoptic segmentation result of the current frame on the fly to the adjacent frames, using geometric information such as camera pose, depth, and optical flow obtained from the VO Module. These two modules contribute to each other through recurrent iterative optimization. Extensive experiments demonstrate that PVO outperforms state-of-the-art methods in both visual odometry and video panoptic segmentation tasks.

Bird’s Eye View (BEV) semantic segmentation is a critical task in autonomous driving. However, existing Transformer-based methods confront difficulties in transforming Perspective View (PV) to BEV due to their unidirectional and posterior interaction mechanisms. To address this issue, we propose a novel Bi-directional and Early Interaction Transformers framework named BAEFormer, consisting of (i) an early-interaction PV-BEV pipeline and (ii) a bi-directional cross-attention mechanism. Moreover, we find that the image feature maps’ resolution in the cross-attention module has a limited effect on the final performance. Under this critical observation, we propose to enlarge the size of input images and downsample the multi-view image features for cross-interaction, further improving the accuracy while keeping the amount of computation controllable. Our proposed method for BEV semantic segmentation achieves state-of-the-art performance in real-time inference speed on the nuScenes dataset, i.e., 38.9 mIoU at 45 FPS on a single A100 GPU.

In recent years, vision-centric perception has flourished in various autonomous driving tasks, including 3D detection, semantic map construction, motion forecasting, and depth estimation. Nevertheless, the latency of vision-centric approaches is too high for practical deployment (e.g., most camera-based 3D detectors have a runtime greater than 300ms). To bridge the gap between ideal researches and real-world applications, it is necessary to quantify the trade-off between performance and efficiency. Traditionally, autonomous-driving perception benchmarks perform the online evaluation, neglecting the inference time delay. To mitigate the problem, we propose the Autonomous-driving StreAming Perception (ASAP) benchmark, which is the first benchmark to evaluate the online performance of vision-centric perception in autonomous driving. On the basis of the 2Hz annotated nuScenes dataset, we first propose an annotation-extending pipeline to generate high-frame-rate labels for the 12Hz raw images. Referring to the practical deployment, the Streaming Perception Under constRained-computation (SPUR) evaluation protocol is further constructed, where the 12Hz inputs are utilized for streaming evaluation under the constraints of different computational resources. In the ASAP benchmark, comprehensive experiment results reveal that the model rank alters under different constraints, suggesting that the model latency and computation budget should be considered as design choices to optimize the practical deployment. To facilitate further research, we establish baselines for camera-based streaming 3D detection, which consistently enhance the streaming performance across various hardware. The ASAP benchmark will be made publicly available.

Multi-task learning has emerged as a powerful paradigm to solve a range of tasks simultaneously with good efficiency in both computation resources and inference time. However, these algorithms are designed for different tasks mostly not within the scope of autonomous driving, thus making it hard to compare multi-task methods in autonomous driving. Aiming to enable the comprehensive evaluation of present multi-task learning methods in autonomous driving, we extensively investigate the performance of popular multi-task methods on the large-scale driving dataset, which covers four common perception tasks, i.e., object detection, semantic segmentation, drivable area segmentation, and lane detection. We provide an in-depth analysis of current multi-task learning methods under different common settings and find out that the existing methods make progress but there is still a large performance gap compared with single-task baselines. To alleviate this dilemma in autonomous driving, we present an effective multi-task framework, VE-Prompt, which introduces visual exemplars via task-specific prompting to guide the model toward learning high-quality task-specific representations. Specifically, we generate visual exemplars based on bounding boxes and color-based markers, which provide accurate visual appearances of target categories and further mitigate the performance gap. Furthermore, we bridge transformer-based encoders and convolutional layers for efficient and accurate unified perception in autonomous driving. Comprehensive experimental results on the diverse self-driving dataset BDD100K show that the VE-Prompt improves the multi-task baseline and further surpasses single-task models.

The prevailing way to test a self-driving vehicle (SDV) in simulation involves non-reactive open-loop replay of real world scenarios. However, in order to safely deploy SDVs to the real world, we need to evaluate them in closed-loop. Towards this goal, we propose to leverage the wealth of interesting scenarios captured in the real world and make them reactive and controllable to enable closed-loop SDV evaluation in what-if situations. In particular, we present MixSim, a hierarchical framework for mixed reality traffic simulation. MixSim explicitly models agent goals as routes along the road network and learns a reactive route-conditional policy. By inferring each agent’s route from the original scenario, MixSim can reactively re-simulate the scenario and enable testing different autonomy systems under the same conditions. Furthermore, by varying each agent’s route, we can expand the scope of testing to what-if situations with realistic variations in agent behaviors or even safety-critical interactions. Our experiments show that MixSim can serve as a realistic, reactive, and controllable digital twin of real world scenarios. For more information, please visit the project website: https://waabi.ai/research/mixsim/

Trajectory prediction is a crucial undertaking in understanding entity movement or human behavior from observed sequences. However, current methods often assume that the observed sequences are complete while ignoring the potential for missing values caused by object occlusion, scope limitation, sensor failure, etc. This limitation inevitably hinders the accuracy of trajectory prediction. To address this issue, our paper presents a unified framework, the Graph-based Conditional Variational Recurrent Neural Network (GC-VRNN), which can perform trajectory imputation and prediction simultaneously. Specifically, we introduce a novel Multi-Space Graph Neural Network (MS-GNN) that can extract spatial features from incomplete observations and leverage missing patterns. Additionally, we employ a Conditional VRNN with a specifically designed Temporal Decay (TD) module to capture temporal dependencies and temporal missing patterns in incomplete trajectories. The inclusion of the TD module allows for valuable information to be conveyed through the temporal flow. We also curate and benchmark three practical datasets for the joint problem of trajectory imputation and prediction. Extensive experiments verify the exceptional performance of our proposed method. As far as we know, this is the first work to address the lack of benchmarks and techniques for trajectory imputation and prediction in a unified manner.

We present MotionDiffuser, a diffusion based representation for the joint distribution of future trajectories over multiple agents. Such representation has several key advantages: first, our model learns a highly multimodal distribution that captures diverse future outcomes. Second, the simple predictor design requires only a single L2 loss training objective, and does not depend on trajectory anchors. Third, our model is capable of learning the joint distribution for the motion of multiple agents in a permutation-invariant manner. Furthermore, we utilize a compressed trajectory representation via PCA, which improves model performance and allows for efficient computation of the exact sample log probability. Subsequently, we propose a general constrained sampling framework that enables controlled trajectory sampling based on differentiable cost functions. This strategy enables a host of applications such as enforcing rules and physical priors, or creating tailored simulation scenarios. MotionDiffuser can be combined with existing backbone architectures to achieve top motion forecasting results. We obtain state-of-the-art results for multi-agent motion prediction on the Waymo Open Motion Dataset.

We propose the first framework to learn control policies for vision-based human-to-robot handovers, a critical task for human-robot interaction. While research in Embodied AI has made significant progress in training robot agents in simulated environments, interacting with humans remains challenging due to the difficulties of simulating humans. Fortunately, recent research has developed realistic simulated environments for human-to-robot handovers. Leveraging this result, we introduce a method that is trained with a human-in-the-loop via a two-stage teacher-student framework that uses motion and grasp planning, reinforcement learning, and self-supervision. We show significant performance gains over baselines on a simulation benchmark, sim-to-sim transfer and sim-to-real transfer.

Training embodied agents in simulation has become mainstream for the embodied AI community. However, these agents often struggle when deployed in the physical world due to their inability to generalize to real-world environments. In this paper, we present Phone2Proc, a method that uses a 10-minute phone scan and conditional procedural generation to create a distribution of training scenes that are semantically similar to the target environment. The generated scenes are conditioned on the wall layout and arrangement of large objects from the scan, while also sampling lighting, clutter, surface textures, and instances of smaller objects with randomized placement and materials. Leveraging just a simple RGB camera, training with Phone2Proc shows massive improvements from 34.7% to 70.7% success rate in sim-to-real ObjectNav performance across a test suite of over 200 trials in diverse real-world environments, including homes, offices, and RoboTHOR. Furthermore, Phone2Proc’s diverse distribution of generated scenes makes agents remarkably robust to changes in the real world, such as human movement, object rearrangement, lighting changes, or clutter.

We propose GazeNeRF, a 3D-aware method for the task of gaze redirection. Existing gaze redirection methods operate on 2D images and struggle to generate 3D consistent results. Instead, we build on the intuition that the face region and eye balls are separate 3D structures that move in a coordinated yet independent fashion. Our method leverages recent advancements in conditional image-based neural radiance fields and proposes a two-branch architecture that predicts volumetric features for the face and eye regions separately. Rigidly transforming the eye features via a 3D rotation matrix provides fine-grained control over the desired gaze angle. The final, redirected image is then attained via differentiable volume compositing. Our experiments show that this architecture outperforms naively conditioned NeRF baselines as well as previous state-of-the-art 2D gaze redirection methods in terms of redirection accuracy and identity preservation. Code and models will be released for research purposes.

Kalman filter (KF) based methods for multi-object tracking (MOT) make an assumption that objects move linearly. While this assumption is acceptable for very short periods of occlusion, linear estimates of motion for prolonged time can be highly inaccurate. Moreover, when there is no measurement available to update Kalman filter parameters, the standard convention is to trust the priori state estimations for posteriori update. This leads to the accumulation of errors during a period of occlusion. The error causes significant motion direction variance in practice. In this work, we show that a basic Kalman filter can still obtain state-of-the-art tracking performance if proper care is taken to fix the noise accumulated during occlusion. Instead of relying only on the linear state estimate (i.e., estimation-centric approach), we use object observations (i.e., the measurements by object detector) to compute a virtual trajectory over the occlusion period to fix the error accumulation of filter parameters. This allows more time steps to correct errors accumulated during occlusion. We name our method Observation-Centric SORT (OC-SORT). It remains Simple, Online, and Real-Time but improves robustness during occlusion and non-linear motion. Given off-the-shelf detections as input, OC-SORT runs at 700+ FPS on a single CPU. It achieves state-of-the-art on multiple datasets, including MOT17, MOT20, KITTI, head tracking, and especially DanceTrack where the object motion is highly non-linear. The code and models are available at https://github.com/noahcao/OC_SORT.

We present ARTrack, an autoregressive framework for visual object tracking. ARTrack tackles tracking as a coordinate sequence interpretation task that estimates object trajectories progressively, where the current estimate is induced by previous states and in turn affects subsequences. This time-autoregressive approach models the sequential evolution of trajectories to keep tracing the object across frames, making it superior to existing template matching based trackers that only consider the per-frame localization accuracy. ARTrack is simple and direct, eliminating customized localization heads and post-processings. Despite its simplicity, ARTrack achieves state-of-the-art performance on prevailing benchmark datasets.

Gait recognition is one of the most critical long-distance identification technologies and increasingly gains popularity in both research and industry communities. Despite the significant progress made in indoor datasets, much evidence shows that gait recognition techniques perform poorly in the wild. More importantly, we also find that some conclusions drawn from indoor datasets cannot be generalized to real applications. Therefore, the primary goal of this paper is to present a comprehensive benchmark study for better practicality rather than only a particular model for better performance. To this end, we first develop a flexible and efficient gait recognition codebase named OpenGait. Based on OpenGait, we deeply revisit the recent development of gait recognition by re-conducting the ablative experiments. Encouragingly,we detect some unperfect parts of certain prior woks, as well as new insights. Inspired by these discoveries, we develop a structurally simple, empirically powerful, and practically robust baseline model, GaitBase. Experimentally, we comprehensively compare GaitBase with many current gait recognition methods on multiple public datasets, and the results reflect that GaitBase achieves significantly strong performance in most cases regardless of indoor or outdoor situations. Code is available at https://github.com/ShiqiYu/OpenGait.

Self-supervised facial representation has recently attracted increasing attention due to its ability to perform face understanding without relying on large-scale annotated datasets heavily. However, analytically, current contrastive-based self-supervised learning (SSL) still performs unsatisfactorily for learning facial representation. More specifically, existing contrastive learning (CL) tends to learn pose-invariant features that cannot depict the pose details of faces, compromising the learning performance. To conquer the above limitation of CL, we propose a novel Pose-disentangled Contrastive Learning (PCL) method for general self-supervised facial representation. Our PCL first devises a pose-disentangled decoder (PDD) with a delicately designed orthogonalizing regulation, which disentangles the pose-related features from the face-aware features; therefore, pose-related and other pose-unrelated facial information could be performed in individual subnetworks and do not affect each other’s training. Furthermore, we introduce a pose-related contrastive learning scheme that learns pose-related information based on data augmentation of the same image, which would deliver more effective face-aware representation for various downstream tasks. We conducted linear evaluation on four challenging downstream facial understanding tasks, i.e., facial expression recognition, face recognition, AU detection and head pose estimation.Experimental results demonstrate that PCL significantly outperforms cutting-edge SSL methods. Our Code is available at https://github.com/DreamMr/PCL.

Generating talking face videos from audio attracts lots of research interest. A few person-specific methods can generate vivid videos but require the target speaker’s videos for training or fine-tuning. Existing person-generic methods have difficulty in generating realistic and lip-synced videos while preserving identity information. To tackle this problem, we propose a two-stage framework consisting of audio-to-landmark generation and landmark-to-video rendering procedures. First, we devise a novel Transformer-based landmark generator to infer lip and jaw landmarks from the audio. Prior landmark characteristics of the speaker’s face are employed to make the generated landmarks coincide with the facial outline of the speaker. Then, a video rendering model is built to translate the generated landmarks into face images. During this stage, prior appearance information is extracted from the lower-half occluded target face and static reference images, which helps generate realistic and identity-preserving visual content. For effectively exploring the prior information of static reference images, we align static reference images with the target face’s pose and expression based on motion fields. Moreover, auditory features are reused to guarantee that the generated face images are well synchronized with the audio. Extensive experiments demonstrate that our method can produce more realistic, lip-synced, and identity-preserving videos than existing person-generic talking face generation methods.

Deepfake detection is gaining significant importance in the research community. While most of the research efforts are focused around high-quality images and videos, deepfake generation algorithms today have the capability to generate low-resolution videos, occluded deepfakes, and multiple-subject deepfakes. In this research, we emulate the real-world scenario of deepfake generation and spreading, and propose the DF-Platter dataset, which contains (i) both low-resolution and high-resolution deepfakes generated using multiple generation techniques and (ii) single-subject and multiple-subject deepfakes, with face images of Indian ethnicity. Faces in the dataset are annotated for various attributes such as gender, age, skin tone, and occlusion. The database is prepared in 116 days with continuous usage of 32 GPUs accounting to 1,800 GB cumulative memory. With over 500 GBs in size, the dataset contains a total of 133,260 videos encompassing three sets. To the best of our knowledge, this is one of the largest datasets containing vast variability and multiple challenges. We also provide benchmark results under multiple evaluation settings using popular and state-of-the-art deepfake detection models. Further, benchmark results under c23 and c40 compression are provided. The results demonstrate a significant performance reduction in the deepfake detection task on low-resolution deepfakes and show that the existing techniques fail drastically on multiple-subject deepfakes. It is our assertion that this database will improve the state-of-the-art by extending the capabilities of deepfake detection algorithms to real-world scenarios. The database is available at: http://iab-rubric.org/df-platter-database.

Modeling sounds emitted from physical object interactions is critical for immersive perceptual experiences in real and virtual worlds. Traditional methods of impact sound synthesis use physics simulation to obtain a set of physics parameters that could represent and synthesize the sound. However, they require fine details of both the object geometries and impact locations, which are rarely available in the real world and can not be applied to synthesize impact sounds from common videos. On the other hand, existing video-driven deep learning-based approaches could only capture the weak correspondence between visual content and impact sounds since they lack of physics knowledge. In this work, we propose a physics-driven diffusion model that can synthesize high-fidelity impact sound for a silent video clip. In addition to the video content, we propose to use additional physics priors to guide the impact sound synthesis procedure. The physics priors include both physics parameters that are directly estimated from noisy real-world impact sound examples without sophisticated setup and learned residual parameters that interpret the sound environment via neural networks. We further implement a novel diffusion model with specific training and inference strategies to combine physics priors and visual information for impact sound synthesis. Experimental results show that our model outperforms several existing systems in generating realistic impact sounds. More importantly, the physics-based representations are fully interpretable and transparent, thus enabling us to perform sound editing flexibly. We encourage the readers to visit our project page to watch demo videos with audio turned on to experience the results.

Conventional methods for human motion synthesis have either been deterministic or have had to struggle with the trade-off between motion diversity vs~motion quality. In response to these limitations, we introduce MoFusion, i.e., a new denoising-diffusion-based framework for high-quality conditional human motion synthesis that can synthesise long, temporally plausible, and semantically accurate motions based on a range of conditioning contexts (such as music and text). We also present ways to introduce well-known kinematic losses for motion plausibility within the motion-diffusion framework through our scheduled weighting strategy. The learned latent space can be used for several interactive motion-editing applications like in-betweening, seed-conditioning, and text-based editing, thus, providing crucial abilities for virtual-character animation and robotics. Through comprehensive quantitative evaluations and a perceptual user study, we demonstrate the effectiveness of MoFusion compared to the state-of-the-art on established benchmarks in the literature. We urge the reader to watch our supplementary video. The source code will be released.

In virtually all event-based vision problems, there is the need to select the most recent events, which are assumed to carry the most relevant information content. To achieve this, at least one of three main strategies is applied, namely: 1) constant temporal decay or fixed time window, 2) constant number of events, and 3) flow-based lifetime of events. However, these strategies suffer from at least one major limitation each. We instead propose a novel decay process for event cameras that adapts to the global scene dynamics and whose latency is in the order of nanoseconds. The main idea is to construct an adaptive quantity that encodes the global scene dynamics, denoted by event activity. The proposed method is evaluated in several event-based vision problems and datasets, consistently improving the corresponding baseline methods’ performance. We thus believe it can have a significant widespread impact on event-based research. Code available: https://github.com/neuromorphic-paris/event_batch.

Most existing RGB-based trackers target low frame rate benchmarks of around 30 frames per second. This setting restricts the tracker’s functionality in the real world, especially for fast motion. Event-based cameras as bioinspired sensors provide considerable potential for high frame rate tracking due to their high temporal resolution. However, event-based cameras cannot offer fine-grained texture information like conventional cameras. This unique complementarity motivates us to combine conventional frames and events for high frame rate object tracking under various challenging conditions. In this paper, we propose an end-to-end network consisting of multi-modality alignment and fusion modules to effectively combine meaningful information from both modalities at different measurement rates. The alignment module is responsible for cross-modality and cross-frame-rate alignment between frame and event modalities under the guidance of the moving cues furnished by events. While the fusion module is accountable for emphasizing valuable features and suppressing noise information by the mutual complement between the two modalities. Extensive experiments show that the proposed approach outperforms state-of-the-art trackers by a significant margin in high frame rate tracking. With the FE240hz dataset, our approach achieves high frame rate tracking up to 240Hz.

Video frame interpolation (VFI) is the task that synthesizes the intermediate frame given two consecutive frames. Most of the previous studies have focused on appropriate frame warping operations and refinement modules for the warped frames. These studies have been conducted on natural videos containing only continuous motions. However, many practical videos contain various unnatural objects with discontinuous motions such as logos, user interfaces and subtitles. We propose three techniques that can make the existing deep learning-based VFI architectures robust to these elements. First is a novel data augmentation strategy called figure-text mixing (FTM) which can make the models learn discontinuous motions during training stage without any extra dataset. Second, we propose a simple but effective module that predicts a map called discontinuity map (D-map), which densely distinguishes between areas of continuous and discontinuous motions. Lastly, we propose loss functions to give supervisions of the discontinuous motion areas which can be applied along with FTM and D-map. We additionally collect a special test benchmark called Graphical Discontinuous Motion (GDM) dataset consisting of some mobile games and chatting videos. Applied to the various state-of-the-art VFI networks, our method significantly improves the interpolation qualities on the videos from not only GDM dataset, but also the existing benchmarks containing only continuous motions such as Vimeo90K, UCF101, and DAVIS.

We present All-Pairs Multi-Field Transforms (AMT), a new network architecture for video frame interpolation. It is based on two essential designs. First, we build bidirectional correlation volumes for all pairs of pixels and use the predicted bilateral flows to retrieve correlations for updating both flows and the interpolated content feature. Second, we derive multiple groups of fine-grained flow fields from one pair of updated coarse flows for performing backward warping on the input frames separately. Combining these two designs enables us to generate promising task-oriented flows and reduce the difficulties in modeling large motions and handling occluded areas during frame interpolation. These qualities promote our model to achieve state-of-the-art performance on various benchmarks with high efficiency. Moreover, our convolution-based model competes favorably compared to Transformer-based models in terms of accuracy and efficiency. Our code is available at https://github.com/MCG-NKU/AMT.

Video frame interpolation has seen important progress in recent years, thanks to developments in several directions. Some works leverage better optical flow methods with improved splatting strategies or additional cues from depth, while others have investigated alternative approaches through direct predictions or transformers. Still, the problem remains unsolved in more challenging conditions such as complex lighting or large motion. In this work, we are bridging the gap towards video production with a novel transformer-based interpolation network architecture capable of estimating the expected error together with the interpolated frame. This offers several advantages that are of key importance for frame interpolation usage: First, we obtained improved visual quality over several datasets. The improvement in terms of quality is also clearly demonstrated through a user study. Second, our method estimates error maps for the interpolated frame, which are essential for real-life applications on longer video sequences where problematic frames need to be flagged. Finally, for rendered content a partial rendering pass of the intermediate frame, guided by the predicted error, can be utilized during the interpolation to generate a new frame of superior quality. Through this error estimation, our method can produce even higher-quality intermediate frames using only a fraction of the time compared to a full rendering.

Video restoration, which aims to restore clear frames from degraded videos, has numerous important applications. The key to video restoration depends on utilizing inter-frame information. However, existing deep learning methods often rely on complicated network architectures, such as optical flow estimation, deformable convolution, and cross-frame self-attention layers, resulting in high computational costs. In this study, we propose a simple yet effective framework for video restoration. Our approach is based on grouped spatial-temporal shift, which is a lightweight and straightforward technique that can implicitly capture inter-frame correspondences for multi-frame aggregation. By introducing grouped spatial shift, we attain expansive effective receptive fields. Combined with basic 2D convolution, this simple framework can effectively aggregate inter-frame information. Extensive experiments demonstrate that our framework outperforms the previous state-of-the-art method, while using less than a quarter of its computational cost, on both video deblurring and video denoising tasks. These results indicate the potential for our approach to significantly reduce computational overhead while maintaining high-quality results. Code is avaliable at https://github.com/dasongli1/Shift-Net.

We propose the Recurrent homography estimation framework using Homography-guided image Warping and Focus transformer (FocusFormer), named RHWF. Both being appropriately absorbed into the recurrent framework, the homography-guided image warping progressively enhances the feature consistency and the attention-focusing mechanism in FocusFormer aggregates the intra-inter correspondence in a global->nonlocal->local manner. Thanks to the above strategies, RHWF ranks top in accuracy on a variety of datasets, including the challenging cross-resolution and cross-modal ones. Meanwhile, benefiting from the recurrent framework, RHWF achieves parameter efficiency despite the transformer architecture. Compared to previous state-of-the-art approaches LocalTrans and IHN, RHWF reduces the mean average corner error (MACE) by about 70% and 38.1% on the MSCOCO dataset, while saving the parameter costs by 86.5% and 24.6%. Similar to the previous works, RHWF can also be arranged in 1-scale for efficiency and 2-scale for accuracy, with the 1-scale RHWF already outperforming most of the previous methods. Source code is available at https://github.com/imdumpl78/RHWF.

We consider the challenging task of training models for image-to-video deblurring, which aims to recover a sequence of sharp images corresponding to a given blurry image input. A critical issue disturbing the training of an image-to-video model is the ambiguity of the frame ordering since both the forward and backward sequences are plausible solutions. This paper proposes an effective self-supervised ordering scheme that allows training high-quality image-to-video deblurring models. Unlike previous methods that rely on order-invariant losses, we assign an explicit order for each video sequence, thus avoiding the order-ambiguity issue. Specifically, we map each video sequence to a vector in a latent high-dimensional space so that there exists a hyperplane such that for every video sequence, the vectors extracted from it and its reversed sequence are on different sides of the hyperplane. The side of the vectors will be used to define the order of the corresponding sequence. Last but not least, we propose a real-image dataset for the image-to-video deblurring problem that covers a variety of popular domains, including face, hand, and street. Extensive experimental results confirm the effectiveness of our method. Code and data are available at https://github.com/VinAIResearch/HyperCUT.git

Multi-modal image registration spatially aligns two images with different distributions. One of its major challenges is that images acquired from different imaging machines have different imaging distributions, making it difficult to focus only on the spatial aspect of the images and ignore differences in distributions. In this study, we developed a self-supervised approach, Indescribable Multi-model Spatial Evaluator (IMSE), to address multi-modal image registration. IMSE creates an accurate multi-modal spatial evaluator to measure spatial differences between two images, and then optimizes registration by minimizing the error predicted of the evaluator. To optimize IMSE performance, we also proposed a new style enhancement method called Shuffle Remap which randomizes the image distribution into multiple segments, and then randomly disorders and remaps these segments, so that the distribution of the original image is changed. Shuffle Remap can help IMSE to predict the difference in spatial location from unseen target distributions. Our results show that IMSE outperformed the existing methods for registration using T1-T2 and CT-MRI datasets. IMSE also can be easily integrated into the traditional registration process, and can provide a convenient way to evaluate and visualize registration results. IMSE also has the potential to be used as a new paradigm for image-to-image translation. Our code is available at https://github.com/Kid-Liet/IMSE.

Images taken in a low light condition with the presence of camera shake suffer from motion blur and photon shot noise. While state-of-the-art image restoration networks show promising results, they are largely limited to well-illuminated scenes and their performance drops significantly when photon shot noise is strong. In this paper, we propose a new blur estimation technique customized for photon-limited conditions. The proposed method employs a gradient-based backpropagation method to estimate the blur kernel. By modeling the blur kernel using a low-dimensional representation with the key points on the motion trajectory, we significantly reduce the search space and improve the regularity of the kernel estimation problem. When plugged into an iterative framework, our novel low-dimensional representation provides improved kernel estimates and hence significantly better deconvolution performance when compared to end-to-end trained neural networks.

Single-chip polarized color photography provides both visual textures and object surface information in one snapshot. However, the use of an additional directional polarizing filter array tends to lower photon count and SNR, when compared to conventional color imaging. As a result, such a bilayer structure usually leads to unpleasant noisy images and undermines performance of polarization analysis, especially in low-light conditions. It is a challenge for traditional image processing pipelines owing to the fact that the physical constraints exerted implicitly in the channels are excessively complicated. In this paper, we propose to tackle this issue through a noise modeling method for realistic data synthesis and a powerful network structure inspired by vision Transformer. A real-world polarized color image dataset of paired raw short-exposed noisy images and long-exposed reference images is captured for experimental evaluation, which has demonstrated the effectiveness of our approaches for data synthesis and polarized color image denoising.

Non-blind deblurring methods achieve decent performance under the accurate blur kernel assumption. Since the kernel uncertainty (i.e. kernel error) is inevitable in practice, semi-blind deblurring is suggested to handle it by introducing the prior of the kernel (or induced) error. However, how to design a suitable prior for the kernel (or induced) error remains challenging. Hand-crafted prior, incorporating domain knowledge, generally performs well but may lead to poor performance when kernel (or induced) error is complex. Data-driven prior, which excessively depends on the diversity and abundance of training data, is vulnerable to out-of-distribution blurs and images. To address this challenge, we suggest a dataset-free deep residual prior for the kernel induced error (termed as residual) expressed by a customized untrained deep neural network, which allows us to flexibly adapt to different blurs and images in real scenarios. By organically integrating the respective strengths of deep priors and hand-crafted priors, we propose an unsupervised semi-blind deblurring model which recovers the latent image from the blurry image and inaccurate blur kernel. To tackle the formulated model, an efficient alternating minimization algorithm is developed. Extensive experiments demonstrate the favorable performance of the proposed method as compared to model-driven and data-driven methods in terms of image quality and the robustness to different types of kernel error.

This paper proposes a new framework for low-light image enhancement by simultaneously conducting the appearance as well as structure modeling. It employs the structural feature to guide the appearance enhancement, leading to sharp and realistic results. The structure modeling in our framework is implemented as the edge detection in low-light images. It is achieved with a modified generative model via designing a structure-aware feature extractor and generator. The detected edge maps can accurately emphasize the essential structural information, and the edge prediction is robust towards the noises in dark areas. Moreover, to improve the appearance modeling, which is implemented with a simple U-Net, a novel structure-guided enhancement module is proposed with structure-guided feature synthesis layers. The appearance modeling, edge detector, and enhancement module can be trained end-to-end. The experiments are conducted on representative datasets (sRGB and RAW domains), showing that our model consistently achieves SOTA performance on all datasets with the same architecture.

Exposure correction task aims to correct the underexposure and its adverse overexposure images to the normal exposure in a single network. As well recognized, the optimization flow is opposite. Despite the great advancement, existing exposure correction methods are usually trained with a mini-batch of both underexposure and overexposure mixed samples and have not explored the relationship between them to solve the optimization inconsistency. In this paper, we introduce a new perspective to conjunct their optimization processes by correlating and constraining the relationship of correction procedure in a mini-batch. The core designs of our framework consist of two steps: 1) formulating the exposure relationship of samples across the batch dimension via a context-irrelevant pretext task. 2) delivering the above sample relationship design as the regularization term within the loss function to promote optimization consistency. The proposed sample relationship design as a general term can be easily integrated into existing exposure correction methods without any computational burden in inference time. Extensive experiments over multiple representative exposure correction benchmarks demonstrate consistent performance gains by introducing our sample relationship design.

Significant progress has been made in self-supervised image denoising (SSID) in the recent few years. However, most methods focus on dealing with spatially independent noise, and they have little practicality on real-world sRGB images with spatially correlated noise. Although pixel-shuffle downsampling has been suggested for breaking the noise correlation, it breaks the original information of images, which limits the denoising performance. In this paper, we propose a novel perspective to solve this problem, i.e., seeking for spatially adaptive supervision for real-world sRGB image denoising. Specifically, we take into account the respective characteristics of flat and textured regions in noisy images, and construct supervisions for them separately. For flat areas, the supervision can be safely derived from non-adjacent pixels, which are much far from the current pixel for excluding the influence of the noise-correlated ones. And we extend the blind-spot network to a blind-neighborhood network (BNN) for providing supervision on flat areas. For textured regions, the supervision has to be closely related to the content of adjacent pixels. And we present a locally aware network (LAN) to meet the requirement, while LAN itself is selectively supervised with the output of BNN. Combining these two supervisions, a denoising network (e.g., U-Net) can be well-trained. Extensive experiments show that our method performs favorably against state-of-the-art SSID methods on real-world sRGB photographs. The code is available at https://github.com/nagejacob/SpatiallyAdaptiveSSID.

Hyperspectral image (HSI) classification aims at assigning a unique label for every pixel to identify categories of different land covers. Existing deep learning models for HSIs are usually performed in a traditional learning paradigm. Being emerging machines, quantum computers are limited in the noisy intermediate-scale quantum (NISQ) era. The quantum theory offers a new paradigm for designing deep learning models. Motivated by the quantum circuit (QC) model, we propose a quantum-inspired spectral-spatial network (QSSN) for HSI feature extraction. The proposed QSSN consists of a phase-prediction module (PPM) and a measurement-like fusion module (MFM) inspired from quantum theory to dynamically fuse spectral and spatial information. Specifically, QSSN uses a quantum representation to represent an HSI cuboid and extracts joint spectral-spatial features using MFM. An HSI cuboid and its phases predicted by PPM are used in the quantum representation. Using QSSN as the building block, we propose an end-to-end quantum-inspired spectral-spatial pyramid network (QSSPN) for HSI feature extraction and classification. In this pyramid framework, QSSPN progressively learns feature representations by cascading QSSN blocks and performs classification with a softmax classifier. It is the first attempt to introduce quantum theory in HSI processing model design. Substantial experiments are conducted on three HSI datasets to verify the superiority of the proposed QSSPN framework over the state-of-the-art methods.

Existing image restoration methods mostly leverage the posterior distribution of natural images. However, they often assume known degradation and also require supervised training, which restricts their adaptation to complex real applications. In this work, we propose the Generative Diffusion Prior (GDP) to effectively model the posterior distributions in an unsupervised sampling manner. GDP utilizes a pre-train denoising diffusion generative model (DDPM) for solving linear inverse, non-linear, or blind problems. Specifically, GDP systematically explores a protocol of conditional guidance, which is verified more practical than the commonly used guidance way. Furthermore, GDP is strength at optimizing the parameters of degradation model during denoising process, achieving blind image restoration. Besides, we devise hierarchical guidance and patch-based methods, enabling the GDP to generate images of arbitrary resolutions. Experimentally, we demonstrate GDP’s versatility on several image datasets for linear problems, such as super-resolution, deblurring, inpainting, and colorization, as well as non-linear and blind issues, such as low-light enhancement and HDR image recovery. GDP outperforms the current leading unsupervised methods on the diverse benchmarks in reconstruction quality and perceptual quality. Moreover, GDP also generalizes well for natural images or synthesized images with arbitrary sizes from various tasks out of the distribution of the ImageNet training set.

Deep learning has become an important tool for reconstructing images in compressive sampling (CS). This paper proposes a ground-truth (GT) free meta-learning method for CS, which leverages both external and internal learning for unsupervised high-quality image reconstruction. The proposed method first trains a deep model via external meta-learning using only CS measurements, and then efficiently adapts the trained model to a test sample for further improvement by exploiting its internal characteristics. The meta-learning and model adaptation are built on an improved Stein’s unbiased risk estimator (iSURE) that provides efficient computation and effective guidance for accurate prediction in the range space of the adjoint of the measurement matrix. To further improve the learning on the null space of the measurement matrix, a modified model-agnostic meta-learning scheme is proposed, along with a null-space-consistent loss and a bias-adaptive deep unrolling network to improve and accelerate model adaption in test time. Experimental results have demonstrated that the proposed GT-free method performs well, and can even compete with supervised learning-based methods.

Very low-resolution face recognition (VLRFR) poses unique challenges, such as tiny regions of interest and poor resolution due to extreme standoff distance or wide viewing angle of the acquisition device. In this paper, we study principled approaches to elevate the recognizability of a face in the embedding space instead of the visual quality. We first formulate a robust learning-based face recognizability measure, namely recognizability index (RI), based on two criteria: (i) proximity of each face embedding against the unrecognizable faces cluster center and (ii) closeness of each face embedding against its positive and negative class prototypes. We then devise an index diversion loss to push the hard-to-recognize face embedding with low RI away from unrecognizable faces cluster to boost the RI, which reflects better recognizability. Additionally, a perceptibility-aware attention mechanism is introduced to attend to the salient recognizable face regions, which offers better explanatory and discriminative content for embedding learning. Our proposed model is trained end-to-end and simultaneously serves recognizability-aware embedding learning and face quality estimation. To address VLRFR, extensive evaluations on three challenging low-resolution datasets and face quality assessment demonstrate the superiority of the proposed model over the state-of-the-art methods.

Year after year, the demand for ever-better smartphone photos continues to grow, in particular in the domain of portrait photography. Manufacturers thus use perceptual quality criteria throughout the development of smartphone cameras. This costly procedure can be partially replaced by automated learning-based methods for image quality assessment (IQA). Due to its subjective nature, it is necessary to estimate and guarantee the consistency of the IQA process, a characteristic lacking in the mean opinion scores (MOS) widely used for crowdsourcing IQA. In addition, existing blind IQA (BIQA) datasets pay little attention to the difficulty of cross-content assessment, which may degrade the quality of annotations. This paper introduces PIQ23, a portrait-specific IQA dataset of 5116 images of 50 predefined scenarios acquired by 100 smartphones, covering a high variety of brands, models, and use cases. The dataset includes individuals of various genders and ethnicities who have given explicit and informed consent for their photographs to be used in public research. It is annotated by pairwise comparisons (PWC) collected from over 30 image quality experts for three image attributes: face detail preservation, face target exposure, and overall image quality. An in-depth statistical analysis of these annotations allows us to evaluate their consistency over PIQ23. Finally, we show through an extensive comparison with existing baselines that semantic information (image context) can be used to improve IQA predictions.

In this paper, we study a real-world JPEG image restoration problem with bit errors on the encrypted bitstream. The bit errors bring unpredictable color casts and block shifts on decoded image contents, which cannot be trivially resolved by existing image restoration methods mainly relying on pre-defined degradation models in the pixel domain. To address these challenges, we propose a robust JPEG decoder, followed by a two-stage compensation and alignment framework to restore bitstream-corrupted JPEG images. Specifically, the robust JPEG decoder adopts an error-resilient mechanism to decode the corrupted JPEG bitstream. The two-stage framework is composed of the self-compensation and alignment (SCA) stage and the guided-compensation and alignment (GCA) stage. The SCA adaptively performs block-wise image color compensation and alignment based on the estimated color and block offsets via image content similarity. The GCA leverages the extracted low-resolution thumbnail from the JPEG header to guide full-resolution pixel-wise image restoration in a coarse-to-fine manner. It is achieved by a coarse-guided pix2pix network and a refine-guided bi-directional Laplacian pyramid fusion network. We conduct experiments on three benchmarks with varying degrees of bit error rates. Experimental results and ablation studies demonstrate the superiority of our proposed method. The code will be released at https://github.com/wenyang001/Two-ACIR.

For modern high-resolution imaging sensors, pixel binning is performed in low-lighting conditions and in case high frame rates are required. To recover the original spatial resolution, single-image super-resolution techniques can be applied for upscaling. To achieve a higher image quality after upscaling, we propose a novel binning concept using tetromino-shaped pixels. It is embedded into the field of compressed sensing and the coherence is calculated to motivate the sensor layouts used. Next, we investigate the reconstruction quality using tetromino pixels for the first time in literature. Instead of using different types of tetrominoes as proposed elsewhere, we show that using a small repeating cell consisting of only four T-tetrominoes is sufficient. For reconstruction, we use a locally fully connected reconstruction (LFCR) network as well as two classical reconstruction methods from the field of compressed sensing. Using the LFCR network in combination with the proposed tetromino layout, we achieve superior image quality in terms of PSNR, SSIM, and visually compared to conventional single-image super-resolution using the very deep super-resolution (VDSR) network. For PSNR, a gain of up to +1.92 dB is achieved.

Neural fields have rapidly been adopted for representing 3D signals, but their application to more classical 2D image-processing has been relatively limited. In this paper, we consider one of the most important operations in image processing: upsampling. In deep learning, learnable upsampling layers have extensively been used for single image super-resolution. We propose to parameterize upsampling kernels as neural fields. This parameterization leads to a compact architecture that obtains a 40-fold reduction in the number of parameters when compared with competing arbitrary-scale super-resolution architectures. When upsampling images of size 256x256 we show that our architecture is 2x-10x more efficient than competing arbitrary-scale super-resolution architectures, and more efficient than sub-pixel convolutions when instantiated to a single-scale model. In the general setting, these gains grow polynomially with the square of the target scale. We validate our method on standard benchmarks showing such efficiency gains can be achieved without sacrifices in super-resolution performance.

Arbitrary-scale image super-resolution (SR) is often tackled using the implicit neural representation (INR) approach, which relies on a position encoding scheme to improve its representation ability. In this paper, we introduce orthogonal position encoding (OPE), an extension of position encoding, and an OPE-Upscale module to replace the INR-based upsampling module for arbitrary-scale image super-resolution. Our OPE-Upscale module takes 2D coordinates and latent code as inputs, just like INR, but does not require any training parameters. This parameter-free feature allows the OPE-Upscale module to directly perform linear combination operations, resulting in continuous image reconstruction and achieving arbitrary-scale image reconstruction. As a concise SR framework, our method is computationally efficient and consumes less memory than state-of-the-art methods, as confirmed by extensive experiments and evaluations. In addition, our method achieves comparable results with state-of-the-art methods in arbitrary-scale image super-resolution. Lastly, we show that OPE corresponds to a set of orthogonal basis, validating our design principle.

Image super-resolution (SR) has attracted increasing attention due to its wide applications. However, current SR methods generally suffer from over-smoothing and artifacts, and most work only with fixed magnifications. This paper introduces an Implicit Diffusion Model (IDM) for high-fidelity continuous image super-resolution. IDM integrates an implicit neural representation and a denoising diffusion model in a unified end-to-end framework, where the implicit neural representation is adopted in the decoding process to learn continuous-resolution representation. Furthermore, we design a scale-controllable conditioning mechanism that consists of a low-resolution (LR) conditioning network and a scaling factor. The scaling factor regulates the resolution and accordingly modulates the proportion of the LR information and generated features in the final output, which enables the model to accommodate the continuous-resolution requirement. Extensive experiments validate the effectiveness of our IDM and demonstrate its superior performance over prior arts.

Salient object detection (SOD) aims to mimic the human visual system (HVS) and cognition mechanisms to identify and segment salient objects. However, due to the complexity of these mechanisms, current methods are not perfect. Accuracy and robustness need to be further improved, particularly in complex scenes with multiple objects and background clutter. To address this issue, we propose a novel approach called Multiple Enhancement Network (MENet) that adopts the boundary sensibility, content integrity, iterative refinement, and frequency decomposition mechanisms of HVS. A multi-level hybrid loss is firstly designed to guide the network to learn pixel-level, region-level, and object-level features. A flexible multiscale feature enhancement module (ME-Module) is then designed to gradually aggregate and refine global or detailed features by changing the size order of the input feature sequence. An iterative training strategy is used to enhance boundary features and adaptive features in the dual-branch decoder of MENet. Comprehensive evaluations on six challenging benchmark datasets show that MENet achieves state-of-the-art results. Both the codes and results are publicly available at https://github.com/yiwangtz/MENet.

Assessing the aesthetics of an image is challenging, as it is influenced by multiple factors including composition, color, style, and high-level semantics. Existing image aesthetic assessment (IAA) methods primarily rely on human-labeled rating scores, which oversimplify the visual aesthetic information that humans perceive. Conversely, user comments offer more comprehensive information and are a more natural way to express human opinions and preferences regarding image aesthetics. In light of this, we propose learning image aesthetics from user comments, and exploring vision-language pretraining methods to learn multimodal aesthetic representations. Specifically, we pretrain an image-text encoder-decoder model with image-comment pairs, using contrastive and generative objectives to learn rich and generic aesthetic semantics without human labels. To efficiently adapt the pretrained model for downstream IAA tasks, we further propose a lightweight rank-based adapter that employs text as an anchor to learn the aesthetic ranking concept. Our results show that our pretrained aesthetic vision-language model outperforms prior works on image aesthetic captioning over the AVA-Captions dataset, and it has powerful zero-shot capability for aesthetic tasks such as zero-shot style classification and zero-shot IAA, surpassing many supervised baselines. With only minimal finetuning parameters using the proposed adapter module, our model achieves state-of-the-art IAA performance over the AVA dataset.

Image cropping aims to find visually appealing crops in an image. Despite the great progress made by previous methods, they are weak in capturing the spatial relationship between crops and aesthetic elements (e.g., salient objects, semantic edges). Besides, due to the high annotation cost of labeled data, the potential of unlabeled data awaits to be excavated. To address the first issue, we propose spatial-aware feature to encode the spatial relationship between candidate crops and aesthetic elements, by feeding the concatenation of crop mask and selectively aggregated feature maps to a light-weighted encoder. To address the second issue, we train a pair-wise ranking classifier on labeled images and transfer such knowledge to unlabeled images to enforce rank consistency. Experimental results on the benchmark datasets show that our proposed method performs favorably against state-of-the-art methods.

Screen content images (SCIs) include many informative components, e.g., texts and graphics. Such content creates sharp edges or homogeneous areas, making a pixel distribution of SCI different from the natural image. Therefore, we need to properly handle the edges and textures to minimize information distortion of the contents when a display device’s resolution differs from SCIs. To achieve this goal, we propose an implicit neural representation using B-splines for screen content image super-resolution (SCI SR) with arbitrary scales. Our method extracts scaling, translating, and smoothing parameters of B-splines. The followed multi-layer perceptron (MLP) uses the estimated B-splines to recover high-resolution SCI. Our network outperforms both a transformer-based reconstruction and an implicit Fourier representation method in almost upscaling factor, thanks to the positive constraint and compact support of the B-spline basis. Moreover, our SR results are recognized as correct text letters with the highest confidence by a pre-trained scene text recognition network. Source code is available at https://github.com/ByeongHyunPak/btc.

GAN inversion and editing via StyleGAN maps an input image into the embedding spaces (W, W^+, and F) to simultaneously maintain image fidelity and meaningful manipulation. From latent space W to extended latent space W^+ to feature space F in StyleGAN, the editability of GAN inversion decreases while its reconstruction quality increases. Recent GAN inversion methods typically explore W^+ and F rather than W to improve reconstruction fidelity while maintaining editability. As W^+ and F are derived from W that is essentially the foundation latent space of StyleGAN, these GAN inversion methods focusing on W^+ and F spaces could be improved by stepping back to W. In this work, we propose to first obtain the proper latent code in foundation latent space W. We introduce contrastive learning to align W and the image space for proper latent code discovery. Then, we leverage a cross-attention encoder to transform the obtained latent code in W into W^+ and F, accordingly. Our experiments show that our exploration of the foundation latent space W improves the representation ability of latent codes in W^+ and features in F, which yields state-of-the-art reconstruction fidelity and editability results on the standard benchmarks. Project page: https://kumapowerliu.github.io/CLCAE.

Arbitrary style transfer has been demonstrated to be efficient in artistic image generation. Previous methods either globally modulate the content feature ignoring local details, or overly focus on the local structure details leading to style leakage. In contrast to the literature, we propose a new scheme “style kernel” that learns spatially adaptive kernel for per-pixel stylization, where the convolutional kernels are dynamically generated from the global style-content aligned feature and then the learned kernels are applied to modulate the content feature at each spatial position. This new scheme allows flexible both global and local interactions between the content and style features such that the wanted styles can be easily transferred to the content image while at the same time the content structure can be easily preserved. To further enhance the flexibility of our style transfer method, we propose a Style Alignment Encoding (SAE) module complemented with a Content-based Gating Modulation (CGM) module for learning the dynamic style kernels in focusing regions. Extensive experiments strongly demonstrate that our proposed method outperforms state-of-the-art methods and exhibits superior performance in terms of visual quality and efficiency.

Advances in representation learning have led to great success in understanding and generating data in various domains. However, in modeling vector graphics data, the pure data-driven approach often yields unsatisfactory results in downstream tasks as existing deep learning methods often require the quantization of SVG parameters and cannot exploit the geometric properties explicitly. In this paper, we propose a transformer-based representation learning model (SVGformer) that directly operates on continuous input values and manipulates the geometric information of SVG to encode outline details and long-distance dependencies. SVGfomer can be used for various downstream tasks: reconstruction, classification, interpolation, retrieval, etc. We have conducted extensive experiments on vector font and icon datasets to show that our model can capture high-quality representation information and outperform the previous state-of-the-art on downstream tasks significantly.

Blind text image super-resolution (SR) is challenging as one needs to cope with diverse font styles and unknown degradation. To address the problem, existing methods perform character recognition in parallel to regularize the SR task, either through a loss constraint or intermediate feature condition. Nonetheless, the high-level prior could still fail when encountering severe degradation. The problem is further compounded given characters of complex structures, e.g., Chinese characters that combine multiple pictographic or ideographic symbols into a single character. In this work, we present a novel prior that focuses more on the character structure. In particular, we learn to encapsulate rich and diverse structures in a StyleGAN and exploit such generative structure priors for restoration. To restrict the generative space of StyleGAN so that it obeys the structure of characters yet remains flexible in handling different font styles, we store the discrete features for each character in a {codebook}. The code subsequently drives the StyleGAN to generate high-resolution structural details to aid text SR. Compared to priors based on character recognition, the proposed structure prior exerts stronger character-specific guidance to restore faithful and precise strokes of a designated character. Extensive experiments on synthetic and real datasets demonstrate the compelling performance of the proposed generative structure prior in facilitating robust text SR. Our code is available at https://github.com/csxmli2016/MARCONet.

Layout is essential for graphic design and poster generation. Recently, applying deep learning models to generate layouts has attracted increasing attention. This paper focuses on using the GAN-based model conditioned on image contents to generate advertising poster graphic layouts, which requires an advertising poster layout dataset with paired product images and graphic layouts. However, the paired images and layouts in the existing dataset are collected by inpainting and annotating posters, respectively. There exists a domain gap between inpainted posters (source domain data) and clean product images (target domain data). Therefore, this paper combines unsupervised domain adaption techniques to design a GAN with a novel pixel-level discriminator (PD), called PDA-GAN, to generate graphic layouts according to image contents. The PD is connected to the shallow level feature map and computes the GAN loss for each input-image pixel. Both quantitative and qualitative evaluations demonstrate that PDA-GAN can achieve state-of-the-art performances and generate high-quality image-aware graphic layouts for advertising posters.

The recent success of text-to-image synthesis has taken the world by storm and captured the general public’s imagination. From a technical standpoint, it also marked a drastic change in the favored architecture to design generative image models. GANs used to be the de facto choice, with techniques like StyleGAN. With DALL-E 2, auto-regressive and diffusion models became the new standard for large-scale generative models overnight. This rapid shift raises a fundamental question: can we scale up GANs to benefit from large datasets like LAION? We find that naively increasing the capacity of the StyleGAN architecture quickly becomes unstable. We introduce GigaGAN, a new GAN architecture that far exceeds this limit, demonstrating GANs as a viable option for text-to-image synthesis. GigaGAN offers three major advantages. First, it is orders of magnitude faster at inference time, taking only 0.13 seconds to synthesize a 512px image. Second, it can synthesize high-resolution images, for example, 16-megapixel images in 3.66 seconds. Finally, GigaGAN supports various latent space editing applications such as latent interpolation, style mixing, and vector arithmetic operations.

Recent progress in diffusion models has revolutionized the popular technology of text-to-image generation. While existing approaches could produce photorealistic high-resolution images with text conditions, there are still several open problems to be solved, which limits the further improvement of image fidelity and text relevancy. In this paper, we propose ERNIE-ViLG 2.0, a large-scale Chinese text-to-image diffusion model, to progressively upgrade the quality of generated images by: (1) incorporating fine-grained textual and visual knowledge of key elements in the scene, and (2) utilizing different denoising experts at different denoising stages. With the proposed mechanisms, ERNIE-ViLG 2.0 not only achieves a new state-of-the-art on MS-COCO with zero-shot FID-30k score of 6.75, but also significantly outperforms recent models in terms of image fidelity and image-text alignment, with side-by-side human evaluation on the bilingual prompt set ViLG-300.

The artistic style within a painting is the means of expression, which includes not only the painting material, colors, and brushstrokes, but also the high-level attributes, including semantic elements and object shapes. Previous arbitrary example-guided artistic image generation methods often fail to control shape changes or convey elements. Pre-trained text-to-image synthesis diffusion probabilistic models have achieved remarkable quality but often require extensive textual descriptions to accurately portray the attributes of a particular painting.The uniqueness of an artwork lies in the fact that it cannot be adequately explained with normal language. Our key idea is to learn the artistic style directly from a single painting and then guide the synthesis without providing complex textual descriptions. Specifically, we perceive style as a learnable textual description of a painting.We propose an inversion-based style transfer method (InST), which can efficiently and accurately learn the key information of an image, thus capturing and transferring the artistic style of a painting. We demonstrate the quality and efficiency of our method on numerous paintings of various artists and styles. Codes are available at https://github.com/zyxElsa/InST.

We present Corgi, a novel method for text-to-image generation. Corgi is based on our proposed shifted diffusion model, which achieves better image embedding generation from input text. Different from the baseline diffusion model used in DALL-E 2, our method seamlessly encodes prior knowledge of the pre-trained CLIP model in its diffusion process by designing a new initialization distribution and a new transition step of the diffusion. Compared to the strong DALL-E 2 baseline, our method performs better in generating image embedding from the text in terms of both efficiency and effectiveness, which consequently results in better text-to-image generation. Extensive large-scale experiments are conducted and evaluated in terms of both quantitative measures and human evaluation, indicating a stronger generation ability of our method compared to existing ones. Furthermore, our model enables semi-supervised and language-free training for text-to-image generation, where only part or none of the images in the training dataset have an associated caption. Trained with only 1.7% of the images being captioned, our semi-supervised model obtains FID results comparable to DALL-E 2 on zero-shot text-to-image generation evaluated on MS-COCO. Corgi also achieves new state-of-the-art results across different datasets on downstream language-free text-to-image generation tasks, outperforming the previous method, Lafite, by a large margin.

Controllable layout generation aims at synthesizing plausible arrangement of element bounding boxes with optional constraints, such as type or position of a specific element. In this work, we try to solve a broad range of layout generation tasks in a single model that is based on discrete state-space diffusion models. Our model, named LayoutDM, naturally handles the structured layout data in the discrete representation and learns to progressively infer a noiseless layout from the initial input, where we model the layout corruption process by modality-wise discrete diffusion. For conditional generation, we propose to inject layout constraints in the form of masking or logit adjustment during inference. We show in the experiments that our LayoutDM successfully generates high-quality layouts and outperforms both task-specific and task-agnostic baselines on several layout tasks.

Unpaired image-to-image translation aims to learn proper mappings that can map images from one domain to another domain while preserving the content of the input image. However, with large enough capacities, the network can learn to map the inputs to any random permutation of images in another domain. Existing methods treat two domains as discrete and propose different assumptions to address this problem. In this paper, we start from a different perspective and consider the paths connecting the two domains. We assume that the optimal path length between the input and output image should be the shortest among all possible paths. Based on this assumption, we propose a new method to allow generating images along the path and present a simple way to encourage the network to find the shortest path without pair information. Extensive experiments on various tasks demonstrate the superiority of our approach.

We present DiffCollage, a compositional diffusion model that can generate large content by leveraging diffusion models trained on generating pieces of the large content. Our approach is based on a factor graph representation where each factor node represents a portion of the content and a variable node represents their overlap. This representation allows us to aggregate intermediate outputs from diffusion models defined on individual nodes to generate content of arbitrary size and shape in parallel without resorting to an autoregressive generation procedure. We apply DiffCollage to various tasks, including infinite image generation, panorama image generation, and long-duration text-guided motion generation. Extensive experimental results with a comparison to strong autoregressive baselines verify the effectiveness of our approach.

Diffusion models are rising as a powerful solution for high-fidelity image generation, which exceeds GANs in quality in many circumstances. However, their slow training and inference speed is a huge bottleneck, blocking them from being used in real-time applications. A recent DiffusionGAN method significantly decreases the models’ running time by reducing the number of sampling steps from thousands to several, but their speeds still largely lag behind the GAN counterparts. This paper aims to reduce the speed gap by proposing a novel wavelet-based diffusion scheme. We extract low-and-high frequency components from both image and feature levels via wavelet decomposition and adaptively handle these components for faster processing while maintaining good generation quality. Furthermore, we propose to use a reconstruction term, which effectively boosts the model training convergence. Experimental results on CelebA-HQ, CIFAR-10, LSUN-Church, and STL-10 datasets prove our solution is a stepping-stone to offering real-time and high-fidelity diffusion models. Our code and pre-trained checkpoints are available at https://github.com/VinAIResearch/WaveDiff.git.

A diffusion probabilistic model (DPM), which constructs a forward diffusion process by gradually adding noise to data points and learns the reverse denoising process to generate new samples, has been shown to handle complex data distribution. Despite its recent success in image synthesis, applying DPMs to video generation is still challenging due to high-dimensional data spaces. Previous methods usually adopt a standard diffusion process, where frames in the same video clip are destroyed with independent noises, ignoring the content redundancy and temporal correlation. This work presents a decomposed diffusion process via resolving the per-frame noise into a base noise that is shared among all frames and a residual noise that varies along the time axis. The denoising pipeline employs two jointly-learned networks to match the noise decomposition accordingly. Experiments on various datasets confirm that our approach, termed as VideoFusion, surpasses both GAN-based and diffusion-based alternatives in high-quality video generation. We further show that our decomposed formulation can benefit from pre-trained image diffusion models and well-support text-conditioned video creation.

We propose the first joint audio-video generation framework that brings engaging watching and listening experiences simultaneously, towards high-quality realistic videos. To generate joint audio-video pairs, we propose a novel Multi-Modal Diffusion model (i.e., MM-Diffusion), with two-coupled denoising autoencoders. In contrast to existing single-modal diffusion models, MM-Diffusion consists of a sequential multi-modal U-Net for a joint denoising process by design. Two subnets for audio and video learn to gradually generate aligned audio-video pairs from Gaussian noises. To ensure semantic consistency across modalities, we propose a novel random-shift based attention block bridging over the two subnets, which enables efficient cross-modal alignment, and thus reinforces the audio-video fidelity for each other. Extensive experiments show superior results in unconditional audio-video generation, and zero-shot conditional tasks (e.g., video-to-audio). In particular, we achieve the best FVD and FAD on Landscape and AIST++ dancing datasets. Turing tests of 10k votes further demonstrate dominant preferences for our model.

The most recent efforts in video matting have focused on eliminating trimap dependency since trimap annotations are expensive and trimap-based methods are less adaptable for real-time applications. Despite the latest tripmap-free methods showing promising results, their performance often degrades when dealing with highly diverse and unstructured videos. We address this limitation by introducing Adaptive Matting for Dynamic Videos, termed AdaM, which is a framework designed for simultaneously differentiating foregrounds from backgrounds and capturing alpha matte details of human subjects in the foreground. Two interconnected network designs are employed to achieve this goal: (1) an encoder-decoder network that produces alpha mattes and intermediate masks which are used to guide the transformer in adaptively decoding foregrounds and backgrounds, and (2) a transformer network in which long- and short-term attention combine to retain spatial and temporal contexts, facilitating the decoding of foreground details. We benchmark and study our methods on recently introduced datasets, showing that our model notably improves matting realism and temporal coherence in complex real-world videos and achieves new best-in-class generalizability. Further details and examples are available at https://github.com/microsoft/AdaM.

Recently, numerous end-to-end optimized image compression neural networks have been developed and proved themselves as leaders in rate-distortion performance. The main strength of these learnt compression methods is in powerful nonlinear analysis and synthesis transforms that can be facilitated by deep neural networks. However, out of operational expediency, most of these end-to-end methods adopt uniform scalar quantizers rather than vector quantizers, which are information-theoretically optimal. In this paper, we present a novel Lattice Vector Quantization scheme coupled with a spatially Adaptive Companding (LVQAC) mapping. LVQ can better exploit the inter-feature dependencies than scalar uniform quantization while being computationally almost as simple as the latter. Moreover, to improve the adaptability of LVQ to source statistics, we couple a spatially adaptive companding (AC) mapping with LVQ. The resulting LVQAC design can be easily embedded into any end-to-end optimized image compression system. Extensive experiments demonstrate that for any end-to-end CNN image compression models, replacing uniform quantizer by LVQAC achieves better rate-distortion performance without significantly increasing the model complexity.

Typical video compression systems consist of two main modules: motion coding and residual coding. This general architecture is adopted by classical coding schemes (such as international standards H.265 and H.266) and deep learning-based coding schemes. We propose a novel B-frame coding architecture based on two-layer Conditional Augmented Normalization Flows (CANF). It has the striking feature of not transmitting any motion information. Our proposed idea of video compression without motion coding offers a new direction for learned video coding. Our base layer is a low-resolution image compressor that replaces the full-resolution motion compressor. The low-resolution coded image is merged with the warped high-resolution images to generate a high-quality image as a conditioning signal for the enhancement-layer image coding in full resolution. One advantage of this architecture is significantly reduced computational complexity due to eliminating the motion information compressor. In addition, we adopt a skip-mode coding technique to reduce the transmitted latent samples. The rate-distortion performance of our scheme is slightly lower than that of the state-of-the-art learned B-frame coding scheme, B-CANF, but outperforms other learned B-frame coding schemes. However, compared to B-CANF, our scheme saves 45% of multiply-accumulate operations (MACs) for encoding and 27% of MACs for decoding. The code is available at https://nycu-clab.github.io.

As deep convolutional neural networks (DNNs) are widely used in various fields of computer vision, leveraging the overfitting ability of the DNN to achieve video resolution upscaling has become a new trend in the modern video delivery system. By dividing videos into chunks and overfitting each chunk with a super-resolution model, the server encodes videos before transmitting them to the clients, thus achieving better video quality and transmission efficiency. However, a large number of chunks are expected to ensure good overfitting quality, which substantially increases the storage and consumes more bandwidth resources for data transmission. On the other hand, decreasing the number of chunks through training optimization techniques usually requires high model capacity, which significantly slows down execution speed. To reconcile such, we propose a novel method for high-quality and efficient video resolution upscaling tasks, which leverages the spatial-temporal information to accurately divide video into chunks, thus keeping the number of chunks as well as the model size to a minimum. Additionally, we advance our method into a single overfitting model by a data-aware joint training technique, which further reduces the storage requirement with negligible quality drop. We deploy our proposed overfitting models on an off-the-shelf mobile phone, and experimental results show that our method achieves real-time video super-resolution with high video quality. Compared with the state-of-the-art, our method achieves 28 fps streaming speed with 41.60 PSNR, which is 14 times faster and 2.29 dB better in the live video resolution upscaling tasks.

Implicit neural representations store videos as neural networks and have performed well for vision tasks such as video compression and denoising. With frame index and/or positional index as input, implicit representations (NeRV, E-NeRV, etc.) reconstruct video frames from fixed and content-agnostic embeddings. Such embedding largely limits the regression capacity and internal generalization for video interpolation. In this paper, we propose a Hybrid Neural Representation for Videos (HNeRV), where learnable and content-adaptive embeddings act as decoder input. Besides the input embedding, we introduce a HNeRV block to make model parameters evenly distributed across the entire network, therefore higher layers (layers near the output) can have more capacity to store high-resolution content and video details. With content-adaptive embedding and re-designed model architecture, HNeRV outperforms implicit methods (NeRV, E-NeRV) in video regression task for both reconstruction quality and convergence speed, and shows better internal generalization. As a simple and efficient video representation, HNeRV also shows decoding advantages for speed, flexibility, and deployment, compared to traditional codecs (H.264, H.265) and learning-based compression methods. Finally, we explore the effectiveness of HNeRV on downstream tasks such as video compression and video inpainting.

This paper proposes a regularizer called Implicit Neural Representation Regularizer (INRR) to improve the generalization ability of the Implicit Neural Representation (INR). The INR is a fully connected network that can represent signals with details not restricted by grid resolution. However, its generalization ability could be improved, especially with non-uniformly sampled data. The proposed INRR is based on learned Dirichlet Energy (DE) that measures similarities between rows/columns of the matrix. The smoothness of the Laplacian matrix is further integrated by parameterizing DE with a tiny INR. INRR improves the generalization of INR in signal representation by perfectly integrating the signal’s self-similarity with the smoothness of the Laplacian matrix. Through well-designed numerical experiments, the paper also reveals a series of properties derived from INRR, including momentum methods like convergence trajectory and multi-scale similarity. Moreover, the proposed method could improve the performance of other signal representation methods.

Continuous convolution has recently gained prominence due to its ability to handle irregularly sampled data and model long-term dependency. Also, the promising experimental results of using large convolutional kernels have catalyzed the development of continuous convolution since they can construct large kernels very efficiently. Leveraging neural networks, more specifically multilayer perceptrons (MLPs), is by far the most prevalent approach to implementing continuous convolution. However, there are a few drawbacks, such as high computational costs, complex hyperparameter tuning, and limited descriptive power of filters. This paper suggests an alternative approach to building a continuous convolution without neural networks, resulting in more computationally efficient and improved performance. We present self-moving point representations where weight parameters freely move, and interpolation schemes are used to implement continuous functions. When applied to construct convolutional kernels, the experimental results have shown improved performance with drop-in replacement in the existing frameworks. Due to its lightweight structure, we are first to demonstrate the effectiveness of continuous convolution in a large-scale setting, e.g., ImageNet, presenting the improvements over the prior arts. Our code is available on https://github.com/sangnekim/SMPConv

Inspired by the success of recent vision transformers and large kernel design in convolutional neural networks (CNNs), in this paper, we analyze and explore essential reasons for their success. We claim two factors that are critical for 3D large-scale scene understanding: a larger receptive field and operations with greater non-linearity. The former is responsible for providing long range contexts and the latter can enhance the capacity of the network. To achieve the above properties, we propose a simple yet effective long range pooling (LRP) module using dilation max pooling, which provides a network with a large adaptive receptive field. LRP has few parameters, and can be readily added to current CNNs. Also, based on LRP, we present an entire network architecture, LRPNet, for 3D understanding. Ablation studies are presented to support our claims, and show that the LRP module achieves better results than large kernel convolution yet with reduced computation, due to its non-linearity. We also demonstrate the superiority of LRPNet on various benchmarks: LRPNet performs the best on ScanNet and surpasses other CNN-based methods on S3DIS and Matterport3D. Code will be avalible at https://github.com/li-xl/LRPNet.

Single domain generalization aims to train a generalizable model with only one source domain to perform well on arbitrary unseen target domains. Image augmentation based on Random Convolutions (RandConv), consisting of one convolution layer randomly initialized for each mini-batch, enables the model to learn generalizable visual representations by distorting local textures despite its simple and lightweight structure. However, RandConv has structural limitations in that the generated image easily loses semantics as the kernel size increases, and lacks the inherent diversity of a single convolution operation. To solve the problem, we propose a Progressive Random Convolution (Pro-RandConv) method that recursively stacks random convolution layers with a small kernel size instead of increasing the kernel size. This progressive approach can not only mitigate semantic distortions by reducing the influence of pixels away from the center in the theoretical receptive field, but also create more effective virtual domains by gradually increasing the style diversity. In addition, we develop a basic random convolution layer into a random convolution block including deformable offsets and affine transformation to support texture and contrast diversification, both of which are also randomly initialized. Without complex generators or adversarial learning, we demonstrate that our simple yet effective augmentation strategy outperforms state-of-the-art methods on single domain generalization benchmarks.

As the core building block of vision transformers, attention is a powerful tool to capture long-range dependency. However, such power comes at a cost: it incurs a huge computation burden and heavy memory footprint as pairwise token interaction across all spatial locations is computed. A series of works attempt to alleviate this problem by introducing handcrafted and content-agnostic sparsity into attention, such as restricting the attention operation to be inside local windows, axial stripes, or dilated windows. In contrast to these approaches, we propose a novel dynamic sparse attention via bi-level routing to enable a more flexible allocation of computations with content awareness. Specifically, for a query, irrelevant key-value pairs are first filtered out at a coarse region level, and then fine-grained token-to-token attention is applied in the union of remaining candidate regions (i.e., routed regions). We provide a simple yet effective implementation of the proposed bi-level routing attention, which utilizes the sparsity to save both computation and memory while involving only GPU-friendly dense matrix multiplications. Built with the proposed bi-level routing attention, a new general vision transformer, named BiFormer, is then presented. As BiFormer attends to a small subset of relevant tokens in a query-adaptive manner without distraction from other irrelevant ones, it enjoys both good performance and high computational efficiency, especially in dense prediction tasks. Empirical results across several computer vision tasks such as image classification, object detection, and semantic segmentation verify the effectiveness of our design. Code is available at https://github.com/rayleizhu/BiFormer.

Vision transformers have achieved significant improvements on various vision tasks but their quadratic interactions between tokens significantly reduce computational efficiency. Many pruning methods have been proposed to remove redundant tokens for efficient vision transformers recently. However, existing studies mainly focus on the token importance to preserve local attentive tokens but completely ignore the global token diversity. In this paper, we emphasize the cruciality of diverse global semantics and propose an efficient token decoupling and merging method that can jointly consider the token importance and diversity for token pruning. According to the class token attention, we decouple the attentive and inattentive tokens. In addition to preserve the most discriminative local tokens, we merge similar inattentive tokens and match homogeneous attentive tokens to maximize the token diversity. Despite its simplicity, our method obtains a promising trade-off between model complexity and classification accuracy. On DeiT-S, our method reduces the FLOPs by 35% with only a 0.2% accuracy drop. Notably, benefiting from maintaining the token diversity, our method can even improve the accuracy of DeiT-T by 0.1% after reducing its FLOPs by 40%.

Recently, whether and how cutting-edge Neuroscience findings can inspire Artificial Intelligence (AI) confuse both communities and draw much discussion. As one of the most critical fields in AI, Computer Vision (CV) also pays much attention to the discussion. To show our ideas and experimental evidence to the discussion, we focus on one of the most broadly researched topics both in Neuroscience and CV fields, i.e., Face Recognition (FR). Neuroscience studies show that face attributes are essential to the human face-recognizing system. How the attributes contribute also be explained by the Neuroscience community. Even though a few CV works improved the FR performance with attribute enhancement, none of them are inspired by the human face-recognizing mechanism nor boosted performance significantly. To show our idea experimentally, we model the biological characteristics of the human face-recognizing system with classical Convolutional Neural Network Operators (CNN Ops) purposely. We name the proposed Biologically-inspired Network as BioNet. Our BioNet consists of two cascade sub-networks, i.e., the Visual Cortex Network (VCN) and the Inferotemporal Cortex Network (ICN). The VCN is modeled with a classical CNN backbone. The proposed ICN comprises three biologically-inspired modules, i.e., the Cortex Functional Compartmentalization, the Compartment Response Transform, and the Response Intensity Modulation. The experiments prove that: 1) The cutting-edge findings about the human face-recognizing system can further boost the CNN-based FR network. 2) With the biological mechanism, both identity-related attributes (e.g., gender) and identity-unrelated attributes (e.g., expression) can benefit the deep FR models. Surprisingly, the identity-unrelated ones contribute even more than the identity-related ones. 3) The proposed BioNet significantly boosts state-of-the-art on standard FR benchmark datasets. For example, BioNet boosts IJB-B@1e-6 from 52.12% to 68.28% and MegaFace from 98.74% to 99.19%. The source code will be released.

The remote photoplethysmography (rPPG) technique can estimate pulse-related metrics (e.g. heart rate and respiratory rate) from facial videos and has a high potential for health monitoring. The latest deep rPPG methods can model in-distribution noise due to head motion, video compression, etc., and estimate high-quality rPPG signals under similar scenarios. However, deep rPPG models may not generalize well to the target test domain with unseen noise and distortions. In this paper, to improve the generalization ability of rPPG models, we propose a dual-bridging network to reduce the domain discrepancy by aligning intermediate domains and synthesizing the target noise in the source domain for better noise reduction. To comprehensively explore the target domain noise, we propose a novel adversarial noise generation in which the noise generator indirectly competes with the noise reducer. To further improve the robustness of the noise reducer, we propose hard noise pattern mining to encourage the generator to learn hard noise patterns contained in the target domain features. We evaluated the proposed method on three public datasets with different types of interferences. Under different cross-domain scenarios, the comprehensive results show the effectiveness of our method.

Scaling properties have been one of the central issues in self-supervised pre-training, especially the data scalability, which has successfully motivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capabilities. However, scaling properties seem to be unintentionally neglected in the recent trending studies on masked image modeling (MIM), and some arguments even suggest that MIM cannot benefit from large-scale data. In this work, we try to break down these preconceptions and systematically study the scaling behaviors of MIM through extensive experiments, with data ranging from 10% of ImageNet-1K to full ImageNet-22K, model parameters ranging from 49-million to one-billion, and training length ranging from 125K to 500K iterations. And our main findings can be summarized in two folds: 1) masked image modeling remains demanding large-scale data in order to scale up computes and model parameters; 2) masked image modeling cannot benefit from more data under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-trained language models or supervised pre-trained vision models. In addition, we reveal several intriguing properties in MIM, such as high sample efficiency in large MIM models and strong correlation between pre-training validation loss and transfer performance. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models. Code and models will be available at https://github.com/microsoft/SimMIM.

Masked image modeling (MIM) has attracted much research attention due to its promising potential for learning scalable visual representations. In typical approaches, models usually focus on predicting specific contents of masked patches, and their performances are highly related to pre-defined mask strategies. Intuitively, this procedure can be considered as training a student (the model) on solving given problems (predict masked patches). However, we argue that the model should not only focus on solving given problems, but also stand in the shoes of a teacher to produce a more challenging problem by itself. To this end, we propose Hard Patches Mining (HPM), a brand-new framework for MIM pre-training. We observe that the reconstruction loss can naturally be the metric of the difficulty of the pre-training task. Therefore, we introduce an auxiliary loss predictor, predicting patch-wise losses first and deciding where to mask next. It adopts a relative relationship learning strategy to prevent overfitting to exact reconstruction loss values. Experiments under various settings demonstrate the effectiveness of HPM in constructing masked images. Furthermore, we empirically find that solely introducing the loss prediction objective leads to powerful representations, verifying the efficacy of the ability to be aware of where is hard to reconstruct.

Existing Masked Image Modeling methods apply fixed mask patterns to guide the self-supervised training. As those patterns resort to different criteria to mask local regions, sticking to a fixed pattern leads to limited vision cues modeling capability. This paper proposes an evolved part-based masking to pursue more general visual cues modeling in self-supervised learning. Our method is based on an adaptive part partition module, which leverages the vision model being trained to construct a part graph, and partitions parts with graph cut. The accuracy of partitioned parts is on par with the capability of the pre-trained model, leading to evolved mask patterns at different training stages. It generates simple patterns at the initial training stage to learn low-level visual cues, which hence evolves to eliminate accurate object parts to reinforce the learning of object semantics and contexts. Our method does not require extra pre-trained models or annotations, and effectively ensures the training efficiency by evolving the training difficulty. Experiment results show that it substantially boosts the performance on various tasks including image classification, object detection, and semantic segmentation. For example, it outperforms the recent MAE by 0.69% on imageNet-1K classification and 1.61% on ADE20K segmentation with the same training epochs.

Graph is a highly generic and diverse representation, suitable for almost any data processing problem. Spectral graph theory has been shown to provide powerful algorithms, backed by solid linear algebra theory. It thus can be extremely instrumental to design deep network building blocks with spectral graph characteristics. For instance, such a network allows the design of optimal graphs for certain tasks or obtaining a canonical orthogonal low-dimensional embedding of the data. Recent attempts to solve this problem were based on minimizing Rayleigh-quotient type losses. We propose a different approach of directly learning the graph’s eigensapce. A severe problem of the direct approach, applied in batch-learning, is the inconsistent mapping of features to eigenspace coordinates in different batches. We analyze the degrees of freedom of learning this task using batches and propose a stable alignment mechanism that can work both with batch changes and with graph-metric changes. We show that our learnt spectral embedding is better in terms of NMI, ACC, Grassman distnace, orthogonality and classification accuracy, compared to SOTA. In addition, the learning is more stable.

Transformer-based architectures have become competitive across a variety of visual domains, most notably images and videos. While prior work studies these modalities in isolation, having a common architecture suggests that one can train a single unified model for multiple visual modalities. Prior attempts at unified modeling typically use architectures tailored for vision tasks, or obtain worse performance compared to single modality models. In this work, we show that masked autoencoding can be used to train a simple Vision Transformer on images and videos, without requiring any labeled data. This single model learns visual representations that are comparable to or better than single-modality representations on both image and video benchmarks, while using a much simpler architecture. Furthermore, this model can be learned by dropping 90% of the image and 95% of the video patches, enabling extremely fast training of huge model architectures. In particular, we show that our single ViT-Huge model can be finetuned to achieve 86.6% on ImageNet and 75.5% on the challenging Something Something-v2 video benchmark, setting a new state-of-the-art.

In this paper we introduce the Temporo-Spatial Vision Transformer (TSViT), a fully-attentional model for general Satellite Image Time Series (SITS) processing based on the Vision Transformer (ViT). TSViT splits a SITS record into non-overlapping patches in space and time which are tokenized and subsequently processed by a factorized temporo-spatial encoder. We argue, that in contrast to natural images, a temporal-then-spatial factorization is more intuitive for SITS processing and present experimental evidence for this claim. Additionally, we enhance the model’s discriminative power by introducing two novel mechanisms for acquisition-time-specific temporal positional encodings and multiple learnable class tokens. The effect of all novel design choices is evaluated through an extensive ablation study. Our proposed architecture achieves state-of-the-art performance, surpassing previous approaches by a significant margin in three publicly available SITS semantic segmentation and classification datasets. All model, training and evaluation codes can be found at https://github.com/michaeltrs/DeepSatModels.

The quality of scene graphs generated by the state-of-the-art (SOTA) models is compromised due to the long-tail nature of the relationships and their parent object pairs. Training of the scene graphs is dominated by the majority relationships of the majority pairs and, therefore, the object-conditional distributions of relationship in the minority pairs are not preserved after the training is converged. Consequently, the biased model performs well on more frequent relationships in the marginal distribution of relationships such as ‘on’ and ‘wearing’, and performs poorly on the less frequent relationships such as ‘eating’ or ‘hanging from’. In this work, we propose virtual evidence incorporated within-triplet Bayesian Network (BN) to preserve the object-conditional distribution of the relationship label and to eradicate the bias created by the marginal probability of the relationships. The insufficient number of relationships in the minority classes poses a significant problem in learning the within-triplet Bayesian network. We address this insufficiency by embedding-based augmentation of triplets where we borrow samples of the minority triplet classes from its neighboring triplets in the semantic space. We perform experiments on two different datasets and achieve a significant improvement in the mean recall of the relationships. We also achieve a better balance between recall and mean recall performance compared to the SOTA de-biasing techniques of scene graph models.

Many videos contain flickering artifacts; common causes of flicker include video processing algorithms, video generation algorithms, and capturing videos under specific situations. Prior work usually requires specific guidance such as the flickering frequency, manual annotations, or extra consistent videos to remove the flicker. In this work, we propose a general flicker removal framework that only receives a single flickering video as input without additional guidance. Since it is blind to a specific flickering type or guidance, we name this “blind deflickering.” The core of our approach is utilizing the neural atlas in cooperation with a neural filtering strategy. The neural atlas is a unified representation for all frames in a video that provides temporal consistency guidance but is flawed in many cases. To this end, a neural network is trained to mimic a filter to learn the consistent features (e.g., color, brightness) and avoid introducing the artifacts in the atlas. To validate our method, we construct a dataset that contains diverse real-world flickering videos. Extensive experiments show that our method achieves satisfying deflickering performance and even outperforms baselines that use extra guidance on a public benchmark. The source code is publicly available at https://chenyanglei.github.io/deflicker.

Shadows in videos are difficult to detect because of the large shadow deformation between frames. In this work, we argue that accounting for shadow deformation is essential when designing a video shadow detection method. To this end, we introduce the shadow deformation attention trajectory (SODA), a new type of video self-attention module, specially designed to handle the large shadow deformations in videos. Moreover, we present a new shadow contrastive learning mechanism (SCOTCH) which aims at guiding the network to learn a unified shadow representation from massive positive shadow pairs across different videos. We demonstrate empirically the effectiveness of our two contributions in an ablation study. Furthermore, we show that SCOTCH and SODA significantly outperforms existing techniques for video shadow detection. Code is available at the project page: https://lihaoliu-cambridge.github.io/scotchandsoda/

We introduce the MAsked Generative VIdeo Transformer, MAGVIT, to tackle various video synthesis tasks with a single model. We introduce a 3D tokenizer to quantize a video into spatial-temporal visual tokens and propose an embedding method for masked video token modeling to facilitate multi-task learning. We conduct extensive experiments to demonstrate the quality, efficiency, and flexibility of MAGVIT. Our experiments show that (i) MAGVIT performs favorably against state-of-the-art approaches and establishes the best-published FVD on three video generation benchmarks, including the challenging Kinetics-600. (ii) MAGVIT outperforms existing methods in inference time by two orders of magnitude against diffusion models and by 60x against autoregressive models. (iii) A single MAGVIT model supports ten diverse generation tasks and generalizes across videos from different visual domains. The source code and trained models will be released to the public at https://magvit.cs.cmu.edu.

Semantic segmentation involves classifying each pixel into one of a pre-defined set of object/stuff classes. Such a fine-grained detection and localization of objects in the scene is challenging by itself. The complexity increases manifold in the presence of blur. With cameras becoming increasingly light-weight and compact, blur caused by motion during capture time has become unavoidable. Most research has focused on improving segmentation performance for sharp clean images and the few works that deal with degradations, consider motion-blur as one of many generic degradations. In this work, we focus exclusively on motion-blur and attempt to achieve robustness for semantic segmentation in its presence. Based on the observation that segmentation annotations can be used to generate synthetic space-variant blur, we propose a Class-Centric Motion-Blur Augmentation (CCMBA) strategy. Our approach involves randomly selecting a subset of semantic classes present in the image and using the segmentation map annotations to blur only the corresponding regions. This enables the network to simultaneously learn semantic segmentation for clean images, images with egomotion blur, as well as images with dynamic scene blur. We demonstrate the effectiveness of our approach for both CNN and Vision Transformer-based semantic segmentation networks on PASCAL VOC and Cityscapes datasets. We also illustrate the improved generalizability of our method to complex real-world blur by evaluating on the commonly used deblurring datasets GoPro and REDS.

This paper tackles the problem of semi-supervised video object segmentation on resource-constrained devices, such as mobile phones. We formulate this problem as a distillation task, whereby we demonstrate that small space-time-memory networks with finite memory can achieve competitive results with state of the art, but at a fraction of the computational cost (32 milliseconds per frame on a Samsung Galaxy S22). Specifically, we provide a theoretically grounded framework that unifies knowledge distillation with supervised contrastive representation learning. These models are able to jointly benefit from both pixel-wise contrastive learning and distillation from a pre-trained teacher. We validate this loss by achieving competitive J&F to state of the art on both the standard DAVIS and YouTube benchmarks, despite running up to ×5 faster, and with ×32 fewer parameters.

Manipulated videos often contain subtle inconsistencies between their visual and audio signals. We propose a video forensics method, based on anomaly detection, that can identify these inconsistencies, and that can be trained solely using real, unlabeled data. We train an autoregressive model to generate sequences of audio-visual features, using feature sets that capture the temporal synchronization between video frames and sound. At test time, we then flag videos that the model assigns low probability. Despite being trained entirely on real videos, our model obtains strong performance on the task of detecting manipulated speech videos. Project site: https://cfeng16.github.io/audio-visual-forensics.

Existing video recognition algorithms always conduct different training pipelines for inputs with different frame numbers, which requires repetitive training operations and multiplying storage costs. If we evaluate the model using other frames which are not used in training, we observe the performance will drop significantly (see Fig.1, which is summarized as Temporal Frequency Deviation phenomenon. To fix this issue, we propose a general framework, named Frame Flexible Network (FFN), which not only enables the model to be evaluated at different frames to adjust its computation, but also reduces the memory costs of storing multiple models significantly. Concretely, FFN integrates several sets of training sequences, involves Multi-Frequency Alignment (MFAL) to learn temporal frequency invariant representations, and leverages Multi-Frequency Adaptation (MFAD) to further strengthen the representation abilities. Comprehensive empirical validations using various architectures and popular benchmarks solidly demonstrate the effectiveness and generalization of FFN (e.g., 7.08/5.15/2.17% performance gain at Frame 4/8/16 on Something-Something V1 dataset over Uniformer). Code is available at https://github.com/BeSpontaneous/FFN.

Recent years have witnessed great progress in deep neural networks for real-time applications. However, most existing works do not explicitly consider the general case where the device’s state and the available resources fluctuate over time, and none of them investigate or address the impact of varying computational resources for online video understanding tasks. This paper proposes a System-status-aware Adaptive Network (SAN) that considers the device’s real-time state to provide high-quality predictions with low delay. Usage of our agent’s policy improves efficiency and robustness to fluctuations of the system status. On two widely used video understanding tasks, SAN obtains state-of-the-art performance while constantly keeping processing delays low. Moreover, training such an agent on various types of hardware configurations is not easy as the labeled training data might not be available, or can be computationally prohibitive. To address this challenging problem, we propose a Meta Self-supervised Adaptation (MSA) method that adapts the agent’s policy to new hardware configurations at test-time, allowing for easy deployment of the model onto other unseen hardware platforms.

While impressive progress has been achieved, video instance segmentation (VIS) methods with per-clip input often fail on challenging videos with occluded objects and crowded scenes. This is mainly because instance queries in these methods cannot encode well the discriminative embeddings of instances, making the query-based segmenter difficult to distinguish those ‘hard’ instances. To address these issues, we propose to mine discriminative query embeddings (MDQE) to segment occluded instances on challenging videos. First, we initialize the positional embeddings and content features of object queries by considering their spatial contextual information and the inter-frame object motion. Second, we propose an inter-instance mask repulsion loss to distance each instance from its nearby non-target instances. The proposed MDQE is the first VIS method with per-clip input that achieves state-of-the-art results on challenging videos and competitive performance on simple videos. In specific, MDQE with ResNet50 achieves 33.0% and 44.5% mask AP on OVIS and YouTube-VIS 2021, respectively. Code of MDQE can be found at https://github.com/MinghanLi/MDQE_CVPR2023.

Unsupervised Domain Adaptation (UDA) of semantic segmentation transfers labeled source knowledge to an unlabeled target domain by relying on accessing both the source and target data. However, the access to source data is often restricted or infeasible in real-world scenarios. Under the source data restrictive circumstances, UDA is less practical. To address this, recent works have explored solutions under the Source-Free Domain Adaptation (SFDA) setup, which aims to adapt a source-trained model to the target domain without accessing source data. Still, existing SFDA approaches use only image-level information for adaptation, making them sub-optimal in video applications. This paper studies SFDA for Video Semantic Segmentation (VSS), where temporal information is leveraged to address video adaptation. Specifically, we propose Spatio-Temporal Pixel-Level (STPL) contrastive learning, a novel method that takes full advantage of spatio-temporal information to tackle the absence of source data better. STPL explicitly learns semantic correlations among pixels in the spatio-temporal space, providing strong self-supervision for adaptation to the unlabeled target domain. Extensive experiments show that STPL achieves state-of-the-art performance on VSS benchmarks compared to current UDA and SFDA approaches. Code is available at: https://github.com/shaoyuanlo/STPL

Animating virtual avatars to make co-speech gestures facilitates various applications in human-machine interaction. The existing methods mainly rely on generative adversarial networks (GANs), which typically suffer from notorious mode collapse and unstable training, thus making it difficult to learn accurate audio-gesture joint distributions. In this work, we propose a novel diffusion-based framework, named Diffusion Co-Speech Gesture (DiffGesture), to effectively capture the cross-modal audio-to-gesture associations and preserve temporal coherence for high-fidelity audio-driven co-speech gesture generation. Specifically, we first establish the diffusion-conditional generation process on clips of skeleton sequences and audio to enable the whole framework. Then, a novel Diffusion Audio-Gesture Transformer is devised to better attend to the information from multiple modalities and model the long-term temporal dependency. Moreover, to eliminate temporal inconsistency, we propose an effective Diffusion Gesture Stabilizer with an annealed noise sampling strategy. Benefiting from the architectural advantages of diffusion models, we further incorporate implicit classifier-free guidance to trade off between diversity and gesture quality. Extensive experiments demonstrate that DiffGesture achieves state-of-the-art performance, which renders coherent gestures with better mode coverage and stronger audio correlations. Code is available at https://github.com/Advocate99/DiffGesture.

Can conversational videos captured from multiple egocentric viewpoints reveal the map of a scene in a cost-efficient way? We seek to answer this question by proposing a new problem: efficiently building the map of a previously unseen 3D environment by exploiting shared information in the egocentric audio-visual observations of participants in a natural conversation. Our hypothesis is that as multiple people (“egos”) move in a scene and talk among themselves, they receive rich audio-visual cues that can help uncover the unseen areas of the scene. Given the high cost of continuously processing egocentric visual streams, we further explore how to actively coordinate the sampling of visual information, so as to minimize redundancy and reduce power use. To that end, we present an audio-visual deep reinforcement learning approach that works with our shared scene mapper to selectively turn on the camera to efficiently chart out the space. We evaluate the approach using a state-of-the-art audio-visual simulator for 3D scenes as well as real-world video. Our model outperforms previous state-of-the-art mapping methods, and achieves an excellent cost-accuracy tradeoff. Project: https://vision.cs.utexas.edu/projects/chat2map.

Sound source localization is a typical and challenging task that predicts the location of sound sources in a video. Previous single-source methods mainly used the audio-visual association as clues to localize sounding objects in each frame. Due to the mixed property of multiple sound sources in the original space, there exist rare multi-source approaches to localizing multiple sources simultaneously, except for one recent work using a contrastive random walk in the graph with images and separated sound as nodes. Despite their promising performance, they can only handle a fixed number of sources, and they cannot learn compact class-aware representations for individual sources. To alleviate this shortcoming, in this paper, we propose a novel audio-visual grouping network, namely AVGN, that can directly learn category-wise semantic features for each source from the input audio mixture and frame to localize multiple sources simultaneously. Specifically, our AVGN leverages learnable audio-visual class tokens to aggregate class-aware source features. Then, the aggregated semantic features for each source can be used as guidance to localize the corresponding visual regions. Compared to existing multi-source methods, our new framework can localize a flexible number of sources and disentangle category-aware audio-visual representations for individual sound sources. We conduct extensive experiments on MUSIC, VGGSound-Instruments, and VGG-Sound Sources benchmarks. The results demonstrate that the proposed AVGN can achieve state-of-the-art sounding object localization performance on both single-source and multi-source scenarios.

We propose a self-supervised approach for learning to perform audio source separation in videos based on natural language queries, using only unlabeled video and audio pairs as training data. A key challenge in this task is learning to associate the linguistic description of a sound-emitting object to its visual features and the corresponding components of the audio waveform, all without access to annotations during training. To overcome this challenge, we adapt off-the-shelf vision-language foundation models to provide pseudo-target supervision via two novel loss functions and encourage a stronger alignment between the audio, visual and natural language modalities. During inference, our approach can separate sounds given text, video and audio input, or given text and audio input alone. We demonstrate the effectiveness of our self-supervised approach on three audio-visual separation datasets, including MUSIC, SOLOS and AudioSet, where we outperform state-of-the-art strongly supervised approaches despite not using object detectors or text labels during training. Finally, we also include samples of our separated audios in the supplemental for reference.

We explore a new task for audio-visual-language modeling called fine-grained audible video description (FAVD). It aims to provide detailed textual descriptions for the given audible videos, including the appearance and spatial locations of each object, the actions of moving objects, and the sounds in videos. Existing visual-language modeling tasks often concentrate on visual cues in videos while undervaluing the language and audio modalities. On the other hand, FAVD requires not only audio-visual-language modeling skills but also paragraph-level language generation abilities. We construct the first fine-grained audible video description benchmark (FAVDBench) to facilitate this research. For each video clip, we first provide a one-sentence summary of the video, ie, the caption, followed by 4-6 sentences describing the visual details and 1-2 audio-related descriptions at the end. The descriptions are provided in both English and Chinese. We create two new metrics for this task: an EntityScore to gauge the completeness of entities in the visual descriptions, and an AudioScore to assess the audio descriptions. As a preliminary approach to this task, we propose an audio-visual-language transformer that extends existing video captioning model with an additional audio branch. We combine the masked language modeling and auto-regressive language modeling losses to optimize our model so that it can produce paragraph-level descriptions. We illustrate the efficiency of our model in audio-visual-language modeling by evaluating it against the proposed benchmark using both conventional captioning metrics and our proposed metrics. We further put our benchmark to the test in video generation models, demonstrating that employing fine-grained video descriptions can create more intricate videos than using captions. Code and dataset are available at https://github.com/OpenNLPLab/FAVDBench. Our online benchmark is available at www.avlbench.opennlplab.cn.

Skeleton-based human action recognition is becoming increasingly important in a variety of fields. Most existing works train a CNN or GCN based backbone to extract spatial-temporal features, and use temporal average/max pooling to aggregate the information. However, these pooling methods fail to capture high-order dynamics information. To address the problem, we propose a plug-and-play module called Koopman pooling, which is a parameterized high-order pooling technique based on Koopman theory. The Koopman operator linearizes a non-linear dynamics system, thus providing a way to represent the complex system through the dynamics matrix, which can be used for classification. We also propose an eigenvalue normalization method to encourage the learned dynamics to be non-decaying and stable. Besides, we also show that our Koopman pooling framework can be easily extended to one-shot action recognition when combined with Dynamic Mode Decomposition. The proposed method is evaluated on three benchmark datasets, namely NTU RGB+D 60, 120 and NW-UCLA. Our experiments clearly demonstrate that Koopman pooling significantly improves the performance under both full-dataset and one-shot settings.

Human action recognition aims at classifying the category of human action from a segment of a video. Recently, people have dived into designing GCN-based models to extract features from skeletons for performing this task, because skeleton representations are much more efficient and robust than other modalities such as RGB frames. However, when employing the skeleton data, some important clues like related items are also discarded. It results in some ambiguous actions that are hard to be distinguished and tend to be misclassified. To alleviate this problem, we propose an auxiliary feature refinement head (FR Head), which consists of spatial-temporal decoupling and contrastive feature refinement, to obtain discriminative representations of skeletons. Ambiguous samples are dynamically discovered and calibrated in the feature space. Furthermore, FR Head could be imposed on different stages of GCNs to build a multi-level refinement for stronger supervision. Extensive experiments are conducted on NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets. Our proposed models obtain competitive results from state-of-the-art methods and can help to discriminate those ambiguous samples. Codes are available at https://github.com/zhysora/FR-Head.

In this paper we introduce a rule-based, compositional, and hierarchical modeling of action using Therbligs as our atoms. Introducing these atoms provides us with a consistent, expressive, contact-centered representation of action. Over the atoms we introduce a differentiable method of rule-based reasoning to regularize for logical consistency. Our approach is complementary to other approaches in that the Therblig-based representations produced by our architecture augment rather than replace existing architectures’ representations. We release the first Therblig-centered annotations over two popular video datasets - EPIC Kitchens 100 and 50-Salads. We also broadly demonstrate benefits to adopting Therblig representations through evaluation on the following tasks: action segmentation, action anticipation, and action recognition - observing an average 10.5%/7.53%/6.5% relative improvement, respectively, over EPIC Kitchens and an average 8.9%/6.63%/4.8% relative improvement, respectively, over 50 Salads. Code and data will be made publicly available.

Despite the success of deep learning in video understanding tasks, processing every frame in a video is computationally expensive and often unnecessary in real-time applications. Frame selection aims to extract the most informative and representative frames to help a model better understand video content. Existing frame selection methods either individually sample frames based on per-frame importance prediction, without considering interaction among frames, or adopt reinforcement learning agents to find representative frames in succession, which are costly to train and may lead to potential stability issues. To overcome the limitations of existing methods, we propose a Search-Map-Search learning paradigm which combines the advantages of heuristic search and supervised learning to select the best combination of frames from a video as one entity. By combining search with learning, the proposed method can better capture frame interactions while incurring a low inference overhead. Specifically, we first propose a hierarchical search method conducted on each training video to search for the optimal combination of frames with the lowest error on the downstream task. A feature mapping function is then learned to map the frames of a video to the representation of its target optimal frame combination. During inference, another search is performed on an unseen video to select a combination of frames whose feature representation is close to the projected feature representation. Extensive experiments based on several action recognition benchmarks demonstrate that our frame selection method effectively improves performance of action recognition models, and significantly outperforms a number of competitive baselines.

Temporal action localization (TAL) requires long-form reasoning to predict actions of various durations and complex content. Given limited GPU memory, training TAL end to end (i.e., from videos to predictions) on long videos is a significant challenge. Most methods can only train on pre-extracted features without optimizing them for the localization problem, consequently limiting localization performance. In this work, to extend the potential in TAL networks, we propose a novel end-to-end method Re2TAL, which rewires pretrained video backbones for reversible TAL. Re2TAL builds a backbone with reversible modules, where the input can be recovered from the output such that the bulky intermediate activations can be cleared from memory during training. Instead of designing one single type of reversible module, we propose a network rewiring mechanism, to transform any module with a residual connection to a reversible module without changing any parameters. This provides two benefits: (1) a large variety of reversible networks are easily obtained from existing and even future model designs, and (2) the reversible models require much less training effort as they reuse the pre-trained parameters of their original non-reversible versions. Re2TAL, only using the RGB modality, reaches 37.01% average mAP on ActivityNet-v1.3, a new state-of-the-art record, and mAP 64.9% at tIoU=0.5 on THUMOS-14, outperforming all other RGB-only methods. Code is available at https://github.com/coolbay/Re2TAL.

Due to the lack of temporal annotation, current Weakly-supervised Temporal Action Localization (WTAL) methods are generally stuck into over-complete or incomplete localization. In this paper, we aim to leverage the text information to boost WTAL from two aspects, i.e., (a) the discriminative objective to enlarge the inter-class difference, thus reducing the over-complete; (b) the generative objective to enhance the intra-class integrity, thus finding more complete temporal boundaries. For the discriminative objective, we propose a Text-Segment Mining (TSM) mechanism, which constructs a text description based on the action class label, and regards the text as the query to mine all class-related segments. Without the temporal annotation of actions, TSM compares the text query with the entire videos across the dataset to mine the best matching segments while ignoring irrelevant ones. Due to the shared sub-actions in different categories of videos, merely applying TSM is too strict to neglect the semantic-related segments, which results in incomplete localization. We further introduce a generative objective named Video-text Language Completion (VLC), which focuses on all semantic-related segments from videos to complete the text sentence. We achieve the state-of-the-art performance on THUMOS14 and ActivityNet1.3. Surprisingly, we also find our proposed method can be seamlessly applied to existing methods, and improve their performances with a clear margin. The code is available at https://github.com/lgzlIlIlI/Boosting-WTAL.

Deep sequence recognition (DSR) models receive increasing attention due to their superior application to various applications. Most DSR models use merely the target sequences as supervision without considering other related sequences, leading to over-confidence in their predictions. The DSR models trained with label smoothing regularize labels by equally and independently smoothing each token, reallocating a small value to other tokens for mitigating overconfidence. However, they do not consider tokens/sequences correlations that may provide more effective information to regularize training and thus lead to sub-optimal performance. In this work, we find tokens/sequences with high perception and semantic correlations with the target ones contain more correlated and effective information and thus facilitate more effective regularization. To this end, we propose a Perception and Semantic aware Sequence Regularization framework, which explore perceptively and semantically correlated tokens/sequences as regularization. Specifically, we introduce a semantic context-free recognition and a language model to acquire similar sequences with high perceptive similarities and semantic correlation, respectively. Moreover, over-confidence degree varies across samples according to their difficulties. Thus, we further design an adaptive calibration intensity module to compute a difficulty score for each samples to obtain finer-grained regularization. Extensive experiments on canonical sequence recognition tasks, including scene text and speech recognition, demonstrate that our method sets novel state-of-the-art results. Code is available at https://github.com/husterpzh/PSSR.

Temporal video segmentation is the get-to-go automatic video analysis, which decomposes a long-form video into smaller components for the following-up understanding tasks. Recent works have studied several levels of granularity to segment a video, such as shot, event, and scene. Those segmentations can help compare the semantics in the corresponding scales, but lack a wider view of larger temporal spans, especially when the video is complex and structured. Therefore, we present two abstractive levels of temporal segmentations and study their hierarchy to the existing fine-grained levels. Accordingly, we collect NewsNet, the largest news video dataset consisting of 1,000 videos in over 900 hours, associated with several tasks for hierarchical temporal video segmentation. Each news video is a collection of stories on different topics, represented as aligned audio, visual, and textual data, along with extensive frame-wise annotations in four granularities. We assert that the study on NewsNet can advance the understanding of complex structured video and benefit more areas such as short-video creation, personalized advertisement, digital instruction, and education. Our dataset and code is publicly available at: https://github.com/NewsNet-Benchmark/NewsNet.

Generating a video given the first several static frames is challenging as it anticipates reasonable future frames with temporal coherence. Besides video prediction, the ability to rewind from the last frame or infilling between the head and tail is also crucial, but they have rarely been explored for video completion. Since there could be different outcomes from the hints of just a few frames, a system that can follow natural language to perform video completion may significantly improve controllability. Inspired by this, we introduce a novel task, text-guided video completion (TVC), which requests the model to generate a video from partial frames guided by an instruction. We then propose Multimodal Masked Video Generation (MMVG) to address this TVC task. During training, MMVG discretizes the video frames into visual tokens and masks most of them to perform video completion from any time point. At inference time, a single MMVG model can address all 3 cases of TVC, including video prediction, rewind, and infilling, by applying corresponding masking conditions. We evaluate MMVG in various video scenarios, including egocentric, animation, and gaming. Extensive experimental results indicate that MMVG is effective in generating high-quality visual appearances with text guidance for TVC.

Computer vision models are excellent at identifying and exploiting regularities in the world. However, it is computationally costly to learn these regularities from scratch. This presents a challenge for low-parameter models, like those running on edge devices (e.g. smartphones). Can the performance of models with low representational power be improved by supplementing training with additional information about these statistical regularities? We explore this in the domains of action recognition and action anticipation, leveraging the fact that actions are typically embedded in stereotypical sequences. We introduce the Event Transition Matrix (ETM), computed from action labels in an untrimmed video dataset, which captures the temporal context of a given action, operationalized as the likelihood that it was preceded or followed by each other action in the set. We show that including information from the ETM during training improves action recognition and anticipation performance on various egocentric video datasets. Through ablation and control studies, we show that the coherent sequence of information captured by our ETM is key to this effect, and we find that the benefit of this explicit representation of temporal context is most pronounced for smaller models. Code, matrices and models are available in our project page: https://camilofosco.com/etm_website.

Most existing text-video retrieval methods focus on cross-modal matching between the visual content of videos and textual query sentences. However, in real-world scenarios, online videos are often accompanied by relevant text information such as titles, tags, and even subtitles, which can be utilized to match textual queries. This insight has motivated us to propose a novel approach to text-video retrieval, where we directly generate associated captions from videos using zero-shot video captioning with knowledge from web-scale pre-trained models (e.g., CLIP and GPT-2). Given the generated captions, a natural question arises: what benefits do they bring to text-video retrieval? To answer this, we introduce Cap4Video, a new framework that leverages captions in three ways: i) Input data: video-caption pairs can augment the training data. ii) Intermediate feature interaction: we perform cross-modal feature interaction between the video and caption to produce enhanced video representations. iii) Output score: the Query-Caption matching branch can complement the original Query-Video matching branch for text-video retrieval. We conduct comprehensive ablation studies to demonstrate the effectiveness of our approach. Without any post-processing, Cap4Video achieves state-of-the-art performance on four standard text-video retrieval benchmarks: MSR-VTT (51.4%), VATEX (66.6%), MSVD (51.8%), and DiDeMo (52.0%). The code is available at https://github.com/whwu95/Cap4Video.

In this work, we introduce Vid2Seq, a multi-modal single-stage dense event captioning model pretrained on narrated videos which are readily-available at scale. The Vid2Seq architecture augments a language model with special time tokens, allowing it to seamlessly predict event boundaries and textual descriptions in the same output sequence. Such a unified model requires large-scale training data, which is not available in current annotated datasets. We show that it is possible to leverage unlabeled narrated videos for dense video captioning, by reformulating sentence boundaries of transcribed speech as pseudo event boundaries, and using the transcribed speech sentences as pseudo event captions. The resulting Vid2Seq model pretrained on the YT-Temporal-1B dataset improves the state of the art on a variety of dense video captioning benchmarks including YouCook2, ViTT and ActivityNet Captions. Vid2Seq also generalizes well to the tasks of video paragraph captioning and video clip captioning, and to few-shot settings. Our code is publicly available at https://antoyang.github.io/vid2seq.html.

Our goal is to learn a video representation that is useful for downstream procedure understanding tasks in instructional videos. Due to the small amount of available annotations, a key challenge in procedure understanding is to be able to extract from unlabeled videos the procedural knowledge such as the identity of the task (e.g., ‘make latte’), its steps (e.g., ‘pour milk’), or the potential next steps given partial progress in its execution. Our main insight is that instructional videos depict sequences of steps that repeat between instances of the same or different tasks, and that this structure can be well represented by a Procedural Knowledge Graph (PKG), where nodes are discrete steps and edges connect steps that occur sequentially in the instructional activities. This graph can then be used to generate pseudo labels to train a video representation that encodes the procedural knowledge in a more accessible form to generalize to multiple procedure understanding tasks. We build a PKG by combining information from a text-based procedural knowledge database and an unlabeled instructional video corpus and then use it to generate training pseudo labels with four novel pre-training objectives. We call this PKG-based pre-training procedure and the resulting model Paprika, Procedure-Aware PRe-training for Instructional Knowledge Acquisition. We evaluate Paprika on COIN and CrossTask for procedure understanding tasks such as task recognition, step recognition, and step forecasting. Paprika yields a video representation that improves over the state of the art: up to 11.23% gains in accuracy in 12 evaluation settings. Implementation is available at https://github.com/salesforce/paprika.

The last several years have witnessed remarkable progress in video-and-language (VidL) understanding. However, most modern VidL approaches use complex and specialized model architectures and sophisticated pretraining protocols, making the reproducibility, analysis and comparisons of these frameworks difficult. Hence, instead of proposing yet another new VidL model, this paper conducts a thorough empirical study demystifying the most important factors in the VidL model design. Among the factors that we investigate are (i) the spatiotemporal architecture design, (ii) the multimodal fusion schemes, (iii) the pretraining objectives, (iv) the choice of pretraining data, (v) pretraining and finetuning protocols, and (vi) dataset and model scaling. Our empirical study reveals that the most important design factors include: temporal modeling, video-to-text multimodal fusion, masked modeling objectives, and joint training on images and videos. Using these empirical insights, we then develop a step-by-step recipe, dubbed VindLU, for effective VidL pretraining. Our final model trained using our recipe achieves comparable or better than state-of-the-art results on several VidL tasks without relying on external CLIP pretraining. In particular, on the text-to-video retrieval task, our approach obtains 61.2% on DiDeMo, and 55.0% on ActivityNet, outperforming current SOTA by 7.8% and 6.1% respectively. Furthermore, our model also obtains state-of-the-art video question-answering results on ActivityNet-QA, MSRVTT-QA, MSRVTT-MC and TVQA. Our code and pretrained models are publicly available at: https://github.com/klauscc/VindLU.

The question of how to best estimate the memorability of visual content is currently a source of debate in the memorability community. In this paper, we propose to explore how different key properties of images and videos affect their consolidation into memory. We analyze the impact of several features and develop a model that emulates the most important parts of a proposed “pathway to memory”: a simple but effective way of representing the different hurdles that new visual content needs to surpass to stay in memory. This framework leads to the construction of our M3-S model, a novel memorability network that processes input videos in a modular fashion. Each module of the network emulates one of the four key steps of the pathway to memory: raw encoding, scene understanding, event understanding and memory consolidation. We find that the different representations learned by our modules are non-trivial and substantially different from each other. Additionally, we observe that certain representations tend to perform better at the task of memorability prediction than others, and we introduce an in-depth ablation study to support our results. Our proposed approach surpasses the state of the art on the two largest video memorability datasets and opens the door to new applications in the field.

Complex relationships of high arity across modality and context dimensions is a critical challenge in the Emotion Recognition in Conversation (ERC) task. Yet, previous works tend to encode multimodal and contextual relationships in a loosely-coupled manner, which may harm relationship modelling. Recently, Graph Neural Networks (GNN) which show advantages in capturing data relations, offer a new solution for ERC. However, existing GNN-based ERC models fail to address some general limits of GNNs, including assuming pairwise formulation and erasing high-frequency signals, which may be trivial for many applications but crucial for the ERC task. In this paper, we propose a GNN-based model that explores multivariate relationships and captures the varying importance of emotion discrepancy and commonality by valuing multi-frequency signals. We empower GNNs to better capture the inherent relationships among utterances and deliver more sufficient multimodal and contextual modelling. Experimental results show that our proposed method outperforms previous state-of-the-art works on two popular multimodal ERC datasets.

Continuous sign language recognition (CSLR) aims to recognize glosses in a sign language video. State-of-the-art methods typically have two modules, a spatial perception module and a temporal aggregation module, which are jointly learned end-to-end. Existing results in [9,20,25,36] have indicated that, as the frontal component of the overall model, the spatial perception module used for spatial feature extraction tends to be insufficiently trained. In this paper, we first conduct empirical studies and show that a shallow temporal aggregation module allows more thorough training of the spatial perception module. However, a shallow temporal aggregation module cannot well capture both local and global temporal context information in sign language. To address this dilemma, we propose a cross-temporal context aggregation (CTCA) model. Specifically, we build a dual-path network that contains two branches for perceptions of local temporal context and global temporal context. We further design a cross-context knowledge distillation learning objective to aggregate the two types of context and the linguistic prior. The knowledge distillation enables the resultant one-branch temporal aggregation module to perceive local-global temporal and semantic context. This shallow temporal perception module structure facilitates spatial perception module learning. Extensive experiments on challenging CSLR benchmarks demonstrate that our method outperforms all state-of-the-art methods.

Large-scale transformer models bring significant improvements for various downstream vision language tasks with a unified architecture. The performance improvements come with increasing model size, resulting in slow inference speed and increased cost for severing. While some certain predictions benefit from the full complexity of the large-scale model, not all of input need the same amount of computation to conduct, potentially leading to computation resource waste. To handle this challenge, early exiting is proposed to adaptively allocate computational power in term of input complexity to improve inference efficiency. The existing early exiting strategies usually adopt output confidence based on intermediate layers as a proxy of input complexity to incur the decision of skipping following layers. However, such strategies cannot apply to encoder in the widely-used unified architecture with both encoder and decoder due to difficulty of output confidence estimation in the encoder. It is suboptimal in term of saving computation power to ignore the early exiting in encoder component. To handle this challenge, we propose a novel early exiting strategy for unified visual language models, which allows dynamically skip the layers in encoder and decoder simultaneously in term of input layer-wise similarities with multiple times of early exiting, namely MuE. By decomposing the image and text modalities in the encoder, MuE is flexible and can skip different layers in term of modalities, advancing the inference efficiency while minimizing performance drop. Experiments on the SNLI-VE and MS COCO datasets show that the proposed approach MuE can reduce inference time by up to 50% and 40% while maintaining 99% and 96% performance respectively.

Previous works for ObjectNav task attempt to learn the association (e.g. relation graph) between the visual inputs and the goal during training. Such association contains the prior knowledge of navigating in training environments, which is denoted as the experience. The experience performs a positive effect on helping the agent infer the likely location of the goal when the layout gap between the unseen environments of the test and the prior knowledge obtained in training is minor. However, when the layout gap is significant, the experience exerts a negative effect on navigation. Motivated by keeping the positive effect and removing the negative effect of the experience, we propose the layout-based soft Total Direct Effect (L-sTDE) framework based on the causal inference to adjust the prediction of the navigation policy. In particular, we propose to calculate the layout gap which is defined as the KL divergence between the posterior and the prior distribution of the object layout. Then the sTDE is proposed to appropriately control the effect of the experience based on the layout gap. Experimental results on AI2THOR, RoboTHOR, and Habitat demonstrate the effectiveness of our method.

Vision-and-Language Navigation (VLN) is the task that requires an agent to navigate through the environment based on natural language instructions. At each step, the agent takes the next action by selecting from a set of navigable locations. In this paper, we aim to take one step further and explore whether the agent can benefit from generating the potential future view during navigation. Intuitively, humans will have an expectation of how the future environment will look like, based on the natural language instructions and surrounding views, which will aid correct navigation. Hence, to equip the agent with this ability to generate the semantics of future navigation views, we first propose three proxy tasks during the agent’s in-domain pre-training: Masked Panorama Modeling (MPM), Masked Trajectory Modeling (MTM), and Action Prediction with Image Generation (APIG). These three objectives teach the model to predict missing views in a panorama (MPM), predict missing steps in the full trajectory (MTM), and generate the next view based on the full instruction and navigation history (APIG), respectively. We then fine-tune the agent on the VLN task with an auxiliary loss that minimizes the difference between the view semantics generated by the agent and the ground truth view semantics of the next step. Empirically, our VLN-SIG achieves the new state-of-the-art on both the Room-to-Room dataset and the CVDN dataset. We further show that our agent learns to fill in missing patches in future views qualitatively, which brings more interpretability over agents’ predicted actions. Lastly, we demonstrate that learning to predict future view semantics also enables the agent to have better performance on longer paths.

Recent studies in Vision-and-Language Navigation (VLN) train RL agents to execute natural-language navigation instructions in photorealistic environments, as a step towards robots that can follow human instructions. However, given the scarcity of human instruction data and limited diversity in the training environments, these agents still struggle with complex language grounding and spatial language understanding. Pre-training on large text and image-text datasets from the web has been extensively explored but the improvements are limited. We investigate large-scale augmentation with synthetic instructions. We take 500+ indoor environments captured in densely-sampled 360 degree panoramas, construct navigation trajectories through these panoramas, and generate a visually-grounded instruction for each trajectory using Marky, a high-quality multilingual navigation instruction generator. We also synthesize image observations from novel viewpoints using an image-to-image GAN. The resulting dataset of 4.2M instruction-trajectory pairs is two orders of magnitude larger than existing human-annotated datasets, and contains a wider variety of environments and viewpoints. To efficiently leverage data at this scale, we train a simple transformer agent with imitation learning. On the challenging RxR dataset, our approach outperforms all existing RL agents, improving the state-of-the-art NDTW from 71.1 to 79.1 in seen environments, and from 64.6 to 66.8 in unseen test environments. Our work points to a new path to improving instruction-following agents, emphasizing large-scale training on near-human quality synthetic instructions.

Humans possess the capacity to reason about the future based on a sparse collection of visual cues acquired over time. In order to emulate this ability, we introduce a novel task called Anticipation Captioning, which generates a caption for an unseen oracle image using a sparsely temporally-ordered set of images. To tackle this new task, we propose a model called A-CAP, which incorporates commonsense knowledge into a pre-trained vision-language model, allowing it to anticipate the caption. Through both qualitative and quantitative evaluations on a customized visual storytelling dataset, A-CAP outperforms other image captioning methods and establishes a strong baseline for anticipation captioning. We also address the challenges inherent in this task.

Recent times have witnessed an increasing number of applications of deep neural networks towards solving tasks that require superior cognitive abilities, e.g., playing Go, generating art, question answering (such as ChatGPT), etc. Such a dramatic progress raises the question: how generalizable are neural networks in solving problems that demand broad skills? To answer this question, we propose SMART: a Simple Multimodal Algorithmic Reasoning Task and the associated SMART-101 dataset, for evaluating the abstraction, deduction, and generalization abilities of neural networks in solving visuo-linguistic puzzles designed specifically for children in the 6--8 age group. Our dataset consists of 101 unique puzzles; each puzzle comprises a picture and a question, and their solution needs a mix of several elementary skills, including arithmetic, algebra, and spatial reasoning, among others. To scale our dataset towards training deep neural networks, we programmatically generate entirely new instances for each puzzle while retaining their solution algorithm. To benchmark the performance on the SMART-101 dataset, we propose a vision-and-language meta-learning model that can incorporate varied state-of-the-art neural backbones. Our experiments reveal that while powerful deep models offer reasonable performances on puzzles in a supervised setting, they are not better than random accuracy when analyzed for generalization -- filling this gap may demand new multimodal learning approaches.

Language models are capable of commonsense reasoning: while domain-specific models can learn from explicit knowledge (e.g. commonsense graphs [6], ethical norms [25]), and larger models like GPT-3 manifest broad commonsense reasoning capacity. Can their knowledge be extended to multimodal inputs such as images and audio without paired domain data? In this work, we propose ESPER (Extending Sensory PErception with Reinforcement learning) which enables text-only pretrained models to address multimodal tasks such as visual commonsense reasoning. Our key novelty is to use reinforcement learning to align multimodal inputs to language model generations without direct supervision: for example, our reward optimization relies only on cosine similarity derived from CLIP and requires no additional paired (image, text) data. Experiments demonstrate that ESPER outperforms baselines and prior work on a variety of multimodal text generation tasks ranging from captioning to commonsense reasoning; these include a new benchmark we collect and release, the ESP dataset, which tasks models with generating the text of several different domains for each image. Our code and data are publicly released at https://github.com/JiwanChung/esper.

Although the impressive performance in visual grounding, the prevailing approaches usually exploit the visual backbone in a passive way, i.e., the visual backbone extracts features with fixed weights without expression-related hints. The passive perception may lead to mismatches (e.g., redundant and missing), limiting further performance improvement. Ideally, the visual backbone should actively extract visual features since the expressions already provide the blueprint of desired visual features. The active perception can take expressions as priors to extract relevant visual features, which can effectively alleviate the mismatches. Inspired by this, we propose an active perception Visual Grounding framework based on Language Adaptive Weights, called VG-LAW. The visual backbone serves as an expression-specific feature extractor through dynamic weights generated for various expressions. Benefiting from the specific and relevant visual features extracted from the language-aware visual backbone, VG-LAW does not require additional modules for cross-modal interaction. Along with a neat multi-task head, VG-LAW can be competent in referring expression comprehension and segmentation jointly. Extensive experiments on four representative datasets, i.e., RefCOCO, RefCOCO+, RefCOCOg, and ReferItGame, validate the effectiveness of the proposed framework and demonstrate state-of-the-art performance.

Large language models (LLMs) have demonstrated excellent zero-shot generalization to new language tasks. However, effective utilization of LLMs for zero-shot visual question-answering (VQA) remains challenging, primarily due to the modality disconnection and task disconnection between LLM and VQA task. End-to-end training on vision and language data may bridge the disconnections, but is inflexible and computationally expensive. To address this issue, we propose Img2Prompt, a plug-and-play module that provides the prompts that can bridge the aforementioned modality and task disconnections, so that LLMs can perform zero-shot VQA tasks without end-to-end training. In order to provide such prompts, we further employ LLM-agnostic models to provide prompts that can describe image content and self-constructed question-answer pairs, which can effectively guide LLM to perform zero-shot VQA tasks. Img2Prompt offers the following benefits: 1) It can flexibly work with various LLMs to perform VQA. 2) Without the needing of end-to-end training, it significantly reduces the cost of deploying LLM for zero-shot VQA tasks. 3) It achieves comparable or better performance than methods relying on end-to-end training. For example, we outperform Flamingo by 5.6% on VQAv2. On the challenging A-OKVQA dataset, our method even outperforms few-shot methods by as much as 20%.

We present Diversity-Aware Meta Visual Prompting (DAM-VP), an efficient and effective prompting method for transferring pre-trained models to downstream tasks with frozen backbone. A challenging issue in visual prompting is that image datasets sometimes have a large data diversity whereas a per-dataset generic prompt can hardly handle the complex distribution shift toward the original pretraining data distribution properly. To address this issue, we propose a dataset Diversity-Aware prompting strategy whose initialization is realized by a Meta-prompt. Specifically, we cluster the downstream dataset into small homogeneity subsets in a diversity-adaptive way, with each subset has its own prompt optimized separately. Such a divide-and-conquer design reduces the optimization difficulty greatly and significantly boosts the prompting performance. Furthermore, all the prompts are initialized with a meta-prompt, which is learned across several datasets. It is a bootstrapped paradigm, with the key observation that the prompting knowledge learned from previous datasets could help the prompt to converge faster and perform better on a new dataset. During inference, we dynamically select a proper prompt for each input, based on the feature distance between the input and each subset. Through extensive experiments, our DAM-VP demonstrates superior efficiency and effectiveness, clearly surpassing previous prompting methods in a series of downstream datasets for different pretraining models. Our code is available at: https://github.com/shikiw/DAM-VP.

Vision-language models (VLMs) can effectively transfer to various vision tasks via prompt learning. Real-world scenarios often require adapting a model to multiple similar yet distinct tasks. Existing methods focus on learning a specific prompt for each task, limiting the ability to exploit potentially shared information from other tasks. Naively training a task-shared prompt using a combination of all tasks ignores fine-grained task correlations. Significant discrepancies across tasks could cause negative transferring. Considering this, we present Hierarchical Prompt (HiPro) learning, a simple and effective method for jointly adapting a pre-trained VLM to multiple downstream tasks. Our method quantifies inter-task affinity and subsequently constructs a hierarchical task tree. Task-shared prompts learned by internal nodes explore the information within the corresponding task group, while task-individual prompts learned by leaf nodes obtain fine-grained information targeted at each task. The combination of hierarchical prompts provides high-quality content of different granularity. We evaluate HiPro on four multi-task learning datasets. The results demonstrate the effectiveness of our method.

Large-scale vision-language models (VLMs) pre-trained on billion-level data have learned general visual representations and broad visual concepts. In principle, the well-learned knowledge structure of the VLMs should be inherited appropriately when being transferred to downstream tasks with limited data. However, most existing efficient transfer learning (ETL) approaches for VLMs either damage or are excessively biased towards the prior knowledge, e.g., prompt tuning (PT) discards the pre-trained text-based classifier and builds a new one while adapter-style tuning (AT) fully relies on the pre-trained features. To address this, we propose a new efficient tuning approach for VLMs named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the prior knowledge of the pre-trained models and new knowledge regarding a target task. Specifically, TaskRes keeps the original classifier weights from the VLMs frozen and obtains a new classifier for the target task by tuning a set of prior-independent parameters as a residual to the original one, which enables reliable prior knowledge preservation and flexible task-specific knowledge exploration. The proposed TaskRes is simple yet effective, which significantly outperforms previous ETL methods (e.g., PT and AT) on 11 benchmark datasets while requiring minimal effort for the implementation. Our code is available at https://github.com/geekyutao/TaskRes.

A fundamental characteristic common to both human vision and natural language is their compositional nature. Yet, despite the performance gains contributed by large vision and language pretraining, we find that--across 7 architectures trained with 4 algorithms on massive datasets--they struggle at compositionality. To arrive at this conclusion, we introduce a new compositionality evaluation benchmark, CREPE, which measures two important aspects of compositionality identified by cognitive science literature: systematicity and productivity. To measure systematicity, CREPE consists of a test dataset containing over 370K image-text pairs and three different seen-unseen splits. The three splits are designed to test models trained on three popular training datasets: CC-12M, YFCC-15M, and LAION-400M. We also generate 325K, 316K, and 309K hard negative captions for a subset of the pairs. To test productivity, CREPE contains 17K image-text pairs with nine different complexities plus 278K hard negative captions with atomic, swapping, and negation foils. The datasets are generated by repurposing the Visual Genome scene graphs and region descriptions and applying handcrafted templates and GPT-3. For systematicity, we find that model performance decreases consistently when novel compositions dominate the retrieval set, with Recall@1 dropping by up to 9%. For productivity, models’ retrieval success decays as complexity increases, frequently nearing random chance at high complexity. These results hold regardless of model and training dataset size.

Humans excel at acquiring knowledge through observation. For example, we can learn to use new tools by watching demonstrations. This skill is fundamental for intelligent systems to interact with the world. A key step to acquire this skill is to identify what part of the object affords each action, which is called affordance grounding. In this paper, we address this problem and propose a framework called LOCATE that can identify matching object parts across images, to transfer knowledge from images where an object is being used (exocentric images used for learning), to images where the object is inactive (egocentric ones used to test). To this end, we first find interaction areas and extract their feature embeddings. Then we learn to aggregate the embeddings into compact prototypes (human, object part, and background), and select the one representing the object part. Finally, we use the selected prototype to guide affordance grounding. We do this in a weakly supervised manner, learning only from image-level affordance and object labels. Extensive experiments demonstrate that our approach outperforms state-of-the-art methods by a large margin on both seen and unseen objects.

Concept-based interpretability methods aim to explain a deep neural network model’s components and predictions using a pre-defined set of semantic concepts. These methods evaluate a trained model on a new, “probe” dataset and correlate the model’s outputs with concepts labeled in that dataset. Despite their popularity, they suffer from limitations that are not well-understood and articulated in the literature. In this work, we identify and analyze three commonly overlooked factors in concept-based explanations. First, we find that the choice of the probe dataset has a profound impact on the generated explanations. Our analysis reveals that different probe datasets lead to very different explanations, suggesting that the generated explanations are not generalizable outside the probe dataset. Second, we find that concepts in the probe dataset are often harder to learn than the target classes they are used to explain, calling into question the correctness of the explanations. We argue that only easily learnable concepts should be used in concept-based explanations. Finally, while existing methods use hundreds or even thousands of concepts, our human studies reveal a much stricter upper bound of 32 concepts or less, beyond which the explanations are much less practically useful. We discuss the implications of our findings and provide suggestions for future development of concept-based interpretability methods. Code for our analysis and user interface can be found at https://github.com/princetonvisualai/OverlookedFactors.

Concept-based explanation aims to provide concise and human-understandable explanations of an image classifier. However, existing concept-based explanation methods typically require a significant amount of manually collected concept-annotated images. This is costly and runs the risk of human biases being involved in the explanation. In this paper, we propose counterfactual explanation with text-driven concepts (CounTEX), where the concepts are defined only from text by leveraging a pre-trained multi-modal joint embedding space without additional concept-annotated datasets. A conceptual counterfactual explanation is generated with text-driven concepts. To utilize the text-driven concepts defined in the joint embedding space to interpret target classifier outcome, we present a novel projection scheme for mapping the two spaces with a simple yet effective implementation. We show that CounTEX generates faithful explanations that provide a semantic understanding of model decision rationale robust to human bias.

A key goal for the advancement of AI is to develop technologies that serve the needs not just of one group but of all communities regardless of their geographical region. In fact, a significant proportion of knowledge is locally shared by people from certain regions but may not apply equally in other regions because of cultural differences. If a model is unaware of regional characteristics, it may lead to performance disparity across regions and result in bias against underrepresented groups. We propose GIVL, a Geographically Inclusive Vision-and-Language Pre-trained model. There are two attributes of geo-diverse visual concepts which can help to learn geo-diverse knowledge: 1) concepts under similar categories have unique knowledge and visual characteristics, 2) concepts with similar visual features may fall in completely different categories. Motivated by the attributes, we design new pre-training objectives Image-Knowledge Matching (IKM) and Image Edit Checking (IEC) to pre-train GIVL. Compared with similar-size models pre-trained with similar scale of data, GIVL achieves state-of-the-art (SOTA) and more balanced performance on geo-diverse V&L tasks. Code and data are released at https://github.com/WadeYin9712/GIVL.

Interpreting and explaining the behavior of deep neural networks is critical for many tasks. Explainable AI provides a way to address this challenge, mostly by providing per-pixel relevance to the decision. Yet, interpreting such explanations may require expert knowledge. Some recent attempts toward interpretability adopt a concept-based framework, giving a higher-level relationship between some concepts and model decisions. This paper proposes Bottleneck Concept Learner (BotCL), which represents an image solely by the presence/absence of concepts learned through training over the target task without explicit supervision over the concepts. It uses self-supervision and tailored regularizers so that learned concepts can be human-understandable. Using some image classification tasks as our testbed, we demonstrate BotCL’s potential to rebuild neural networks for better interpretability.

In this paper, we extend scene understanding to include that of human sketch. The result is a complete trilogy of scene representation from three diverse and complementary modalities -- sketch, photo, and text. Instead of learning a rigid three-way embedding and be done with it, we focus on learning a flexible joint embedding that fully supports the “optionality” that this complementarity brings. Our embedding supports optionality on two axis: (i) optionality across modalities -- use any combination of modalities as query for downstream tasks like retrieval, (ii) optionality across tasks -- simultaneously utilising the embedding for either discriminative (e.g., retrieval) or generative tasks (e.g., captioning). This provides flexibility to end-users by exploiting the best of each modality, therefore serving the very purpose behind our proposal of a trilogy at the first place. First, a combination of information-bottleneck and conditional invertible neural networks disentangle the modality-specific component from modality-agnostic in sketch, photo, and text. Second, the modality-agnostic instances from sketch, photo, and text are synergised using a modified cross-attention. Once learned, we show our embedding can accommodate a multi-facet of scene-related tasks, including those enabled for the first time by the inclusion of sketch, all without any task-specific modifications. Project Page: http://www.pinakinathc.me/scenetrilogy

3D visual language reasoning plays an important role in effective human-computer interaction. The current approaches for 3D visual reasoning are task-specific, and lack pre-training methods to learn generic representations that can transfer across various tasks. Despite the encouraging progress in vision-language pre-training for image-text data, 3D-language pre-training is still an open issue due to limited 3D-language paired data, highly sparse and irregular structure of point clouds and ambiguities in spatial relations of 3D objects with viewpoint changes. In this paper, we present a generic 3D-language pre-training approach, that tackles multiple facets of 3D-language reasoning by learning universal representations. Our learning objective constitutes two main parts. 1) Context aware spatial-semantic alignment to establish fine-grained correspondence between point clouds and texts. It reduces relational ambiguities by aligning 3D spatial relationships with textual semantic context. 2) Mutual 3D-Language Masked modeling to enable cross-modality information exchange. Instead of reconstructing sparse 3D points for which language can hardly provide cues, we propose masked proposal reasoning to learn semantic class and mask-invariant representations. Our proposed 3D-language pre-training method achieves promising results once adapted to various downstream tasks, including 3D visual grounding, 3D dense captioning and 3D question answering. Our codes are available at https://github.com/leolyj/3D-VLP

This paper presents a simple yet effective framework MaskCLIP, which incorporates a newly proposed masked self-distillation into contrastive language-image pretraining. The core idea of masked self-distillation is to distill representation from a full image to the representation predicted from a masked image. Such incorporation enjoys two vital benefits. First, masked self-distillation targets local patch representation learning, which is complementary to vision-language contrastive focusing on text-related representation. Second, masked self-distillation is also consistent with vision-language contrastive from the perspective of training objective as both utilize the visual encoder for feature aligning, and thus is able to learn local semantics getting indirect supervision from the language. We provide specially designed experiments with a comprehensive analysis to validate the two benefits. Symmetrically, we also introduce the local semantic supervision into the text branch, which further improves the pretraining performance. With extensive experiments, we show that MaskCLIP, when applied to various challenging downstream tasks, achieves superior results in linear probing, finetuning, and zero-shot performance with the guidance of the language encoder. We will release the code and data after the publication.

Multimodal models are becoming increasingly effective, in part due to unified components, such as the Transformer architecture. However, multimodal models still often consist of many task- and modality-specific pieces and training procedures. For example, CLIP (Radford et al., 2021) trains independent text and image towers via a contrastive loss. We explore an additional unification: the use of a pure pixel-based model to perform image, text, and multimodal tasks. Our model is trained with contrastive loss alone, so we call it CLIP-Pixels Only (CLIPPO). CLIPPO uses a single encoder that processes both regular images and text rendered as images. CLIPPO performs image-based tasks such as retrieval and zero-shot image classification almost as well as CLIP-style models, with half the number of parameters and no text-specific tower or embedding. When trained jointly via image-text contrastive learning and next-sentence contrastive learning, CLIPPO can perform well on natural language understanding tasks, without any word-level loss (language modelling or masked language modelling), outperforming pixel-based prior work. Surprisingly, CLIPPO can obtain good accuracy in visual question answering, simply by rendering the question and image together. Finally, we exploit the fact that CLIPPO does not require a tokenizer to show that it can achieve strong performance on multilingual multimodal retrieval without modifications. Code and pretrained models are available at https://github.com/google-research/big_vision.

Dominant pre-training works for image-text retrieval adopt “dual-encoder” architecture to enable high efficiency, where two encoders are used to extract image and text representations and contrastive learning is employed for global alignment. However, coarse-grained global alignment ignores detailed semantic associations between image and text. In this work, we propose a novel proxy task, named Visual-Language Error Modeling (ViLEM), to inject detailed image-text association into “dual-encoder” model by “proofreading” each word in the text against the corresponding image. Specifically, we first edit the image-paired text to automatically generate diverse plausible negative texts with pre-trained language models. ViLEM then enforces the model to discriminate the correctness of each word in the plausible negative texts and further correct the wrong words via resorting to image information. Furthermore, we propose a multi-granularity interaction framework to perform ViLEM via interacting text features with both global and local image features, which associates local text semantics with both high-level visual context and multi-level local visual information. Our method surpasses state-of-the-art “dual-encoder” methods by a large margin on the image-text retrieval task and significantly improves discriminativeness to local textual semantics. Our model can also generalize well to video-text retrieval.

Contrastive language-image pre-training (CLIP) serves as a de-facto standard to align images and texts. Nonetheless, the loose correlation between images and texts of web-crawled data renders the contrastive objective data inefficient and craving for a large training batch size. In this work, we explore the validity of non-contrastive language-image pre-training (nCLIP) and study whether nice properties exhibited in visual self-supervised models can emerge. We empirically observe that the non-contrastive objective nourishes representation learning while sufficiently underperforming under zero-shot recognition. Based on the above study, we further introduce xCLIP, a multi-tasking framework combining CLIP and nCLIP, and show that nCLIP aids CLIP in enhancing feature semantics. The synergy between two objectives lets xCLIP enjoy the best of both worlds: superior performance in both zero-shot transfer and representation learning. Systematic evaluation is conducted spanning a wide variety of downstream tasks including zero-shot classification, out-of-domain classification, retrieval, visual representation learning, and textual representation learning, showcasing a consistent performance gain and validating the effectiveness of xCLIP.

A great deal of progress has been made in image captioning, driven by research into how to encode the image using pre-trained models. This includes visual encodings (e.g. image grid features or detected objects) and more recently textual encodings (e.g. image tags or text descriptions of image regions). As more advanced encodings are available and incorporated, it is natural to ask: how to efficiently and effectively leverage the heterogeneous set of encodings? In this paper, we propose to regard the encodings as augmented views of the input image. The image captioning model encodes each view independently with a shared encoder efficiently, and a contrastive loss is incorporated across the encoded views in a novel way to improve their representation quality and the model’s data efficiency. Our proposed hierarchical decoder then adaptively weighs the encoded views according to their effectiveness for caption generation by first aggregating within each view at the token level, and then across views at the view level. We demonstrate significant performance improvements of +5.6% CIDEr on MS-COCO and +12.9% CIDEr on Flickr30k compared to state of the arts,

Fashion representation learning involves the analysis and understanding of various visual elements at different granularities and the interactions among them. Existing works often learn fine-grained fashion representations at the attribute-level without considering their relationships and inter-dependencies across different classes. In this work, we propose to learn an attribute and class specific fashion representation duet to better model such attribute relationships and inter-dependencies by leveraging prior knowledge about the taxonomy of fashion attributes and classes. Through two sub-networks for the attributes and classes, respectively, our proposed an embedding network progressively learn and refine the visual representation of a fashion image to improve its robustness for fashion retrieval. A multi-granularity loss consisting of attribute-level and class-level losses is proposed to introduce appropriate inductive bias to learn across different granularities of the fashion representations. Experimental results on three benchmark datasets demonstrate the effectiveness of our method, which outperforms the state-of-the-art methods with a large margin.

This paper aims to establish a generic multi-modal foundation model that has the scalable capability to massive downstream applications in E-commerce. Recently, large-scale vision-language pretraining approaches have achieved remarkable advances in the general domain. However, due to the significant differences between natural and product images, directly applying these frameworks for modeling image-level representations to E-commerce will be inevitably sub-optimal. To this end, we propose an instance-centric multi-modal pretraining paradigm called ECLIP in this work. In detail, we craft a decoder architecture that introduces a set of learnable instance queries to explicitly aggregate instance-level semantics. Moreover, to enable the model to focus on the desired product instance without reliance on expensive manual annotations, two specially configured pretext tasks are further proposed. Pretrained on the 100 million E-commerce-related data, ECLIP successfully extracts more generic, semantic-rich, and robust representations. Extensive experimental results show that, without further fine-tuning, ECLIP surpasses existing methods by a large margin on a broad range of downstream tasks, demonstrating the strong transferability to real-world E-commerce applications.

Learning compact image embeddings that yield semantic similarities between images and that generalize to unseen test classes, is at the core of deep metric learning (DML). Finding a mapping from a rich, localized image feature map onto a compact embedding vector is challenging: Although similarity emerges between tuples of images, DML approaches marginalize out information in an individual image before considering another image to which similarity is to be computed. Instead, we propose during training to condition the embedding of an image on the image we want to compare it to. Rather than embedding by a simple pooling as in standard DML, we use cross-attention so that one image can identify relevant features in the other image. Consequently, the attention mechanism establishes a hierarchy of conditional embeddings that gradually incorporates information about the tuple to steer the representation of an individual image. The cross-attention layers bridge the gap between the original unconditional embedding and the final similarity and allow backpropagtion to update encodings more directly than through a lossy pooling layer. At test time we use the resulting improved unconditional embeddings, thus requiring no additional parameters or computational overhead. Experiments on established DML benchmarks show that our cross-attention conditional embedding during training improves the underlying standard DML pipeline significantly so that it outperforms the state-of-the-art.

In asymmetric retrieval systems, models with different capacities are deployed on platforms with different computational and storage resources. Despite the great progress, existing approaches still suffer from a dilemma between retrieval efficiency and asymmetric accuracy due to the low capacity of the lightweight query model. In this work, we propose an Asymmetric Feature Fusion (AFF) paradigm, which advances existing asymmetric retrieval systems by considering the complementarity among different features just at the gallery side. Specifically, it first embeds each gallery image into various features, e.g., local features and global features. Then, a dynamic mixer is introduced to aggregate these features into a compact embedding for efficient search. On the query side, only a single lightweight model is deployed for feature extraction. The query model and dynamic mixer are jointly trained by sharing a momentum-updated classifier. Notably, the proposed paradigm boosts the accuracy of asymmetric retrieval without introducing any extra overhead to the query side. Exhaustive experiments on various landmark retrieval datasets demonstrate the superiority of our paradigm.

Multi-modal image-text models such as CLIP and LiT have demonstrated impressive performance on image classification benchmarks and their zero-shot generalization ability is particularly exciting. While the top-5 zero-shot accuracies of these models are very high, the top-1 accuracies are much lower (over 25% gap in some cases). We investigate the reasons for this performance gap and find that many of the failure cases are caused by ambiguity in the text prompts. First, we develop a simple and efficient zero-shot post-hoc method to identify images whose top-1 prediction is likely to be incorrect, by measuring consistency of the predictions w.r.t. multiple prompts and image transformations. We show that our procedure better predicts mistakes, outperforming the popular max logit baseline on selective prediction tasks. Next, we propose a simple and efficient way to improve accuracy on such uncertain images by making use of the WordNet hierarchy; specifically we augment the original class by incorporating its parent and children from the semantic label hierarchy, and plug the augmentation into text prompts. We conduct experiments on both CLIP and LiT models with five different ImageNet- based datasets. For CLIP, our method improves the top-1 accuracy by 17.13% on the uncertain subset and 3.6% on the entire ImageNet validation set. We also show that our method improves across ImageNet shifted datasets, four other datasets, and other model architectures such as LiT. Our proposed method is hyperparameter-free, requires no additional model training and can be easily scaled to other large multi-modal architectures. Code is available at https://github.com/gyhandy/Hierarchy-CLIP.

Despite the growing demand for tuning foundation vision transformers (FViTs) on downstream tasks, fully unleashing FViTs’ potential under data-limited scenarios (e.g., few-shot tuning) remains a challenge due to FViTs’ data-hungry nature. Common data augmentation techniques fall short in this context due to the limited features contained in the few-shot tuning data. To tackle this challenge, we first identify an opportunity for FViTs in few-shot tuning: pretrained FViTs themselves have already learned highly representative features from large-scale pretraining data, which are fully preserved during widely used parameter-efficient tuning. We thus hypothesize that leveraging those learned features to augment the tuning data can boost the effectiveness of few-shot FViT tuning. To this end, we propose a framework called Hint-based Data Augmentation (Hint-Aug), which aims to boost FViT in few-shot tuning by augmenting the over-fitted parts of tuning samples with the learned features of pretrained FViTs. Specifically, Hint-Aug integrates two key enablers: (1) an Attentive Over-fitting Detector (AOD) to detect over-confident patches of foundation ViTs for potentially alleviating their over-fitting on the few-shot tuning data and (2) a Confusion-based Feature Infusion (CFI) module to infuse easy-to-confuse features from the pretrained FViTs with the over-confident patches detected by the above AOD in order to enhance the feature diversity during tuning. Extensive experiments and ablation studies on five datasets and three parameter-efficient tuning techniques consistently validate Hint-Aug’s effectiveness: 0.04%~32.91% higher accuracy over the state-of-the-art (SOTA) data augmentation method under various low-shot settings. For example, on the Pet dataset, Hint-Aug achieves a 2.22% higher accuracy with 50% less training data over SOTA data augmentation methods.

Selecting appropriate datasets is critical in modern computer vision. However, no general-purpose tools exist to evaluate the extent to which two datasets differ. For this, we propose representing images -- and by extension datasets -- using Distributions of Neuron Activations (DNAs). DNAs fit distributions, such as histograms or Gaussians, to activations of neurons in a pre-trained feature extractor through which we pass the image(s) to represent. This extractor is frozen for all datasets, and we rely on its generally expressive power in feature space. By comparing two DNAs, we can evaluate the extent to which two datasets differ with granular control over the comparison attributes of interest, providing the ability to customise the way distances are measured to suit the requirements of the task at hand. Furthermore, DNAs are compact, representing datasets of any size with less than 15 megabytes. We demonstrate the value of DNAs by evaluating their applicability on several tasks, including conditional dataset comparison, synthetic image evaluation, and transfer learning, and across diverse datasets, ranging from synthetic cat images to celebrity faces and urban driving scenes.

3D dense captioning aims to generate multiple captions localized with their associated object regions. Existing methods follow a sophisticated “detect-then-describe” pipeline equipped with numerous hand-crafted components. However, these hand-crafted components would yield suboptimal performance given cluttered object spatial and class distributions among different scenes. In this paper, we propose a simple-yet-effective transformer framework Vote2Cap-DETR based on recent popular DEtection TRansformer (DETR). Compared with prior arts, our framework has several appealing advantages: 1) Without resorting to numerous hand-crafted components, our method is based on a full transformer encoder-decoder architecture with a learnable vote query driven object decoder, and a caption decoder that produces the dense captions in a set-prediction manner. 2) In contrast to the two-stage scheme, our method can perform detection and captioning in one-stage. 3) Without bells and whistles, extensive experiments on two commonly used datasets, ScanRefer and Nr3D, demonstrate that our Vote2Cap-DETR surpasses current state-of-the-arts by 11.13% and 7.11% in CIDEr@0.5IoU, respectively. Codes will be released soon.

Table structure recognition aims to extract the logical and physical structure of unstructured table images into a machine-readable format. The latest end-to-end image-to-text approaches simultaneously predict the two structures by two decoders, where the prediction of the physical structure (the bounding boxes of the cells) is based on the representation of the logical structure. However, as the logical representation lacks the local visual information, the previous methods often produce imprecise bounding boxes. To address this issue, we propose an end-to-end sequential modeling framework for table structure recognition called VAST. It contains a novel coordinate sequence decoder triggered by the representation of the non-empty cell from the logical structure decoder. In the coordinate sequence decoder, we model the bounding box coordinates as a language sequence, where the left, top, right and bottom coordinates are decoded sequentially to leverage the inter-coordinate dependency. Furthermore, we propose an auxiliary visual-alignment loss to enforce the logical representation of the non-empty cells to contain more local visual details, which helps produce better cell bounding boxes. Extensive experiments demonstrate that our proposed method can achieve state-of-the-art results in both logical and physical structure recognition. The ablation study also validates that the proposed coordinate sequence decoder and the visual-alignment loss are the keys to the success of our method.

We present Region-aware Open-vocabulary Vision Transformers (RO-ViT) -- a contrastive image-text pretraining recipe to bridge the gap between image-level pretraining and open-vocabulary object detection. At the pretraining phase, we propose to randomly crop and resize regions of positional embeddings instead of using the whole image positional embeddings. This better matches the use of positional embeddings at region-level in the detection finetuning phase. In addition, we replace the common softmax cross entropy loss in contrastive learning with focal loss to better learn the informative yet difficult examples. Finally, we leverage recent advances in novel object proposals to improve open-vocabulary detection finetuning. We evaluate our full model on the LVIS and COCO open-vocabulary detection benchmarks and zero-shot transfer. RO-ViT achieves a state-of-the-art 32.1 APr on LVIS, surpassing the best existing approach by +5.8 points in addition to competitive zero-shot transfer detection. Surprisingly, RO-ViT improves the image-level representation as well and achieves the state of the art on 9 out of 12 metrics on COCO and Flickr image-text retrieval benchmarks, outperforming competitive approaches with larger models.

Recent object detection approaches rely on pretrained vision-language models for image-text alignment. However, they fail to detect the Mobile User Interface (MUI) element since it contains additional OCR information, which describes its content and function but is often ignored. In this paper, we develop a new MUI element detection dataset named MUI-zh and propose an Adaptively Prompt Tuning (APT) module to take advantage of discriminating OCR information. APT is a lightweight and effective module to jointly optimize category prompts across different modalities. For every element, APT uniformly encodes its visual features and OCR descriptions to dynamically adjust the representation of frozen category prompts. We evaluate the effectiveness of our plug-and-play APT upon several existing CLIP-based detectors for both standard and open-vocabulary MUI element detection. Extensive experiments show that our method achieves considerable improvements on two datasets. The datasets is available at github.com/antmachineintelligence/MUI-zh.

We tackle open-world semantic segmentation, which aims at learning to segment arbitrary visual concepts in images, by using only image-text pairs without dense annotations. Existing open-world segmentation methods have shown impressive advances by employing contrastive learning (CL) to learn diverse visual concepts and transferring the learned image-level understanding to the segmentation task. However, these CL-based methods suffer from a train-test discrepancy, since it only considers image-text alignment during training, whereas segmentation requires region-text alignment during testing. In this paper, we proposed a novel Text-grounded Contrastive Learning (TCL) framework that enables a model to directly learn region-text alignment. Our method generates a segmentation mask for a given text, extracts text-grounded image embedding from the masked region, and aligns it with text embedding via TCL. By learning region-text alignment directly, our framework encourages a model to directly improve the quality of generated segmentation masks. In addition, for a rigorous and fair comparison, we present a unified evaluation protocol with widely used 8 semantic segmentation datasets. TCL achieves state-of-the-art zero-shot segmentation performances with large margins in all datasets. Code is available at https://github.com/kakaobrain/tcl.

Recently, CLIP has been applied to pixel-level zero-shot learning tasks via a wo-stage scheme. The general idea is to first generate class-agnostic region proposals and then feed the cropped proposal regions to CLIP to utilize its image-level zero-shot classification capability. While effective, such a scheme requires two image encoders, one for proposal generation and one for CLIP, leading to a complicated pipeline and high computational cost. In this work, we pursue a simpler-and-efficient one-stage solution that directly extends CLIP’s zero-shot prediction capability from image to pixel level. Our investigation starts with a straightforward extension as our baseline that generates semantic masks by comparing the similarity between text and patch embeddings extracted from CLIP. However, such a paradigm could heavily overfit the seen classes and fail to generalize to unseen classes. To handle this issue, we propose three simple-but-effective designs and figure out that they can significantly retain the inherent zero-shot capacity of CLIP and improve pixel-level generalization ability. Incorporating those modifications leads to an efficient zero-shot semantic segmentation system called ZegCLIP. Through extensive experiments on three public benchmarks, ZegCLIP demonstrates superior performance, outperforming the state-of-the-art methods by a large margin under both “inductive” and “transductive” zero-shot settings. In addition, compared with the two-stage method, our one-stage ZegCLIP achieves a speedup of about 5 times faster during inference. We release the code at https://github.com/ZiqinZhou66/ZegCLIP.git.

Open-vocabulary object detection aims to provide object detectors trained on a fixed set of object categories with the generalizability to detect objects described by arbitrary text queries. Previous methods adopt knowledge distillation to extract knowledge from Pretrained Vision-and-Language Models (PVLMs) and transfer it to detectors. However, due to the non-adaptive proposal cropping and single-level feature mimicking processes, they suffer from information destruction during knowledge extraction and inefficient knowledge transfer. To remedy these limitations, we propose an Object-Aware Distillation Pyramid (OADP) framework, including an Object-Aware Knowledge Extraction (OAKE) module and a Distillation Pyramid (DP) mechanism. When extracting object knowledge from PVLMs, the former adaptively transforms object proposals and adopts object-aware mask attention to obtain precise and complete knowledge of objects. The latter introduces global and block distillation for more comprehensive knowledge transfer to compensate for the missing relation information in object distillation. Extensive experiments show that our method achieves significant improvement compared to current methods. Especially on the MS-COCO dataset, our OADP framework reaches 35.6 mAP^N50, surpassing the current state-of-the-art method by 3.3 mAP^N50. Code is anonymously provided in the supplementary materials.

Compositional Zero-Shot Learning (CZSL) aims to train models to recognize novel compositional concepts based on learned concepts such as attribute-object combinations. One of the challenges is to model attributes interacted with different objects, e.g., the attribute “wet” in “wet apple” and “wet cat” is different. As a solution, we provide analysis and argue that attributes are conditioned on the recognized object and input image and explore learning conditional attribute embeddings by a proposed attribute learning framework containing an attribute hyper learner and an attribute base learner. By encoding conditional attributes, our model enables to generate flexible attribute embeddings for generalization from seen to unseen compositions. Experiments on CZSL benchmarks, including the more challenging C-GQA dataset, demonstrate better performances compared with other state-of-the-art approaches and validate the importance of learning conditional attributes.

Existing semantic segmentation approaches are often limited by costly pixel-wise annotations and predefined classes. In this work, we present CLIP-S^4 that leverages self-supervised pixel representation learning and vision-language models to enable various semantic segmentation tasks (e.g., unsupervised, transfer learning, language-driven segmentation) without any human annotations and unknown class information. We first learn pixel embeddings with pixel-segment contrastive learning from different augmented views of images. To further improve the pixel embeddings and enable language-driven semantic segmentation, we design two types of consistency guided by vision-language models: 1) embedding consistency, aligning our pixel embeddings to the joint feature space of a pre-trained vision-language model, CLIP; and 2) semantic consistency, forcing our model to make the same predictions as CLIP over a set of carefully designed target classes with both known and unknown prototypes. Thus, CLIP-S^4 enables a new task of class-free semantic segmentation where no unknown class information is needed during training. As a result, our approach shows consistent and substantial performance improvement over four popular benchmarks compared with the state-of-the-art unsupervised and language-driven semantic segmentation methods. More importantly, our method outperforms these methods on unknown class recognition by a large margin.

Visual place recognition (VPR) is usually considered as a specific image retrieval problem. Limited by existing training frameworks, most deep learning-based works cannot extract sufficiently stable global features from RGB images and rely on a time-consuming re-ranking step to exploit spatial structural information for better performance. In this paper, we propose StructVPR, a novel training architecture for VPR, to enhance structural knowledge in RGB global features and thus improve feature stability in a constantly changing environment. Specifically, StructVPR uses segmentation images as a more definitive source of structural knowledge input into a CNN network and applies knowledge distillation to avoid online segmentation and inference of seg-branch in testing. Considering that not all samples contain high-quality and helpful knowledge, and some even hurt the performance of distillation, we partition samples and weigh each sample’s distillation loss to enhance the expected knowledge precisely. Finally, StructVPR achieves impressive performance on several benchmarks using only global retrieval and even outperforms many two-stage approaches by a large margin. After adding additional re-ranking, ours achieves state-of-the-art performance while maintaining a low computational cost.

Domain adaptive panoptic segmentation aims to mitigate data annotation challenge by leveraging off-the-shelf annotated data in one or multiple related source domains. However, existing studies employ two separate networks for instance segmentation and semantic segmentation which lead to excessive network parameters as well as complicated and computationally intensive training and inference processes. We design UniDAformer, a unified domain adaptive panoptic segmentation transformer that is simple but can achieve domain adaptive instance segmentation and semantic segmentation simultaneously within a single network. UniDAformer introduces Hierarchical Mask Calibration (HMC) that rectifies inaccurate predictions at the level of regions, superpixels and pixels via online self-training on the fly. It has three unique features: 1) it enables unified domain adaptive panoptic adaptation; 2) it mitigates false predictions and improves domain adaptive panoptic segmentation effectively; 3) it is end-to-end trainable with a much simpler training and inference pipeline. Extensive experiments over multiple public benchmarks show that UniDAformer achieves superior domain adaptive panoptic segmentation as compared with the state-of-the-art.

We study universal zero-shot segmentation in this work to achieve panoptic, instance, and semantic segmentation for novel categories without any training samples. Such zero-shot segmentation ability relies on inter-class relationships in semantic space to transfer the visual knowledge learned from seen categories to unseen ones. Thus, it is desired to well bridge semantic-visual spaces and apply the semantic relationships to visual feature learning. We introduce a generative model to synthesize features for unseen categories, which links semantic and visual spaces as well as address the issue of lack of unseen training data. Furthermore, to mitigate the domain gap between semantic and visual spaces, firstly, we enhance the vanilla generator with learned primitives, each of which contains fine-grained attributes related to categories, and synthesize unseen features by selectively assembling these primitives. Secondly, we propose to disentangle the visual feature into the semantic-related part and the semantic-unrelated part that contains useful visual classification clues but is less relevant to semantic representation. The inter-class relationships of semantic-related visual features are then required to be aligned with those in semantic space, thereby transferring semantic knowledge to visual feature learning. The proposed approach achieves impressively state-of-the-art performance on zero-shot panoptic segmentation, instance segmentation, and semantic segmentation.

Despite the progress in semantic image synthesis, it remains a challenging problem to generate photo-realistic parts from input semantic map. Integrating part segmentation map can undoubtedly benefit image synthesis, but is bothersome and inconvenient to be provided by users. To improve part synthesis, this paper presents to infer Parts from Object ShapE (iPOSE) and leverage it for improving semantic image synthesis. However, albeit several part segmentation datasets are available, part annotations are still not provided for many object categories in semantic image synthesis. To circumvent it, we resort to few-shot regime to learn a PartNet for predicting the object part map with the guidance of pre-defined support part maps. PartNet can be readily generalized to handle a new object category when a small number (e.g., 3) of support part maps for this category are provided. Furthermore, part semantic modulation is presented to incorporate both inferred part map and semantic map for image synthesis. Experiments show that our iPOSE not only generates objects with rich part details, but also enables to control the image synthesis flexibly. And our iPOSE performs favorably against the state-of-the-art methods in terms of quantitative and qualitative evaluation. Our code will be publicly available at https://github.com/csyxwei/iPOSE.

In this work, we present a robust approach for joint part and object segmentation. Specifically, we reformulate object and part segmentation as an optimization problem and build a hierarchical feature representation including pixel, part, and object-level embeddings to solve it in a bottom-up clustering manner. Pixels are grouped into several clusters where the part-level embeddings serve as cluster centers. Afterwards, object masks are obtained by compositing the part proposals. This bottom-up interaction is shown to be effective in integrating information from lower semantic levels to higher semantic levels. Based on that, our novel approach Compositor produces part and object segmentation masks simultaneously while improving the mask quality. Compositor achieves state-of-the-art performance on PartImageNet and Pascal-Part by outperforming previous methods by around 0.9% and 1.3% on PartImageNet, 0.4% and 1.7% on Pascal-Part in terms of part and object mIoU and demonstrates better robustness against occlusion by around 4.4% and 7.1% on part and object respectively.

This paper introduces a generalized few-shot segmentation framework with a straightforward training process and an easy-to-optimize inference phase. In particular, we propose a simple yet effective model based on the well-known InfoMax principle, where the Mutual Information (MI) between the learned feature representations and their corresponding predictions is maximized. In addition, the terms derived from our MI-based formulation are coupled with a knowledge distillation term to retain the knowledge on base classes. With a simple training process, our inference model can be applied on top of any segmentation network trained on base classes. The proposed inference yields substantial improvements on the popular few-shot segmentation benchmarks, PASCAL-5^i and COCO-20^i. Particularly, for novel classes, the improvement gains range from 7% to 26% (PASCAL-5^i) and from 3% to 12% (COCO-20^i) in the 1-shot and 5-shot scenarios, respectively. Furthermore, we propose a more challenging setting, where performance gaps are further exacerbated. Our code is publicly available at https://github.com/sinahmr/DIaM.

The representative instance segmentation methods mostly segment different object instances with a mask of the fixed resolution, e.g., 28× 28 grid. However, a low-resolution mask loses rich details, while a high-resolution mask incurs quadratic computation overhead. It is a challenging task to predict the optimal binary mask for each instance. In this paper, we propose to dynamically select suitable masks for different object proposals. First, a dual-level Feature Pyramid Network (FPN) with adaptive feature aggregation is developed to gradually increase the mask grid resolution, ensuring high-quality segmentation of objects. Specifically, an efficient region-level top-down path (r-FPN) is introduced to incorporate complementary contextual and detailed information from different stages of image-level FPN (i-FPN). Then, to alleviate the increase of computation and memory costs caused by using large masks, we develop a Mask Switch Module (MSM) with negligible computational cost to select the most suitable mask resolution for each instance, achieving high efficiency while maintaining high segmentation accuracy. Without bells and whistles, the proposed method, namely DynaMask, brings consistent and noticeable performance improvements over other state-of-the-arts at a moderate computation overhead. The source code: https://github.com/lslrh/DynaMask.

Discriminative representation is essential to keep a unique identifier for each target in Multiple object tracking (MOT). Some recent MOT methods extract features of the bounding box region or the center point as identity embeddings. However, when targets are occluded, these coarse-grained global representations become unreliable. To this end, we propose exploring diverse fine-grained representation, which describes appearance comprehensively from global and local perspectives. This fine-grained representation requires high feature resolution and precise semantic information. To effectively alleviate the semantic misalignment caused by indiscriminate contextual information aggregation, Flow Alignment FPN (FAFPN) is proposed for multi-scale feature alignment aggregation. It generates semantic flow among feature maps from different resolutions to transform their pixel positions. Furthermore, we present a Multi-head Part Mask Generator (MPMG) to extract fine-grained representation based on the aligned feature maps. Multiple parallel branches of MPMG allow it to focus on different parts of targets to generate local masks without label supervision. The diverse details in target masks facilitate fine-grained representation. Eventually, benefiting from a Shuffle-Group Sampling (SGS) training strategy with positive and negative samples balanced, we achieve state-of-the-art performance on MOT17 and MOT20 test sets. Even on DanceTrack, where the appearance of targets is extremely similar, our method significantly outperforms ByteTrack by 5.0% on HOTA and 5.6% on IDF1. Extensive experiments have proved that diverse fine-grained representation makes Re-ID great again in MOT.

The DETR-like segmentors have underpinned the most recent breakthroughs in semantic segmentation, which end-to-end train a set of queries representing the class prototypes or target segments. Recently, masked attention is proposed to restrict each query to only attend to the foreground regions predicted by the preceding decoder block for easier optimization. Although promising, it relies on the learnable parameterized positional queries which tend to encode the dataset statistics, leading to inaccurate localization for distinct individual queries. In this paper, we propose a simple yet effective query design for semantic segmentation termed Dynamic Focus-aware Positional Queries (DFPQ), which dynamically generates positional queries conditioned on the cross-attention scores from the preceding decoder block and the positional encodings for the corresponding image features, simultaneously. Therefore, our DFPQ preserves rich localization information for the target segments and provides accurate and fine-grained positional priors. In addition, we propose to efficiently deal with high-resolution cross-attention by only aggregating the contextual tokens based on the low-resolution cross-attention scores to perform local relation aggregation. Extensive experiments on ADE20K and Cityscapes show that with the two modifications on Mask2former, our framework achieves SOTA performance and outperforms Mask2former by clear margins of 1.1%, 1.9%, and 1.1% single-scale mIoU with ResNet-50, Swin-T, and Swin-B backbones on the ADE20K validation set, respectively. Source code is available at https://github.com/ziplab/FASeg.

Correctness of instance segmentation constitutes counting the number of objects, correctly localizing all predictions and classifying each localized prediction. Average Precision is the de-facto metric used to measure all these constituents of segmentation. However, this metric does not penalize duplicate predictions in the high-recall range, and cannot distinguish instances that are localized correctly but categorized incorrectly. This weakness has inadvertently led to network designs that achieve significant gains in AP but also introduce a large number of false positives. We therefore cannot rely on AP to choose a model that provides an optimal tradeoff between false positives and high recall. To resolve this dilemma, we review alternative metrics in the literature and propose two new measures to explicitly measure the amount of both spatial and categorical duplicate predictions. We also propose a Semantic Sorting and NMS module to remove these duplicates based on a pixel occupancy matching scheme. Experiments show that modern segmentation networks have significant gains in AP, but also contain a considerable amount of duplicates. Our Semantic Sorting and NMS can be added as a plug-and-play module to mitigate hedged predictions and preserve AP.

Generalized few-shot semantic segmentation (GFSS) distinguishes pixels of base and novel classes from the background simultaneously, conditioning on sufficient data of base classes and a few examples from novel class. A typical GFSS approach has two training phases: base class learning and novel class updating. Nevertheless, such a stand-alone updating process often compromises the well-learnt features and results in performance drop on base classes. In this paper, we propose a new idea of leveraging Projection onto Orthogonal Prototypes (POP), which updates features to identify novel classes without compromising base classes. POP builds a set of orthogonal prototypes, each of which represents a semantic class, and makes the prediction for each class separately based on the features projected onto its prototype. Technically, POP first learns prototypes on base data, and then extends the prototype set to novel classes. The orthogonal constraint of POP encourages the orthogonality between the learnt prototypes and thus mitigates the influence on base class features when generalizing to novel prototypes. Moreover, we capitalize on the residual of feature projection as the background representation to dynamically fit semantic shifting (i.e., background no longer includes the pixels of novel classes in updating phase). Extensive experiments on two benchmarks demonstrate that our POP achieves superior performances on novel classes without sacrificing much accuracy on base classes. Notably, POP outperforms the state-of-the-art fine-tuning by 3.93% overall mIoU on PASCAL-5i in 5-shot scenario.

In Weakly Supervised Semantic Segmentation (WSSS), Class Activation Maps (CAMs) usually 1) do not cover the whole object and 2) be activated on irrelevant regions. To address the issues, we propose a novel WSSS framework via adversarial learning of a classifier and an image reconstructor. When an image is perfectly decomposed into class-wise segments, information (i.e., color or texture) of a single segment could not be inferred from the other segments. Therefore, inferability between the segments can represent the preciseness of segmentation. We quantify the inferability as a reconstruction quality of one segment from the other segments. If one segment could be reconstructed from the others, then the segment would be imprecise. To bring this idea into WSSS, we simultaneously train two models: a classifier generating CAMs that decompose an image into segments and a reconstructor that measures the inferability between the segments. As in GANs, while being alternatively trained in an adversarial manner, two networks provide positive feedback to each other. We verify the superiority of the proposed framework with extensive ablation studies. Our method achieves new state-of-the-art performances on both PASCAL VOC 2012 and MS COCO 2014. The code is available at https://github.com/sangrockEG/ACR.

Semi-supervised learning improves data efficiency of deep models by leveraging unlabeled samples to alleviate the reliance on a large set of labeled samples. These successes concentrate on the pixel-wise consistency by using convolutional neural networks (CNNs) but fail to address both global learning capability and class-level features for unlabeled data. Recent works raise a new trend that Trans- former achieves superior performance on the entire feature map in various tasks. In this paper, we unify the current dominant Mean-Teacher approaches by reconciling intra- model and inter-model properties for semi-supervised segmentation to produce a novel algorithm, SemiCVT, that absorbs the quintessence of CNNs and Transformer in a comprehensive way. Specifically, we first design a parallel CNN-Transformer architecture (CVT) with introducing an intra-model local-global interaction schema (LGI) in Fourier domain for full integration. The inter-model class- wise consistency is further presented to complement the class-level statistics of CNNs and Transformer in a cross- teaching manner. Extensive empirical evidence shows that SemiCVT yields consistent improvements over the state-of- the-art methods in two public benchmarks.

Recent studies on semi-supervised semantic segmentation (SSS) have seen fast progress. Despite their promising performance, current state-of-the-art methods tend to increasingly complex designs at the cost of introducing more network components and additional training procedures. Differently, in this work, we follow a standard teacher-student framework and propose AugSeg, a simple and clean approach that focuses mainly on data perturbations to boost the SSS performance. We argue that various data augmentations should be adjusted to better adapt to the semi-supervised scenarios instead of directly applying these techniques from supervised learning. Specifically, we adopt a simplified intensity-based augmentation that selects a random number of data transformations with uniformly sampling distortion strengths from a continuous space. Based on the estimated confidence of the model on different unlabeled samples, we also randomly inject labelled information to augment the unlabeled samples in an adaptive manner. Without bells and whistles, our simple AugSeg can readily achieve new state-of-the-art performance on SSS benchmarks under different partition protocols.

In this paper, we introduce a novel learning scheme named weakly semi-supervised instance segmentation (WSSIS) with point labels for budget-efficient and high-performance instance segmentation. Namely, we consider a dataset setting consisting of a few fully-labeled images and a lot of point-labeled images. Motivated by the main challenge of semi-supervised approaches mainly derives from the trade-off between false-negative and false-positive instance proposals, we propose a method for WSSIS that can effectively leverage the budget-friendly point labels as a powerful weak supervision source to resolve the challenge. Furthermore, to deal with the hard case where the amount of fully-labeled data is extremely limited, we propose a MaskRefineNet that refines noise in rough masks. We conduct extensive experiments on COCO and BDD100K datasets, and the proposed method achieves promising results comparable to those of the fully-supervised model, even with 50% of the fully labeled COCO data (38.8% vs. 39.7%). Moreover, when using as little as 5% of fully labeled COCO data, our method shows significantly superior performance over the state-of-the-art semi-supervised learning method (33.7% vs. 24.9%). The code is available at https://github.com/clovaai/PointWSSIS.

Exemplar-based class-incremental learning (CIL) finetunes the model with all samples of new classes but few-shot exemplars of old classes in each incremental phase, where the “few-shot” abides by the limited memory budget. In this paper, we break this “few-shot” limit based on a simple yet surprisingly effective idea: compressing exemplars by downsampling non-discriminative pixels and saving “many-shot” compressed exemplars in the memory. Without needing any manual annotation, we achieve this compression by generating 0-1 masks on discriminative pixels from class activation maps (CAM). We propose an adaptive mask generation model called class-incremental masking (CIM) to explicitly resolve two difficulties of using CAM: 1) transforming the heatmaps of CAM to 0-1 masks with an arbitrary threshold leads to a trade-off between the coverage on discriminative pixels and the quantity of exemplars, as the total memory is fixed; and 2) optimal thresholds vary for different object classes, which is particularly obvious in the dynamic environment of CIL. We optimize the CIM model alternatively with the conventional CIL model through a bilevel optimization problem. We conduct extensive experiments on high-resolution CIL benchmarks including Food-101, ImageNet-100, and ImageNet-1000, and show that using the compressed exemplars by CIM can achieve a new state-of-the-art CIL accuracy, e.g., 4.8 percentage points higher than FOSTER on 10-Phase ImageNet-1000. Our code is available at https://github.com/xfflzl/CIM-CIL.

Annotating new datasets for machine learning tasks is tedious, time-consuming, and costly. For segmentation applications, the burden is particularly high as manual delineations of relevant image content are often extremely expensive or can only be done by experts with domain-specific knowledge. Thanks to developments in transfer learning and training with weak supervision, segmentation models can now also greatly benefit from annotations of different kinds. However, for any new domain application looking to use weak supervision, the dataset builder still needs to define a strategy to distribute full segmentation and other weak annotations. Doing so is challenging, however, as it is a priori unknown how to distribute an annotation budget for a given new dataset. To this end, we propose a novel approach to determine annotation strategies for segmentation datasets, whereby estimating what proportion of segmentation and classification annotations should be collected given a fixed budget. To do so, our method sequentially determines proportions of segmentation and classification annotations to collect for budget-fractions by modeling the expected improvement of the final segmentation model. We show in our experiments that our approach yields annotations that perform very close to the optimal for a number of different annotation budgets and datasets.

Self-supervised learning (SSL) strategies have demonstrated remarkable performance in various recognition tasks. However, both our preliminary investigation and recent studies suggest that they may be less effective in learning representations for fine-grained visual recognition (FGVR) since many features helpful for optimizing SSL objectives are not suitable for characterizing the subtle differences in FGVR. To overcome this issue, we propose learning an additional screening mechanism to identify discriminative clues commonly seen across instances and classes, dubbed as common rationales in this paper. Intuitively, common rationales tend to correspond to the discriminative patterns from the key parts of foreground objects. We show that a common rationale detector can be learned by simply exploiting the GradCAM induced from the SSL objective without using any pre-trained object parts or saliency detectors, making it seamlessly to be integrated with the existing SSL process. Specifically, we fit the GradCAM with a branch with limited fitting capacity, which allows the branch to capture the common rationales and discard the less common discriminative patterns. At the test stage, the branch generates a set of spatial weights to selectively aggregate features representing an instance. Extensive experimental results on four visual tasks demonstrate that the proposed method can lead to a significant improvement in different evaluation settings.

Combining multiple datasets enables performance boost on many computer vision tasks. But similar trend has not been witnessed in object detection when combining multiple datasets due to two inconsistencies among detection datasets: taxonomy difference and domain gap. In this paper, we address these challenges by a new design (named Detection Hub) that is dataset-aware and category-aligned. It not only mitigates the dataset inconsistency but also provides coherent guidance for the detector to learn across multiple datasets. In particular, the dataset-aware design is achieved by learning a dataset embedding that is used to adapt object queries as well as convolutional kernels in detection heads. The categories across datasets are semantically aligned into a unified space by replacing one-hot category representations with word embedding and leveraging the semantic coherence of language embedding. Detection Hub fulfills the benefits of large data on object detection. Experiments demonstrate that joint training on multiple datasets achieves significant performance gains over training on each dataset alone. Detection Hub further achieves SoTA performance on UODB benchmark with wide variety of datasets.

Recently, AutoFlow has shown promising results on learning a training set for optical flow, but requires ground truth labels in the target domain to compute its search metric. Observing a strong correlation between the ground truth search metric and self-supervised losses, we introduce self-supervised AutoFlow to handle real-world videos without ground truth labels. Using self-supervised loss as the search metric, our self-supervised AutoFlow performs on par with AutoFlow on Sintel and KITTI where ground truth is available, and performs better on the real-world DAVIS dataset. We further explore using self-supervised AutoFlow in the (semi-)supervised setting and obtain competitive results against the state of the art.

This paper presents a DETR-based method for cross-domain weakly supervised object detection (CDWSOD), aiming at adapting the detector from source to target domain through weak supervision. We think DETR has strong potential for CDWSOD due to an insight: the encoder and the decoder in DETR are both based on the attention mechanism and are thus capable of aggregating semantics across the entire image. The aggregation results, i.e., image-level predictions, can naturally exploit the weak supervision for domain alignment. Such motivated, we propose DETR with additional Global Aggregation (DETR-GA), a CDWSOD detector that simultaneously makes “instance-level + image-level” predictions and utilizes “strong + weak” supervisions. The key point of DETR-GA is very simple: for the encoder / decoder, we respectively add multiple class queries / a foreground query to aggregate the semantics into image-level predictions. Our query-based aggregation has two advantages. First, in the encoder, the weakly-supervised class queries are capable of roughly locating the corresponding positions and excluding the distraction from non-relevant regions. Second, through our design, the object queries and the foreground query in the decoder share consensus on the class semantics, therefore making the strong and weak supervision mutually benefit each other for domain alignment. Extensive experiments on four popular cross-domain benchmarks show that DETR-GA significantly improves CSWSOD and advances the states of the art (e.g., 29.0% --> 79.4% mAP on PASCAL VOC --> Clipart_all dataset).

In this paper, we formally address universal object detection, which aims to detect every scene and predict every category. The dependence on human annotations, the limited visual information, and the novel categories in the open world severely restrict the universality of traditional detectors. We propose UniDetector, a universal object detector that has the ability to recognize enormous categories in the open world. The critical points for the universality of UniDetector are: 1) it leverages images of multiple sources and heterogeneous label spaces for training through the alignment of image and text spaces, which guarantees sufficient information for universal representations. 2) it generalizes to the open world easily while keeping the balance between seen and unseen classes, thanks to abundant information from both vision and language modalities. 3) it further promotes the generalization ability to novel categories through our proposed decoupling training manner and probability calibration. These contributions allow UniDetector to detect over 7k categories, the largest measurable category size so far, with only about 500 classes participating in training. Our UniDetector behaves the strong zero-shot generalization ability on large-vocabulary datasets like LVIS, ImageNetBoxes, and VisualGenome - it surpasses the traditional supervised baselines by more than 4% on average without seeing any corresponding images. On 13 public detection datasets with various scenes, UniDetector also achieves state-of-the-art performance with only a 3% amount of training data.

Open World Object Detection (OWOD) is a new and challenging computer vision task that bridges the gap between classic object detection (OD) benchmarks and object detection in the real world. In addition to detecting and classifying seen/labeled objects, OWOD algorithms are expected to detect novel/unknown objects - which can be classified and incrementally learned. In standard OD, object proposals not overlapping with a labeled object are automatically classified as background. Therefore, simply applying OD methods to OWOD fails as unknown objects would be predicted as background. The challenge of detecting unknown objects stems from the lack of supervision in distinguishing unknown objects and background object proposals. Previous OWOD methods have attempted to overcome this issue by generating supervision using pseudo-labeling - however, unknown object detection has remained low. Probabilistic/generative models may provide a solution for this challenge. Herein, we introduce a novel probabilistic framework for objectness estimation, where we alternate between probability distribution estimation and objectness likelihood maximization of known objects in the embedded feature space - ultimately allowing us to estimate the objectness probability of different proposals. The resulting Probabilistic Objectness transformer-based open-world detector, PROB, integrates our framework into traditional object detection models, adapting them for the open-world setting. Comprehensive experiments on OWOD benchmarks show that PROB outperforms all existing OWOD methods in both unknown object detection (~2x unknown recall) and known object detection (~ mAP). Our code is available at https://github.com/orrzohar/PROB.

Open world object detection (OWOD) has attracted extensive attention due to its practicability in the real world. Previous OWOD works manually designed unknown-discover strategies to select unknown proposals from the background, suffering from uncertainties without appropriate priors. In this paper, we claim the learning of object detection could be seen as an object-level feature-entanglement process, where unknown traits are propagated to the known proposals through convolutional operations and could be distilled to benefit unknown recognition without manual selection. Therefore, we propose a simple yet effective Annealing-based Label-Transfer framework, which sufficiently explores the known proposals to alleviate the uncertainties. Specifically, a Label-Transfer Learning paradigm is introduced to decouple the known and unknown features, while a Sawtooth Annealing Scheduling strategy is further employed to rebuild the decision boundaries of the known and unknown classes, thus promoting both known and unknown recognition. Moreover, previous OWOD works neglected the trade-off of known and unknown performance, and we thus introduce a metric called Equilibrium Index to comprehensively evaluate the effectiveness of the OWOD models. To the best of our knowledge, this is the first OWOD work without manual unknown selection. Extensive experiments conducted on the common-used benchmark validate that our model achieves superior detection performance (200% unknown mAP improvement with the even higher known detection performance) compared to other state-of-the-art methods. Our code is available at https://github.com/DIG-Beihang/ALLOW.git.

The task of key-points detection and description is to estimate the stable location and discriminative representation of local features, which is essential for image matching. However, either the rough hard positive or negative labels generated from one-to-one correspondences among images bring indistinguishable samples, called pseudo positives or negatives, which act as inconsistent supervisions while learning key-points used for matching. Such pseudo-labeled samples prevent deep neural networks from learning discriminative descriptions for accurate matching. To tackle this challenge, we propose to learn transformation-predictive representations with self-supervised contrastive learning. We maximize the similarity between corresponded views of the same 3D point (landmark) by using none of the negative sample pairs (including true and pseudo negatives) and avoiding collapsing solutions. Then we design a learnable label prediction mechanism to soften the hard positive labels into soft continuous targets. The aggressively updated soft labels extensively deal with the training bottleneck (derived from the label noise of pseudo positives) and make the model can be trained under a stronger augmentation paradigm. Our self-supervised method outperforms the state-of-the-art on the standard image matching benchmarks by noticeable margins and shows excellent generalization capability on multiple downstream tasks.

Deep neural networks (DNNs) have enabled astounding progress in several vision-based problems. Despite showing high predictive accuracy, recently, several works have revealed that they tend to provide overconfident predictions and thus are poorly calibrated. The majority of the works addressing the miscalibration of DNNs fall under the scope of classification and consider only in-domain predictions. However, there is little to no progress in studying the calibration of DNN-based object detection models, which are central to many vision-based safety-critical applications. In this paper, inspired by the train-time calibration methods, we propose a novel auxiliary loss formulation that explicitly aims to align the class confidence of bounding boxes with the accurateness of predictions (i.e. precision). Since the original formulation of our loss depends on the counts of true positives and false positives in a minibatch, we develop a differentiable proxy of our loss that can be used during training with other application-specific loss functions. We perform extensive experiments on challenging in-domain and out-domain scenarios with six benchmark datasets including MS-COCO, Cityscapes, Sim10k, and BDD100k. Our results reveal that our train-time loss surpasses strong calibration baselines in reducing calibration error for both in and out-domain scenarios. Our source code and pre-trained models are available at https://github.com/akhtarvision/bpc_calibration

Object detection at night is a challenging problem due to the absence of night image annotations. Despite several domain adaptation methods, achieving high-precision results remains an issue. False-positive error propagation is still observed in methods using the well-established student-teacher framework, particularly for small-scale and low-light objects. This paper proposes a two-phase consistency unsupervised domain adaptation network, 2PCNet, to address these issues. The network employs high-confidence bounding-box predictions from the teacher in the first phase and appends them to the student’s region proposals for the teacher to re-evaluate in the second phase, resulting in a combination of high and low confidence pseudo-labels. The night images and pseudo-labels are scaled-down before being used as input to the student, providing stronger small-scale pseudo-labels. To address errors that arise from low-light regions and other night-related attributes in images, we propose a night-specific augmentation pipeline called NightAug. This pipeline involves applying random augmentations, such as glare, blur, and noise, to daytime images. Experiments on publicly available datasets demonstrate that our method achieves superior results to state-of-the-art methods by 20%, and to supervised models trained directly on the target data.

Recently, CLIP-guided image synthesis has shown appealing performance on adapting a pre-trained source-domain generator to an unseen target domain. It does not require any target-domain samples but only the textual domain labels. The training is highly efficient, e.g., a few minutes. However, existing methods still have some limitations in the quality of generated images and may suffer from the mode collapse issue. A key reason is that a fixed adaptation direction is applied for all cross-domain image pairs, which leads to identical supervision signals. To address this issue, we propose an Image-specific Prompt Learning (IPL) method, which learns specific prompt vectors for each source-domain image. This produces a more precise adaptation direction for every cross-domain image pair, endowing the target-domain generator with greatly enhanced flexibility. Qualitative and quantitative evaluations on various domains demonstrate that IPL effectively improves the quality and diversity of synthesized images and alleviates the mode collapse. Moreover, IPL is independent of the structure of the generative model, such as generative adversarial networks or diffusion models. Code is available at https://github.com/Picsart-AI-Research/IPL-Zero-Shot-Generative-Model-Adaptation.

Open-set Unsupervised Video Domain Adaptation (OUVDA) deals with the task of adapting an action recognition model from a labelled source domain to an unlabelled target domain that contains “target-private” categories, which are present in the target but absent in the source. In this work we deviate from the prior work of training a specialized open-set classifier or weighted adversarial learning by proposing to use pre-trained Language and Vision Models (CLIP). The CLIP is well suited for OUVDA due to its rich representation and the zero-shot recognition capabilities. However, rejecting target-private instances with the CLIP’s zero-shot protocol requires oracle knowledge about the target-private label names. To circumvent the impossibility of the knowledge of label names, we propose AutoLabel that automatically discovers and generates object-centric compositional candidate target-private class names. Despite its simplicity, we show that CLIP when equipped with AutoLabel can satisfactorily reject the target-private instances, thereby facilitating better alignment between the shared classes of the two domains. The code is available.

In semi-supervised medical image segmentation, there exist empirical mismatch problems between labeled and unlabeled data distribution. The knowledge learned from the labeled data may be largely discarded if treating labeled and unlabeled data separately or training labeled and unlabeled data in an inconsistent manner. We propose a straightforward method for alleviating the problem -- copy-pasting labeled and unlabeled data bidirectionally, in a simple Mean Teacher architecture. The method encourages unlabeled data to learn comprehensive common semantics from the labeled data in both inward and outward directions. More importantly, the consistent learning procedure for labeled and unlabeled data can largely reduce the empirical distribution gap. In detail, we copy-paste a random crop from a labeled image (foreground) onto an unlabeled image (background) and an unlabeled image (foreground) onto a labeled image (background), respectively. The two mixed images are fed into a Student network. It is trained by the generated supervisory signal via bidirectional copy-pasting between the predictions of the unlabeled images from the Teacher and the label maps of the labeled images. We explore several design choices of how to copy-paste to make it more effective for minimizing empirical distribution gaps between labeled and unlabeled data. We reveal that the simple mechanism of copy-pasting bidirectionally between labeled and unlabeled data is good enough and the experiments show solid gains (e.g., over 21% Dice improvement on ACDC dataset with 5% labeled data) compared with other state-of-the-arts on various semi-supervised medical image segmentation datasets.

Anatomical consistency in biomarker segmentation is crucial for many medical image analysis tasks. A promising paradigm for achieving anatomically consistent segmentation via deep networks is incorporating pixel connectivity, a basic concept in digital topology, to model inter-pixel relationships. However, previous works on connectivity modeling have ignored the rich channel-wise directional information in the latent space. In this work, we demonstrate that effective disentanglement of directional sub-space from the shared latent space can significantly enhance the feature representation in the connectivity-based network. To this end, we propose a directional connectivity modeling scheme for segmentation that decouples, tracks, and utilizes the directional information across the network. Experiments on various public medical image segmentation benchmarks show the effectiveness of our model as compared to the state-of-the-art methods. Code is available at https://github.com/Zyun-Y/DconnNet.

Collective insights from a group of experts have always proven to outperform an individual’s best diagnostic for clinical tasks. For the task of medical image segmentation, existing research on AI-based alternatives focuses more on developing models that can imitate the best individual rather than harnessing the power of expert groups. In this paper, we introduce a single diffusion model-based approach that produces multiple plausible outputs by learning a distribution over group insights. Our proposed model generates a distribution of segmentation masks by leveraging the inherent stochastic sampling process of diffusion using only minimal additional learning. We demonstrate on three different medical image modalities- CT, ultrasound, and MRI that our model is capable of producing several possible variants while capturing the frequencies of their occurrences. Comprehensive results show that our proposed approach outperforms existing state-of-the-art ambiguous segmentation networks in terms of accuracy while preserving naturally occurring variation. We also propose a new metric to evaluate the diversity as well as the accuracy of segmentation predictions that aligns with the interest of clinical practice of collective insights. Implementation code will be released publicly after the review process.

Processing giga-pixel whole slide histopathology images (WSI) is a computationally expensive task. Multiple instance learning (MIL) has become the conventional approach to process WSIs, in which these images are split into smaller patches for further processing. However, MIL-based techniques ignore explicit information about the individual cells within a patch. In this paper, by defining the novel concept of shared-context processing, we designed a multi-modal Graph Transformer that uses the cellular graph within the tissue to provide a single representation for a patient while taking advantage of the hierarchical structure of the tissue, enabling a dynamic focus between cell-level and tissue-level information. We benchmarked the performance of our model against multiple state-of-the-art methods in survival prediction and showed that ours can significantly outperform all of them including hierarchical vision Transformer (ViT). More importantly, we show that our model is strongly robust to missing information to an extent that it can achieve the same performance with as low as 20% of the data. Finally, in two different cancer datasets, we demonstrated that our model was able to stratify the patients into low-risk and high-risk groups while other state-of-the-art methods failed to achieve this goal. We also publish a large dataset of immunohistochemistry (IHC) images containing 1,600 tissue microarray (TMA) cores from 188 patients along with their survival information, making it one of the largest publicly available datasets in this context.

In clinical scenarios, multi-specialist consultation could significantly benefit the diagnosis, especially for intricate cases. This inspires us to explore a “multi-expert joint diagnosis” mechanism to upgrade the existing “single expert” framework commonly seen in the current literature. To this end, we propose METransformer, a method to realize this idea with a transformer-based backbone. The key design of our method is the introduction of multiple learnable “expert” tokens into both the transformer encoder and decoder. In the encoder, each expert token interacts with both vision tokens and other expert tokens to learn to attend different image regions for image representation. These expert tokens are encouraged to capture complementary information by an orthogonal loss that minimizes their overlap. In the decoder, each attended expert token guides the cross-attention between input words and visual tokens, thus influencing the generated report. A metrics-based expert voting strategy is further developed to generate the final report. By the multi-experts concept, our model enjoys the merits of an ensemble-based approach but through a manner that is computationally more efficient and supports more sophisticated interactions among experts. Experimental results demonstrate the promising performance of our proposed model on two widely used benchmarks. Last but not least, the framework-level innovation makes our work ready to incorporate advances on existing “single-expert” models to further improve its performance.

Deep neural networks have demonstrated promising performance on image recognition tasks. However, they may heavily rely on confounding factors, using irrelevant artifacts or bias within the dataset as the cue to improve performance. When a model performs decision-making based on these spurious correlations, it can become untrustable and lead to catastrophic outcomes when deployed in the real-world scene. In this paper, we explore and try to solve this problem in the context of skin cancer diagnosis. We introduce a human-in-the-loop framework in the model training process such that users can observe and correct the model’s decision logic when confounding behaviors happen. Specifically, our method can automatically discover confounding factors by analyzing the co-occurrence behavior of the samples. It is capable of learning confounding concepts using easily obtained concept exemplars. By mapping the blackbox model’s feature representation onto an explainable concept space, human users can interpret the concept and intervene via first order-logic instruction. We systematically evaluate our method on our newly crafted, well-controlled skin lesion dataset and several public skin lesion datasets. Experiments show that our method can effectively detect and remove confounding factors from datasets without any prior knowledge about the category distribution and does not require fully annotated concept labels. We also show that our method enables the model to focus on clinicalrelated concepts, improving the model’s performance and trustworthiness during model inference.

The core of out-of-distribution (OOD) detection is to learn the in-distribution (ID) representation, which is distinguishable from OOD samples. Previous work applied recognition-based methods to learn the ID features, which tend to learn shortcuts instead of comprehensive representations. In this work, we find surprisingly that simply using reconstruction-based methods could boost the performance of OOD detection significantly. We deeply explore the main contributors of OOD detection and find that reconstruction-based pretext tasks have the potential to provide a generally applicable and efficacious prior, which benefits the model in learning intrinsic data distributions of the ID dataset. Specifically, we take Masked Image Modeling as a pretext task for our OOD detection framework (MOOD). Without bells and whistles, MOOD outperforms previous SOTA of one-class OOD detection by 5.7%, multi-class OOD detection by 3.0%, and near-distribution OOD detection by 2.1%. It even defeats the 10-shot-per-class outlier exposure OOD detection, although we do not include any OOD samples for our detection.

Existing multi-view representation learning methods typically follow a specific-to-uniform pipeline, extracting latent features from each view and then fusing or aligning them to obtain the unified object representation. However, the manually pre-specified fusion functions and aligning criteria could potentially degrade the quality of the derived representation. To overcome them, we propose a novel uniform-to-specific multi-view learning framework from a meta-learning perspective, where the unified representation no longer involves manual manipulation but is automatically derived from a meta-learner named MetaViewer. Specifically, we formulated the extraction and fusion of view-specific latent features as a nested optimization problem and solved it by using a bi-level optimization scheme. In this way, MetaViewer automatically fuses view-specific features into a unified one and learns the optimal fusion scheme by observing reconstruction processes from the unified to the specific over all views. Extensive experimental results in downstream classification and clustering tasks demonstrate the efficiency and effectiveness of the proposed method.

The success of existing multi-view clustering relies on the assumption of sample integrity across multiple views. However, in real-world scenarios, samples of multi-view are partially available due to data corruption or sensor failure, which leads to incomplete multi-view clustering study (IMVC). Although several attempts have been proposed to address IMVC, they suffer from the following drawbacks: i) Existing methods mainly adopt cross-view contrastive learning forcing the representations of each sample across views to be exactly the same, which might ignore view discrepancy and flexibility in representations; ii) Due to the absence of non-observed samples across multiple views, the obtained prototypes of clusters might be unaligned and biased, leading to incorrect fusion. To address the above issues, we propose a Cross-view Partial Sample and Prototype Alignment Network (CPSPAN) for Deep Incomplete Multi-view Clustering. Firstly, unlike existing contrastive-based methods, we adopt pair-observed data alignment as ‘proxy supervised signals’ to guide instance-to-instance correspondence construction among views. Then, regarding of the shifted prototypes in IMVC, we further propose a prototype alignment module to achieve incomplete distribution calibration across views. Extensive experimental results showcase the effectiveness of our proposed modules, attaining noteworthy performance improvements when compared to existing IMVC competitors on benchmark datasets.

Recently, with the advent of Metaverse and AI Generated Content, cross-modal retrieval becomes popular with a burst of 2D and 3D data. However, this problem is challenging given the heterogeneous structure and semantic discrepancies. Moreover, imperfect annotations are ubiquitous given the ambiguous 2D and 3D content, thus inevitably producing noisy labels to degrade the learning performance. To tackle the problem, this paper proposes a robust 2D-3D retrieval framework (RONO) to robustly learn from noisy multimodal data. Specifically, one novel Robust Discriminative Center Learning mechanism (RDCL) is proposed in RONO to adaptively distinguish clean and noisy samples for respectively providing them with positive and negative optimization directions, thus mitigating the negative impact of noisy labels. Besides, we present a Shared Space Consistency Learning mechanism (SSCL) to capture the intrinsic information inside the noisy data by minimizing the cross-modal and semantic discrepancy between common space and label space simultaneously. Comprehensive mathematical analyses are given to theoretically prove the noise tolerance of the proposed method. Furthermore, we conduct extensive experiments on four 3D-model multimodal datasets to verify the effectiveness of our method by comparing it with 15 state-of-the-art methods. Code is available at https://github.com/penghu-cs/RONO.

Out-of-distribution (OOD) generalization is an important issue for Graph Neural Networks (GNNs). Recent works employ different graph editions to generate augmented environments and learn an invariant GNN for generalization. However, the graph structural edition inevitably alters the graph label. This causes the label shift in augmentations and brings inconsistent predictive relationships among augmented environments. To address this issue, we propose LiSA, which generates label-invariant augmentations to facilitate graph OOD generalization. Instead of resorting to graph editions, LiSA exploits Label-invariant Subgraphs of the training graphs to construct Augmented environments. Specifically, LiSA first designs the variational subgraph generators to efficiently extract locally predictive patterns and construct multiple label-invariant subgraphs. Then, the subgraphs produced by different generators are collected to build different augmented environments. To promote diversity among augmented environments, LiSA further introduces a tractable energy-based regularization to enlarge pair-wise distances between the distributions of environments. In this manner, LiSA generates diverse augmented environments with a consistent predictive relationship to facilitate learning an invariant GNN. Extensive experiments on node-level and graph-level OOD benchmarks show that LiSA achieves impressive generalization performance with different GNN backbones. Code is available on https://github.com/Samyu0304/LiSA.

When it comes to deploying deep vision models, the behavior of these systems must be explicable to ensure confidence in their reliability and fairness. A common approach to evaluate deep learning models is to build a labeled test set with attributes of interest and assess how well it performs. However, creating a balanced test set (i.e., one that is uniformly sampled over all the important traits) is often time-consuming, expensive, and prone to mistakes. The question we try to address is: can we evaluate the sensitivity of deep learning models to arbitrary visual attributes without an annotated test set? This paper argues the case that Zero-shot Model Diagnosis (ZOOM) is possible without the need for a test set nor labeling. To avoid the need for test sets, our system relies on a generative model and CLIP. The key idea is enabling the user to select a set of prompts (relevant to the problem) and our system will automatically search for semantic counterfactual images (i.e., synthesized images that flip the prediction in the case of a binary classifier) using the generative model. We evaluate several visual tasks (classification, key-point detection, and segmentation) in multiple visual domains to demonstrate the viability of our methodology. Extensive experiments demonstrate that our method is capable of producing counterfactual images and offering sensitivity analysis for model diagnosis without the need for a test set.

Confidence-based pseudo-labeling is among the dominant approaches in semi-supervised learning (SSL). It relies on including high-confidence predictions made on unlabeled data as additional targets to train the model. We propose ProtoCon, a novel SSL method aimed at the less-explored label-scarce SSL where such methods usually underperform. ProtoCon refines the pseudo-labels by leveraging their nearest neighbours’ information. The neighbours are identified as the training proceeds using an online clustering approach operating in an embedding space trained via a prototypical loss to encourage well-formed clusters. The online nature of ProtoCon allows it to utilise the label history of the entire dataset in one training cycle to refine labels in the following cycle without the need to store image embeddings. Hence, it can seamlessly scale to larger datasets at a low cost. Finally, ProtoCon addresses the poor training signal in the initial phase of training (due to fewer confident predictions) by introducing an auxiliary self-supervised loss. It delivers significant gains and faster convergence over state-of-the-art across 5 datasets, including CIFARs, ImageNet and DomainNet.

Learning with noisy labels (LNL) aims to ensure model generalization given a label-corrupted training set. In this work, we investigate a rarely studied scenario of LNL on fine-grained datasets (LNL-FG), which is more practical and challenging as large inter-class ambiguities among fine-grained classes cause more noisy labels. We empirically show that existing methods that work well for LNL fail to achieve satisfying performance for LNL-FG, arising the practical need of effective solutions for LNL-FG. To this end, we propose a novel framework called stochastic noise-tolerated supervised contrastive learning (SNSCL) that confronts label noise by encouraging distinguishable representation. Specifically, we design a noise-tolerated supervised contrastive learning loss that incorporates a weight-aware mechanism for noisy label correction and selectively updating momentum queue lists. By this mechanism, we mitigate the effects of noisy anchors and avoid inserting noisy labels into the momentum-updated queue. Besides, to avoid manually-defined augmentation strategies in contrastive learning, we propose an efficient stochastic module that samples feature embeddings from a generated distribution, which can also enhance the representation ability of deep models. SNSCL is general and compatible with prevailing robust LNL strategies to improve their performance for LNL-FG. Extensive experiments demonstrate the effectiveness of SNSCL.

Learning from noisy data is a challenging task that significantly degenerates the model performance. In this paper, we present TCL, a novel twin contrastive learning model to learn robust representations and handle noisy labels for classification. Specifically, we construct a Gaussian mixture model (GMM) over the representations by injecting the supervised model predictions into GMM to link label-free latent variables in GMM with label-noisy annotations. Then, TCL detects the examples with wrong labels as the out-of-distribution examples by another two-component GMM, taking into account the data distribution. We further propose a cross-supervision with an entropy regularization loss that bootstraps the true targets from model predictions to handle the noisy labels. As a result, TCL can learn discriminative representations aligned with estimated labels through mixup and contrastive learning. Extensive experimental results on several standard benchmarks and real-world datasets demonstrate the superior performance of TCL. In particular, TCL achieves 7.5% improvements on CIFAR-10 with 90% noisy label---an extremely noisy scenario. The source code is available at https://github.com/Hzzone/TCL.

Visual Question Answering models have been shown to suffer from language biases, where the model learns a correlation between the question and the answer, ignoring the image. While early works attempted to use question-only models or data augmentations to reduce this bias, we propose an adaptive margin loss approach having two components. The first component considers the frequency of answers within a question type in the training data, which addresses the concern of the class-imbalance causing the language biases. However, it does not take into account the answering difficulty of the samples, which impacts their learning. We address this through the second component, where instance-specific margins are learnt, allowing the model to distinguish between samples of varying complexity. We introduce a bias-injecting component to our model, and compute the instance-specific margins from the confidence of this component. We combine these with the estimated margins to consider both answer-frequency and task-complexity in the training loss. We show that, while the margin loss is effective for out-of-distribution (ood) data, the bias-injecting component is essential for generalising to in-distribution (id) data. Our proposed approach, Robust Margin Loss for Visual Question Answering (RMLVQA) improves upon the existing state-of-the-art results when compared to augmentation-free methods on benchmark VQA datasets suffering from language biases, while maintaining competitive performance on id data, making our method the most robust one among all comparable methods.

The task of Visual Question Answering (VQA) is known to be plagued by the issue of VQA models exploiting biases within the dataset to make its final prediction. Various previous ensemble based debiasing methods have been proposed where an additional model is purposefully trained to be biased in order to train a robust target model. However, these methods compute the bias for a model simply from the label statistics of the training data or from single modal branches. In this work, in order to better learn the bias a target VQA model suffers from, we propose a generative method to train the bias model directly from the target model, called GenB. In particular, GenB employs a generative network to learn the bias in the target model through a combination of the adversarial objective and knowledge distillation. We then debias our target model with GenB as a bias model, and show through extensive experiments the effects of our method on various VQA bias datasets including VQA-CP2, VQA-CP1, GQA-OOD, and VQA-CE, and show state-of-the-art results with the LXMERT architecture on VQA-CP2.

Although machines have surpassed humans on visual recognition problems, they are still limited to providing closed-set answers. Unlike machines, humans can cognize novel categories at the first observation. Novel category discovery (NCD) techniques, transferring knowledge from seen categories to distinguish unseen categories, aim to bridge the gap. However, current NCD methods assume a transductive learning and offline inference paradigm, which restricts them to a pre-defined query set and renders them unable to deliver instant feedback. In this paper, we study on-the-fly category discovery (OCD) aimed at making the model instantaneously aware of novel category samples (i.e., enabling inductive learning and streaming inference). We first design a hash coding-based expandable recognition model as a practical baseline. Afterwards, noticing the sensitivity of hash codes to intra-category variance, we further propose a novel Sign-Magnitude dIsentangLEment (SMILE) architecture to alleviate the disturbance it brings. Our experimental results demonstrate the superiority of SMILE against our baseline model and prior art. Our code will be made publicly available. Our code is available at https://github.com/PRIS-CV/On-the-fly-Category-Discovery.

This paper introduces submodel co-training, a regularization method related to co-training, self-distillation and stochastic depth. Given a neural network to be trained, for each sample we implicitly instantiate two altered networks, “submodels”, with stochastic depth: i.e. activating only a subset of the layers and skipping others. Each network serves as a soft teacher to the other, by providing a cross-entropy loss that complements the regular softmax cross-entropy loss provided by the one-hot label. Our approach, dubbed “cosub”, uses a single set of weights, and does not involve a pre-trained external model or temporal averaging. Experimentally, we show that submodel co-training is effective to train backbones for recognition tasks such as image classification and semantic segmentation, and that our approach is compatible with multiple recent architectures, including RegNet, PiT, and Swin. We report new state-of-the-art results for vision transformers trained on ImageNet only. For instance, a ViT-B pre-trained with cosub on Imagenet-21k achieves 87.4% top-1 acc. on Imagenet-val.

This work presents two astonishing findings on neural networks learned for large-scale image classification. 1) Given a well-trained model, the logits predicted for some category can be directly obtained by linearly combining the predictions of a few other categories, which we call neural dependency. 2) Neural dependencies exist not only within a single model, but even between two independently learned models, regardless of their architectures. Towards a theoretical analysis of such phenomena, we demonstrate that identifying neural dependencies is equivalent to solving the Covariance Lasso (CovLasso) regression problem proposed in this paper. Through investigating the properties of the problem solution, we confirm that neural dependency is guaranteed by a redundant logit covariance matrix, which condition is easily met given massive categories, and that neural dependency is sparse, which implies one category relates to only a few others. We further empirically show the potential of neural dependencies in understanding internal data correlations, generalizing models to unseen categories, and improving model robustness with a dependency-derived regularize. Code to exactly reproduce the results in this work will be released publicly.

In unsupervised domain adaptation (UDA), a model trained on source data (e.g. synthetic) is adapted to target data (e.g. real-world) without access to target annotation. Most previous UDA methods struggle with classes that have a similar visual appearance on the target domain as no ground truth is available to learn the slight appearance differences. To address this problem, we propose a Masked Image Consistency (MIC) module to enhance UDA by learning spatial context relations of the target domain as additional clues for robust visual recognition. MIC enforces the consistency between predictions of masked target images, where random patches are withheld, and pseudo-labels that are generated based on the complete image by an exponential moving average teacher. To minimize the consistency loss, the network has to learn to infer the predictions of the masked regions from their context. Due to its simple and universal concept, MIC can be integrated into various UDA methods across different visual recognition tasks such as image classification, semantic segmentation, and object detection. MIC significantly improves the state-of-the-art performance across the different recognition tasks for synthetic-to-real, day-to-nighttime, and clear-to-adverse-weather UDA. For instance, MIC achieves an unprecedented UDA performance of 75.9 mIoU and 92.8% on GTA-to-Cityscapes and VisDA-2017, respectively, which corresponds to an improvement of +2.1 and +3.0 percent points over the previous state of the art. The implementation is available at https://github.com/lhoyer/MIC.

Unsupervised domain adaptation (UDA) in semantic segmentation transfers the knowledge of the source domain to the target one to improve the adaptability of the segmentation model in the target domain. The need to access labeled source data makes UDA unable to handle adaptation scenarios involving privacy, property rights protection, and confidentiality. In this paper, we focus on unsupervised model adaptation (UMA), also called source-free domain adaptation, which adapts a source-trained model to the target domain without accessing source data. We find that the online self-training method has the potential to be deployed in UMA, but the lack of source domain loss will greatly weaken the stability and adaptability of the method. We analyze the two possible reasons for the degradation of online self-training, i.e. inopportune updates of the teacher model and biased knowledge from source-trained model. Based on this, we propose a dynamic teacher update mechanism and a training-consistency based resampling strategy to improve the stability and adaptability of online self training. On multiple model adaptation benchmarks, our method obtains new state-of-the-art performance, which is comparable or even better than state-of-the-art UDA methods.

Unsupervised Domain Adaptation Regression (DAR) aims to bridge the domain gap between a labeled source dataset and an unlabelled target dataset for regression problems. Recent works mostly focus on learning a deep feature encoder by minimizing the discrepancy between source and target features. In this work, we present a different perspective for the DAR problem by analyzing the closed-form ordinary least square (OLS) solution to the linear regressor in the deep domain adaptation context. Rather than aligning the original feature embedding space, we propose to align the inverse Gram matrix of the features, which is motivated by its presence in the OLS solution and the Gram matrix’s ability to capture the feature correlations. Specifically, we propose a simple yet effective DAR method which leverages the pseudo-inverse low-rank property to align the scale and angle in a selected subspace generated by the pseudo-inverse Gram matrix of the two domains. We evaluate our method on three domain adaptation regression benchmarks. Experimental results demonstrate that our method achieves state-of-the-art performance. Our code is available at https://github.com/ismailnejjar/DARE-GRAM.

We propose Equiangular Basis Vectors~(EBVs) for classification tasks. In deep neural networks, models usually end with a k-way fully connected layer with softmax to handle different classification tasks. The learning objective of these methods can be summarized as mapping the learned feature representations to the samples’ label space. While in metric learning approaches, the main objective is to learn a transformation function that maps training data points from the original space to a new space where similar points are closer while dissimilar points become farther apart. Different from previous methods, our EBVs generate normalized vector embeddings as “predefined classifiers” which are required to not only be with the equal status between each other, but also be as orthogonal as possible. By minimizing the spherical distance of the embedding of an input between its categorical EBV in training, the predictions can be obtained by identifying the categorical EBV with the smallest distance during inference. Various experiments on the ImageNet-1K dataset and other downstream tasks demonstrate that our method outperforms the general fully connected classifier while it does not introduce huge additional computation compared with classical metric learning methods. Our EBVs won the first place in the 2022 DIGIX Global AI Challenge, and our code is open-source and available at https://github.com/NJUST-VIPGroup/Equiangular-Basis-Vectors.

This paper explores the tasks of leveraging auxiliary modalities which are only available at training to enhance multimodal representation learning through cross-modal Knowledge Distillation (KD). The widely adopted mutual information maximization-based objective leads to a short-cut solution of the weak teacher, i.e., achieving the maximum mutual information by simply making the teacher model as weak as the student model. To prevent such a weak solution, we introduce an additional objective term, i.e., the mutual information between the teacher and the auxiliary modality model. Besides, to narrow down the information gap between the student and teacher, we further propose to minimize the conditional entropy of the teacher given the student. Novel training schemes based on contrastive learning and adversarial learning are designed to optimize the mutual information and the conditional entropy, respectively. Experimental results on three popular multimodal benchmark datasets have shown that the proposed method outperforms a range of state-of-the-art approaches for video recognition, video retrieval and emotion classification.

Domain generalization (DG) is a principal task to evaluate the robustness of computer vision models. Many previous studies have used normalization for DG. In normalization, statistics and normalized features are regarded as style and content, respectively. However, it has a content variation problem when removing style because the boundary between content and style is unclear. This study addresses this problem from the frequency domain perspective, where amplitude and phase are considered as style and content, respectively. First, we verify the quantitative phase variation of normalization through the mathematical derivation of the Fourier transform formula. Then, based on this, we propose a novel normalization method, PCNorm, which eliminates style only as the preserving content through spectral decomposition. Furthermore, we propose advanced PCNorm variants, CCNorm and SCNorm, which adjust the degrees of variations in content and style, respectively. Thus, they can learn domain-agnostic representations for DG. With the normalization methods, we propose ResNet-variant models, DAC-P and DAC-SC, which are robust to the domain gap. The proposed models outperform other recent DG methods. The DAC-SC achieves an average state-of-the-art performance of 65.6% on five datasets: PACS, VLCS, Office-Home, DomainNet, and TerraIncognita.

Test-time adaptation harnesses test inputs to improve the accuracy of a model trained on source data when tested on shifted target data. Most methods update the source model by (re-)training on each target domain. While re-training can help, it is sensitive to the amount and order of the data and the hyperparameters for optimization. We update the target data instead, and project all test inputs toward the source domain with a generative diffusion model. Our diffusion-driven adaptation (DDA) method shares its models for classification and generation across all domains, training both on source then freezing them for all targets, to avoid expensive domain-wise re-training. We augment diffusion with image guidance and classifier self-ensembling to automatically decide how much to adapt. Input adaptation by DDA is more robust than model adaptation across a variety of corruptions, models, and data regimes on the ImageNet-C benchmark. With its input-wise updates, DDA succeeds where model adaptation degrades on too little data (small batches), on dependent data (correlated orders), or on mixed data (multiple corruptions).

Improving the generalization ability of Deep Neural Networks (DNNs) is critical for their practical uses, which has been a longstanding challenge. Some theoretical studies have uncovered that DNNs have preferences for some frequency components in the learning process and indicated that this may affect the robustness of learned features. In this paper, we propose Deep Frequency Filtering (DFF) for learning domain-generalizable features, which is the first endeavour to explicitly modulate the frequency components of different transfer difficulties across domains in the latent space during training. To achieve this, we perform Fast Fourier Transform (FFT) for the feature maps at different layers, then adopt a light-weight module to learn attention masks from the frequency representations after FFT to enhance transferable components while suppressing the components not conducive to generalization. Further, we empirically compare the effectiveness of adopting different types of attention designs for implementing DFF. Extensive experiments demonstrate the effectiveness of our proposed DFF and show that applying our DFF on a plain baseline outperforms the state-of-the-art methods on different domain generalization tasks, including close-set classification and open-set retrieval.

Despite recent advances in implicit neural representations (INRs), it remains challenging for a coordinate-based multi-layer perceptron (MLP) of INRs to learn a common representation across data instances and generalize it for unseen instances. In this work, we introduce a simple yet effective framework for generalizable INRs that enables a coordinate-based MLP to represent complex data instances by modulating only a small set of weights in an early MLP layer as an instance pattern composer; the remaining MLP weights learn pattern composition rules to learn common representations across instances. Our generalizable INR framework is fully compatible with existing meta-learning and hypernetworks in learning to predict the modulated weight for unseen instances. Extensive experiments demonstrate that our method achieves high performance on a wide range of domains such as an audio, image, and 3D object, while the ablation study validates our weight modulation.

We study the problem of how to train a “personalization-friendly” model such that given only the task descriptions, the model can be adapted to different end-users’ needs, e.g., for accurately classifying different subsets of objects. One baseline approach is to train a “generic” model for classifying a wide range of objects, followed by class selection. In our experiments, we however found it suboptimal, perhaps because the model’s weights are kept frozen without being personalized. To address this drawback, we propose Train-once-for-All PERsonalization (TAPER), a framework that is trained just once and can later customize a model for different end-users given their task descriptions. TAPER learns a set of “basis” models and a mixer predictor, such that given the task description, the weights (not the predictions!) of the basis models can be on the fly combined into a single “personalized” model. Via extensive experiments on multiple recognition tasks, we show that TAPER consistently outperforms the baseline methods in achieving a higher personalized accuracy. Moreover, we show that TAPER can synthesize a much smaller model to achieve comparable performance to a huge generic model, making it “deployment-friendly” to resource-limited end devices. Interestingly, even without end-users’ task descriptions, TAPER can still be specialized to the deployed context based on its past predictions, making it even more “personalization-friendly”.

Optimization in multi-task learning (MTL) is more challenging than single-task learning (STL), as the gradient from different tasks can be contradictory. When tasks are related, it can be beneficial to share some parameters among them (cooperation). However, some tasks require additional parameters with expertise in a specific type of data or discrimination (specialization). To address the MTL challenge, we propose Mod-Squad, a new model that is Modularized into groups of experts (a ‘Squad’). This structure allows us to formalize cooperation and specialization as the process of matching experts and tasks. We optimize this matching process during the training of a single model. Specifically, we incorporate mixture of experts (MoE) layers into a transformer model, with a new loss that incorporates the mutual dependence between tasks and experts. As a result, only a small set of experts are activated for each task. This prevents the sharing of the entire backbone model between all tasks, which strengthens the model, especially when the training set size and the number of tasks scale up. More interestingly, for each task, we can extract the small set of experts as a standalone model that maintains the same performance as the large model. Extensive experiments on the Taskonomy dataset with 13 vision tasks and the PASCALContext dataset with 5 vision tasks show the superiority of our approach. The project page can be accessed at https://vis-www.cs.umass.edu/mod-squad.

Few-Shot Class-Incremental Learning (FSCIL) aims to continually learn novel classes based on only few training samples, which poses a more challenging task than the well-studied Class-Incremental Learning (CIL) due to data scarcity. While knowledge distillation, a prevailing technique in CIL, can alleviate the catastrophic forgetting of older classes by regularizing outputs between current and previous model, it fails to consider the overfitting risk of novel classes in FSCIL. To adapt the powerful distillation technique for FSCIL, we propose a novel distillation structure, by taking the unique challenge of overfitting into account. Concretely, we draw knowledge from two complementary teachers. One is the model trained on abundant data from base classes that carries rich general knowledge, which can be leveraged for easing the overfitting of current novel classes. The other is the updated model from last incremental session that contains the adapted knowledge of previous novel classes, which is used for alleviating their forgetting. To combine the guidances, an adaptive strategy conditioned on the class-wise semantic similarities is introduced. Besides, for better preserving base class knowledge when accommodating novel concepts, we adopt a two-branch network with an attention-based aggregation module to dynamically merge predictions from two complementary branches. Extensive experiments on 3 popular FSCIL datasets: mini-ImageNet, CIFAR100 and CUB200 validate the effectiveness of our method by surpassing existing works by a significant margin.

Self-supervised learning (SSL) has made remarkable progress in visual representation learning. Some studies combine SSL with knowledge distillation (SSL-KD) to boost the representation learning performance of small models. In this study, we propose a Multi-mode Online Knowledge Distillation method (MOKD) to boost self-supervised visual representation learning. Different from existing SSL-KD methods that transfer knowledge from a static pre-trained teacher to a student, in MOKD, two different models learn collaboratively in a self-supervised manner. Specifically, MOKD consists of two distillation modes: self-distillation and cross-distillation modes. Among them, self-distillation performs self-supervised learning for each model independently, while cross-distillation realizes knowledge interaction between different models. In cross-distillation, a cross-attention feature search strategy is proposed to enhance the semantic feature alignment between different models. As a result, the two models can absorb knowledge from each other to boost their representation learning performance. Extensive experimental results on different backbones and datasets demonstrate that two heterogeneous models can benefit from MOKD and outperform their independently trained baseline. In addition, MOKD also outperforms existing SSL-KD methods for both the student and teacher models.

The problem of class incremental learning (CIL) is considered. State-of-the-art approaches use a dynamic architecture based on network expansion (NE), in which a task expert is added per task. While effective from a computational standpoint, these methods lead to models that grow quickly with the number of tasks. A new NE method, dense network expansion (DNE), is proposed to achieve a better trade-off between accuracy and model complexity. This is accomplished by the introduction of dense connections between the intermediate layers of the task expert networks, that enable the transfer of knowledge from old to new tasks via feature sharing and reusing. This sharing is implemented with a cross-task attention mechanism, based on a new task attention block (TAB), that fuses information across tasks. Unlike traditional attention mechanisms, TAB operates at the level of the feature mixing and is decoupled with spatial attentions. This is shown more effective than a joint spatial-and-task attention for CIL. The proposed DNE approach can strictly maintain the feature space of old classes while growing the network and feature scale at a much slower rate than previous methods. In result, it outperforms the previous SOTA methods by a margin of 4% in terms of accuracy, with similar or even smaller model scale.

Previous knowledge distillation methods have shown their impressive performance on model compression tasks, however, it is hard to explain how the knowledge they transferred helps to improve the performance of the student network. In this work, we focus on proposing a knowledge distillation method that has both high interpretability and competitive performance. We first revisit the structure of mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions of input is critical for CNN to perform classification. Furthermore, we demonstrate that this capacity can be obtained and enhanced by transferring class activation maps. Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD). Different from previous KD methods, we explore and present several properties of the knowledge transferred by our method, which not only improve the interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Code is available at: https://github.com/GzyAftermath/CAT-KD.

Existing continual learning (CL) research regards catastrophic forgetting (CF) as almost the only challenge. This paper argues for another challenge in class-incremental learning (CIL), which we call cross-task class discrimination (CTCD),~i.e., how to establish decision boundaries between the classes of the new task and old tasks with no (or limited) access to the old task data.~CTCD is implicitly and partially dealt with by replay-based methods. A replay method saves a small amount of data (replay data) from previous tasks. When a batch of current task data arrives, the system jointly trains the new data and some sampled replay data. The replay data enables the system to partially learn the decision boundaries between the new classes and the old classes as the amount of the saved data is small. However, this paper argues that the replay approach also has a dynamic training bias issue which reduces the effectiveness of the replay data in solving the CTCD problem. A novel optimization objective with a gradient-based adaptive method is proposed to dynamically deal with the problem in the online CL process. Experimental results show that the new method achieves much better results in online CL.

Current evaluations of Continual Learning (CL) methods typically assume that there is no constraint on training time and computation. This is an unrealistic assumption for any real-world setting, which motivates us to propose: a practical real-time evaluation of continual learning, in which the stream does not wait for the model to complete training before revealing the next data for predictions. To do this, we evaluate current CL methods with respect to their computational costs. We conduct extensive experiments on CLOC, a large-scale dataset containing 39 million time-stamped images with geolocation labels. We show that a simple baseline outperforms state-of-the-art CL methods under this evaluation, questioning the applicability of existing methods in realistic settings. In addition, we explore various CL components commonly used in the literature, including memory sampling strategies and regularization approaches. We find that all considered methods fail to be competitive against our simple baseline. This surprisingly suggests that the majority of existing CL literature is tailored to a specific class of streams that is not practical. We hope that the evaluation we provide will be the first step towards a paradigm shift to consider the computational cost in the development of online continual learning methods.

Knowledge distillation (KD) is an effective training strategy to improve the lightweight student models under the guidance of cumbersome teachers. However, the large architecture difference across the teacher-student pairs limits the distillation gains. In contrast to previous adaptive distillation methods to reduce the teacher-student gap, we explore a novel training-free framework to search for the best student architectures for a given teacher. Our work first empirically show that the optimal model under vanilla training cannot be the winner in distillation. Secondly, we find that the similarity of feature semantics and sample relations between random-initialized teacher-student networks have good correlations with final distillation performances. Thus, we efficiently measure similarity matrixs conditioned on the semantic activation maps to select the optimal student via an evolutionary algorithm without any training. In this way, our student architecture search for Distillation WithOut Training (DisWOT) significantly improves the performance of the model in the distillation stage with at least 180× training acceleration. Additionally, we extend similarity metrics in DisWOT as new distillers and KD-based zero-proxies. Our experiments on CIFAR, ImageNet and NAS-Bench-201 demonstrate that our technique achieves state-of-the-art results on different search spaces. Our project and code are available at https://lilujunai.github.io/DisWOT-CVPR2023/.

Computer vision models suffer from a phenomenon known as catastrophic forgetting when learning novel concepts from continuously shifting training data. Typical solutions for this continual learning problem require extensive rehearsal of previously seen data, which increases memory costs and may violate data privacy. Recently, the emergence of large-scale pre-trained vision transformer models has enabled prompting approaches as an alternative to data-rehearsal. These approaches rely on a key-query mechanism to generate prompts and have been found to be highly resistant to catastrophic forgetting in the well-established rehearsal-free continual learning setting. However, the key mechanism of these methods is not trained end-to-end with the task sequence. Our experiments show that this leads to a reduction in their plasticity, hence sacrificing new task accuracy, and inability to benefit from expanded parameter capacity. We instead propose to learn a set of prompt components which are assembled with input-conditioned weights to produce input-conditioned prompts, resulting in a novel attention-based end-to-end key-query scheme. Our experiments show that we outperform the current SOTA method DualPrompt on established benchmarks by as much as 4.5% in average final accuracy. We also outperform the state of art by as much as 4.4% accuracy on a continual learning benchmark which contains both class-incremental and domain-incremental task shifts, corresponding to many practical settings. Our code is available at https://github.com/GT-RIPL/CODA-Prompt

This paper presents a simple yet effective approach that improves continual test-time adaptation (TTA) in a memory-efficient manner. TTA may primarily be conducted on edge devices with limited memory, so reducing memory is crucial but has been overlooked in previous TTA studies. In addition, long-term adaptation often leads to catastrophic forgetting and error accumulation, which hinders applying TTA in real-world deployments. Our approach consists of two components to address these issues. First, we present lightweight meta networks that can adapt the frozen original networks to the target domain. This novel architecture minimizes memory consumption by decreasing the size of intermediate activations required for backpropagation. Second, our novel self-distilled regularization controls the output of the meta networks not to deviate significantly from the output of the frozen original networks, thereby preserving well-trained knowledge from the source domain. Without additional memory, this regularization prevents error accumulation and catastrophic forgetting, resulting in stable performance even in long-term test-time adaptation. We demonstrate that our simple yet effective strategy outperforms other state-of-the-art methods on various benchmarks for image classification and semantic segmentation tasks. Notably, our proposed method with ResNet-50 and WideResNet-40 takes 86% and 80% less memory than the recent state-of-the-art method, CoTTA.

In contrast to the natural capabilities of humans to learn new tasks in a sequential fashion, neural networks are known to suffer from catastrophic forgetting, where the model’s performances on old tasks drop dramatically after being optimized for a new task. Since then, the continual learning (CL) community has proposed several solutions aiming to equip the neural network with the ability to learn the current task (plasticity) while still achieving high accuracy on the previous tasks (stability). Despite remarkable improvements, the plasticity-stability trade-off is still far from being solved, and its underlying mechanism is poorly understood. In this work, we propose Auxiliary Network Continual Learning (ANCL), a novel method that applies an additional auxiliary network which promotes plasticity to the continually learned model which mainly focuses on stability. More concretely, the proposed framework materializes in a regularizer that naturally interpolates between plasticity and stability, surpassing strong baselines on task incremental and class incremental scenarios. Through extensive analyses on ANCL solutions, we identify some essential principles beneath the stability-plasticity trade-off.

Based on the weight-sharing mechanism, one-shot NAS methods train a supernet and then inherit the pre-trained weights to evaluate sub-models, largely reducing the search cost. However, several works have pointed out that the shared weights suffer from different gradient descent directions during training. And we further find that large gradient variance occurs during supernet training, which degrades the supernet ranking consistency. To mitigate this issue, we propose to explicitly minimize the gradient variance of the supernet training by jointly optimizing the sampling distributions of PAth and DAta (PA&DA). We theoretically derive the relationship between the gradient variance and the sampling distributions, and reveal that the optimal sampling probability is proportional to the normalized gradient norm of path and training data. Hence, we use the normalized gradient norm as the importance indicator for path and training data, and adopt an importance sampling strategy for the supernet training. Our method only requires negligible computation cost for optimizing the sampling distributions of path and data, but achieves lower gradient variance during supernet training and better generalization performance for the supernet, resulting in a more consistent NAS. We conduct comprehensive comparisons with other improved approaches in various search spaces. Results show that our method surpasses others with more reliable ranking performance and higher accuracy of searched architectures, showing the effectiveness of our method. Code is available at https://github.com/ShunLu91/PA-DA.

Dataset Distillation (DD), a newly emerging field, aims at generating much smaller but efficient synthetic training datasets from large ones. Existing DD methods based on gradient matching achieve leading performance; however, they are extremely computationally intensive as they require continuously optimizing a dataset among thousands of randomly initialized models. In this paper, we assume that training the synthetic data with diverse models leads to better generalization performance. Thus we propose two model augmentation techniques, i.e. using early-stage models and parameter perturbation to learn an informative synthetic set with significantly reduced training cost. Extensive experiments demonstrate that our method achieves up to 20× speedup and comparable performance on par with state-of-the-art methods.

In this paper, we propose multi-agent automated machine learning (MA2ML) with the aim to effectively handle joint optimization of modules in automated machine learning (AutoML). MA2ML takes each machine learning module, such as data augmentation (AUG), neural architecture search (NAS), or hyper-parameters (HPO), as an agent and the final performance as the reward, to formulate a multi-agent reinforcement learning problem. MA2ML explicitly assigns credit to each agent according to its marginal contribution to enhance cooperation among modules, and incorporates off-policy learning to improve search efficiency. Theoretically, MA2ML guarantees monotonic improvement of joint optimization. Extensive experiments show that MA2ML yields the state-of-the-art top-1 accuracy on ImageNet under constraints of computational cost, e.g., 79.7%/80.5% with FLOPs fewer than 600M/800M. Extensive ablation studies verify the benefits of credit assignment and off-policy learning of MA2ML.

We propose a new approach to learned optimization where we represent the computation of an optimizer’s update step using a neural network. The parameters of the optimizer are then learned by training on a set of optimization tasks with the objective to perform minimization efficiently. Our innovation is a new neural network architecture, Optimus, for the learned optimizer inspired by the classic BFGS algorithm. As in BFGS, we estimate a preconditioning matrix as a sum of rank-one updates but use a Transformer-based neural network to predict these updates jointly with the step length and direction. In contrast to several recent learned optimization-based approaches, our formulation allows for conditioning across the dimensions of the parameter space of the target problem while remaining applicable to optimization tasks of variable dimensionality without retraining. We demonstrate the advantages of our approach on a benchmark composed of objective functions traditionally used for the evaluation of optimization algorithms, as well as on the real world-task of physics-based visual reconstruction of articulated 3d human motion.

We consider the problem of solving LP relaxations of MAP-MRF inference problems, and in particular the method proposed recently in (Swoboda, Kolmogorov 2019; Kolmogorov, Pock 2021). As a key computational subroutine, it uses a variant of the Frank-Wolfe (FW) method to minimize a smooth convex function over a combinatorial polytope. We propose an efficient implementation of this subproutine based on in-face Frank-Wolfe directions, introduced in (Freund et al. 2017) in a different context. More generally, we define an abstract data structure for a combinatorial subproblem that enables in-face FW directions, and describe its specialization for tree-structured MAP-MRF inference subproblems. Experimental results indicate that the resulting method is the current state-of-art LP solver for some classes of problems. Our code is available at pub.ist.ac.at/~vnk/papers/IN-FACE-FW.html.

Instead of searching the entire network directly, current NAS approaches increasingly search for multiple relatively small cells to reduce search costs. A major challenge is to jointly measure the similarity of cell micro-architectures and the difference in macro-architectures between different cell-based networks. Recently, optimal transport (OT) has been successfully applied to NAS as it can capture the operational and structural similarity across various networks. However, existing OT-based NAS methods either ignore the cell similarity or focus solely on searching for a single cell architecture. To address these issues, we propose a hierarchical optimal transport metric called HOTNN for measuring the similarity of different networks. In HOTNN, the cell-level similarity computes the OT distance between cells in various networks by considering the similarity of each node and the differences in the information flow costs between node pairs within each cell in terms of operational and structural information. The network-level similarity calculates OT distance between networks by considering both the cell-level similarity and the variation in the global position of each cell within their respective networks. We then explore HOTNN in a Bayesian optimization framework called HOTNAS, and demonstrate its efficacy in diverse tasks. Extensive experiments demonstrate that HOTNAS can discover network architectures with better performance in multiple modular cell-based search spaces.

The inverted index is a widely used data structure to avoid the infeasible exhaustive search. It accelerates retrieval significantly by splitting the database into multiple disjoint sets and restricts distance computation to a small fraction of the database. Moreover, it even improves search quality by allowing quantizers to exploit the compact distribution of residual vector space. However, we firstly point out a problem that an existing deep learning-based quantizer hardly benefits from the residual vector space, unlike conventional shallow quantizers. To cope with this problem, we introduce a novel disentangled representation learning for unsupervised neural quantization. Similar to the concept of residual vector space, the proposed method enables more compact latent space by disentangling information of the inverted index from the vectors. Experimental results on large-scale datasets confirm that our method outperforms the state-of-the-art retrieval systems by a large margin.

We present FFCV, a library for easy, fast, resource-efficient training of machine learning models. FFCV speeds up model training by eliminating (often subtle) data bottlenecks from the training process. In particular, we combine techniques such as an efficient file storage format, caching, data pre-loading, asynchronous data transfer, and just-in-time compilation to (a) make data loading and transfer significantly more efficient, ensuring that GPUs can reach full utilization; and (b) offload as much data processing as possible to the CPU asynchronously, freeing GPU up capacity for training. Using FFCV, we train ResNet-18 and ResNet-50 on the ImageNet dataset with a state-of-the-art tradeoff between accuracy and training time. For example, across the range of ResNet-50 models we test, we obtain the same accuracy as the best baselines in half the time. We demonstrate FFCV’s performance, ease-of-use, extensibility, and ability to adapt to resource constraints through several case studies.

To design fast neural networks, many works have been focusing on reducing the number of floating-point operations (FLOPs). We observe that such reduction in FLOPs, however, does not necessarily lead to a similar level of reduction in latency. This mainly stems from inefficiently low floating-point operations per second (FLOPS). To achieve faster networks, we revisit popular operators and demonstrate that such low FLOPS is mainly due to frequent memory access of the operators, especially the depthwise convolution. We hence propose a novel partial convolution (PConv) that extracts spatial features more efficiently, by cutting down redundant computation and memory access simultaneously. Building upon our PConv, we further propose FasterNet, a new family of neural networks, which attains substantially higher running speed than others on a wide range of devices, without compromising on accuracy for various vision tasks. For example, on ImageNet-1k, our tiny FasterNet-T0 is 2.8×, 3.3×, and 2.4× faster than MobileViT-XXS on GPU, CPU, and ARM processors, respectively, while being 2.9% more accurate. Our large FasterNet-L achieves impressive 83.5% top-1 accuracy, on par with the emerging Swin-B, while having 36% higher inference throughput on GPU, as well as saving 37% compute time on CPU. Code is available at https://github.com/JierunChen/FasterNet.

Generative DNNs are a powerful tool for image synthesis, but they are limited by their computational load. On the other hand, given a trained model and a task, e.g. faces generation within a range of characteristics, the output image quality will be unevenly distributed among images with different characteristics. It follows, that we might restrain the model’s complexity on some instances, maintaining a high quality. We propose a method for diminishing computations by adding so-called early exit branches to the original architecture, and dynamically switching the computational path depending on how difficult it will be to render the output. We apply our method on two different SOTA models performing generative tasks: generation from a semantic map, and cross reenactment of face expressions; showing it is able to output images with custom lower quality thresholds. For a threshold of LPIPS <=0.1, we diminish their computations by up to a half. This is especially relevant for real-time applications such as synthesis of faces, when quality loss needs to be contained, but most of the inputs need fewer computations than the complex instances.

Predictions made by deep learning models are prone to data perturbations, adversarial attacks, and out-of-distribution inputs. To build a trusted AI system, it is therefore critical to accurately quantify the prediction uncertainties. While current efforts focus on improving uncertainty quantification accuracy and efficiency, there is a need to identify uncertainty sources and take actions to mitigate their effects on predictions. Therefore, we propose to develop explainable and actionable Bayesian deep learning methods to not only perform accurate uncertainty quantification but also explain the uncertainties, identify their sources, and propose strategies to mitigate the uncertainty impacts. Specifically, we introduce a gradient-based uncertainty attribution method to identify the most problematic regions of the input that contribute to the prediction uncertainty. Compared to existing methods, the proposed UA-Backprop has competitive accuracy, relaxed assumptions, and high efficiency. Moreover, we propose an uncertainty mitigation strategy that leverages the attribution results as attention to further improve the model performance. Both qualitative and quantitative evaluations are conducted to demonstrate the effectiveness of our proposed methods.

Deep learning with noisy labels is challenging and inevitable in many circumstances. Existing methods reduce the impact of noise samples by reducing loss weights of uncertain samples or by filtering out potential noise samples, which highly rely on the model’s superior discriminative power for identifying noise samples. However, in the training stage, the trainee model is imperfect will miss many noise samples, which cause continuous damage to the model training. Consequently, there is a large performance gap between existing anti-noise models trained with noisy samples and models trained with clean samples. In this paper, we put forward a Gradient Switching Strategy (GSS) to prevent the continuous damage of noise samples to the classifier. Theoretical analysis shows that the damage comes from the misleading gradient direction computed from the noise samples. The trainee model will deviate from the correct optimization direction under the influence of the accumulated misleading gradient of noise samples. To address this problem, the proposed GSS alleviates the damage by switching the current gradient direction of each sample to a new direction selected from a gradient direction pool, which contains all-class gradient directions with different probabilities. During training, the trainee model is optimized along switched gradient directions generated by GSS, which assigns higher probabilities to potential principal directions for high-confidence samples. Conversely, uncertain samples have a relatively uniform probability distribution for all gradient directions, which can cancel out the misleading gradient directions. Extensive experiments show that a model trained with GSS can achieve comparable performance with a model trained with clean data. Moreover, the proposed GSS is pluggable for existing frameworks for noisy-label learning. This work can provide a new perspective for future noisy-label learning.

Zero-shot quantization is a promising approach for developing lightweight deep neural networks when data is inaccessible owing to various reasons, including cost and issues related to privacy. By exploiting the learned parameters (µ and sigma) of batch normalization layers in an FP32-pre-trained model, zero-shot quantization schemes focus on generating synthetic data. Subsequently, they distill knowledge from the pre-trained model (teacher) to the quantized model (student) such that the quantized model can be optimized with the synthetic dataset. However, thus far, zero-shot quantization has primarily been discussed in the context of quantization-aware training methods, which require task-specific losses and long-term optimization as much as retraining. We thus introduce a post-training quantization scheme for zero-shot quantization that produces high-quality quantized networks within a few hours. Furthermore, we propose a framework called GENIE that generates data suited for quantization. With the data synthesized by GENIE, we can produce robust quantized models without real datasets, which is comparable to few-shot quantization. We also propose a post-training quantization algorithm to enhance the performance of quantized models. By combining them, we can bridge the gap between zero-shot and few-shot quantization while significantly improving the quantization performance compared to that of existing approaches. In other words, we can obtain a unique state-of-the-art zero-shot quantization approach.

Reliable confidence estimation for deep neural classifiers is a challenging yet fundamental requirement in high-stakes applications. Unfortunately, modern deep neural networks are often overconfident for their erroneous predictions. In this work, we exploit the easily available outlier samples, i.e., unlabeled samples coming from non-target classes, for helping detect misclassification errors. Particularly, we find that the well-known Outlier Exposure, which is powerful in detecting out-of-distribution (OOD) samples from unknown classes, does not provide any gain in identifying misclassification errors. Based on these observations, we propose a novel method called OpenMix, which incorporates open-world knowledge by learning to reject uncertain pseudo-samples generated via outlier transformation. OpenMix significantly improves confidence reliability under various scenarios, establishing a strong and unified framework for detecting both misclassified samples from known classes and OOD samples from unknown classes.

Rising concerns about privacy and anonymity preservation of deep learning models have facilitated research in data-free learning. Primarily based on data-free knowledge distillation, models developed in this area so far have only been able to operate in a single modality, performing the same kind of task as that of the teacher. For the first time, we propose Data-Free Sketch-Based Image Retrieval (DF-SBIR), a cross-modal data-free learning setting, where teachers trained for classification in a single modality have to be leveraged by students to learn a cross-modal metric-space for retrieval. The widespread availability of pre-trained classification models, along with the difficulty in acquiring paired photo-sketch datasets for SBIR justify the practicality of this setting. We present a methodology for DF-SBIR, which can leverage knowledge from models independently trained to perform classification on photos and sketches. We evaluate our model on the Sketchy, TU-Berlin, and QuickDraw benchmarks, designing a variety of baselines based on existing data-free learning literature, and observe that our method surpasses all of them by significant margins. Our method also achieves mAPs competitive with data-dependent approaches, all the while requiring no training data. Implementation is available at https://github.com/abhrac/data-free-sbir.

Generative adversarial network (GAN) is formulated as a two-player game between a generator (G) and a discriminator (D), where D is asked to differentiate whether an image comes from real data or is produced by G. Under such a formulation, D plays as the rule maker and hence tends to dominate the competition. Towards a fairer game in GANs, we propose a new paradigm for adversarial training, which makes G assign a task to D as well. Specifically, given an image, we expect D to extract representative features that can be adequately decoded by G to reconstruct the input. That way, instead of learning freely, D is urged to align with the view of G for domain classification. Experimental results on various datasets demonstrate the substantial superiority of our approach over the baselines. For instance, we improve the FID of StyleGAN2 from 4.30 to 2.55 on LSUN Bedroom and from 4.04 to 2.82 on LSUN Church. We believe that the pioneering attempt present in this work could inspire the community with better designed generator-leading tasks for GAN improvement. Project page is at https://ezioby.github.io/glead/.

Recently, there has been a significant advancement in image generation technology, known as GAN. It can easily generate realistic fake images, leading to an increased risk of abuse. However, most image detectors suffer from sharp performance drops in unseen domains. The key of fake image detection is to develop a generalized representation to describe the artifacts produced by generation models. In this work, we introduce a novel detection framework, named Learning on Gradients (LGrad), designed for identifying GAN-generated images, with the aim of constructing a generalized detector with cross-model and cross-data. Specifically, a pretrained CNN model is employed as a transformation model to convert images into gradients. Subsequently, we leverage these gradients to present the generalized artifacts, which are fed into the classifier to ascertain the authenticity of the images. In our framework, we turn the data-dependent problem into a transformation-model-dependent problem. To the best of our knowledge, this is the first study to utilize gradients as the representation of artifacts in GAN-generated images. Extensive experiments demonstrate the effectiveness and robustness of gradients as generalized artifact representations. Our detector achieves a new state-of-the-art performance with a remarkable gain of 11.4%. The code is released at https://github.com/chuangchuangtan/LGrad.

Vision transformers use [CLS] tokens to predict image classes. Their explainability visualization has been studied using relevant information from [CLS] tokens or focusing on attention scores during self-attention. Such visualization, however, is challenging because of the dependence of the structure of a vision transformer on skip connections and attention operators, the instability of non-linearities in the learning process, and the limited reflection of self-attention scores on relevance. We argue that the output vectors for each input patch token in a vision transformer retain the image information of each patch location, which can facilitate the prediction of an image class. In this paper, we propose ICE (Adversarial Normalization: I Can visualize Everything), a novel method that enables a model to directly predict a class for each patch in an image; thus, advancing the effective visualization of the explainability of a vision transformer. Our method distinguishes background from foreground regions by predicting background classes for patches that do not determine image classes. We used the DeiT-S model, the most representative model employed in studies, on the explainability visualization of vision transformers. On the ImageNet-Segmentation dataset, ICE outperformed all explainability visualization methods for four cases depending on the model size. We also conducted quantitative and qualitative analyses on the tasks of weakly-supervised object localization and unsupervised object discovery. On the CUB-200-2011 and PASCALVOC07/12 datasets, ICE achieved comparable performance to the state-of-the-art methods. We incorporated ICE into the encoder of DeiT-S and improved efficiency by 44.01% on the ImageNet dataset over that achieved by the original DeiT-S model. We showed performance on the accuracy and efficiency comparable to EViT, the state-of-the-art pruning model, demonstrating the effectiveness of ICE. The code is available at https://github.com/Hanyang-HCC-Lab/ICE.

Enormous hand images with reliable annotations are collected through marker-based MoCap. Unfortunately, degradations caused by markers limit their application in hand appearance reconstruction. A clear appearance recovery insight is an image-to-image translation trained with unpaired data. However, most frameworks fail because there exists structure inconsistency from a degraded hand to a bare one. The core of our approach is to first disentangle the bare hand structure from those degraded images and then wrap the appearance to this structure with a dual adversarial discrimination (DAD) scheme. Both modules take full advantage of the semi-supervised learning paradigm: The structure disentanglement benefits from the modeling ability of ViT, and the translator is enhanced by the dual discrimination on both translation processes and translation results. Comprehensive evaluations have been conducted to prove that our framework can robustly recover photo-realistic hand appearance from diverse marker-contained and even object-occluded datasets. It provides a novel avenue to acquire bare hand appearance data for other downstream learning problems.

Weakly-supervised Video Anomaly Detection is the task of detecting frame-level anomalies using video-level labeled training data. It is difficult to explore class representative features using minimal supervision of weak labels with a single backbone branch. Furthermore, in real-world scenarios, the boundary between normal and abnormal is ambiguous and varies depending on the situation. For example, even for the same motion of running person, the abnormality varies depending on whether the surroundings are a playground or a roadway. Therefore, our aim is to extract discriminative features by widening the relative gap between classes’ features from a single branch. In the proposed Class-Activate Feature Learning (CLAV), the features are extracted as per the weights that are implicitly activated depending on the class, and the gap is then enlarged through relative distance learning. Furthermore, as the relationship between context and motion is important in order to identify the anomalies in complex and diverse scenes, we propose a Context--Motion Interrelation Module (CoMo), which models the relationship between the appearance of the surroundings and motion, rather than utilizing only temporal dependencies or motion information. The proposed method shows SOTA performance on four benchmarks including large-scale real-world datasets, and we demonstrate the importance of relational information by analyzing the qualitative results and generalization ability.

Reconstruction-based anomaly detection models achieve their purpose by suppressing the generalization ability for anomaly. However, diverse normal patterns are consequently not well reconstructed as well. Although some efforts have been made to alleviate this problem by modeling sample diversity, they suffer from shortcut learning due to undesired transmission of abnormal information. In this paper, to better solve the tradeoff problem, we propose Diversity-Measurable Anomaly Detection (DMAD) framework to enhance reconstruction diversity while avoid the undesired generalization on anomalies. To this end, we design Pyramid Deformation Module (PDM), which models diverse normals and measures the severity of anomaly by estimating multi-scale deformation fields from reconstructed reference to original input. Integrated with an information compression module, PDM essentially decouples deformation from prototypical embedding and makes the final anomaly score more reliable. Experimental results on both surveillance videos and industrial images demonstrate the effectiveness of our method. In addition, DMAD works equally well in front of contaminated data and anomaly-like normal samples.

When facing changing environments in the real world, the lightweight model on client devices suffer from severe performance drop under distribution shifts. The main limitations of existing device model lie in: (1) unable to update due to the computation limit of the device, (2) limited generalization ability of the lightweight model. Meanwhile, recent large models have shown strong generalization capability on cloud while they can not be deployed on client devices due to the poor computation constraint. To enable the device model to deal with changing environments, we propose a new learning paradigm of Cloud-Device Collaborative Continual Adaptation. To encourage collaboration between cloud and device and improve the generalization of device model, we propose an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model in such paradigm. Specifically, we first design the Uncertainty Guided Sampling (UGS) to screen out challenging data continuously and transmit the most out-of-distribution samples from the device to the cloud. To further transfer the generalization capability of the large model on the cloud to the device model, we propose a Visual Prompt Learning Strategy with Uncertainty guided updating (VPLU) to specifically deal with the selected samples with more distribution shifts. Then, we transmit the visual prompts to the device and concatenate them with the incoming data to pull the device testing distribution closer to the cloud training distribution. We conduct extensive experiments on two object detection datasets with continually changing environments. Our proposed U-VPA teacher-student framework outperforms previous state-of-the-art test time adaptation and device-cloud collaboration methods. The code and datasets will be released.

Personalized federated learning (pFL) collaboratively trains personalized models, which provides a customized model solution for individual clients in the presence of heterogeneous distributed local data. Although many recent studies have applied various algorithms to enhance personalization in pFL, they mainly focus on improving the performance from averaging or top perspective. However, part of the clients may fall into poor performance and are not clearly discussed. Therefore, how to prevent these poor clients should be considered critically. Intuitively, these poor clients may come from biased universal information shared with others. To address this issue, we propose a novel pFL strategy, called Personalize Locally, Generalize Universally (PLGU). PLGU generalizes the fine-grained universal information and moderates its biased performance by designing a Layer-Wised Sharpness Aware Minimization (LWSAM) algorithm while keeping the personalization local. Specifically, we embed our proposed PLGU strategy into two pFL schemes concluded in this paper: with/without a global model, and present the training procedures in detail. Through in-depth study, we show that the proposed PLGU strategy achieves competitive generalization bounds on both considered pFL schemes. Our extensive experimental results show that all the proposed PLGU based-algorithms achieve state-of-the-art performance.

The Federated Learning (FL) paradigm is known to face challenges under heterogeneous client data. Local training on non-iid distributed data results in deflected local optimum, which causes the client models drift further away from each other and degrades the aggregated global model’s performance. A natural solution is to gather all client data onto the server, such that the server has a global view of the entire data distribution. Unfortunately, this reduces to regular training, which compromises clients’ privacy and conflicts with the purpose of FL. In this paper, we put forth an idea to collect and leverage global knowledge on the server without hindering data privacy. We unearth such knowledge from the dynamics of the global model’s trajectory. Specifically, we first reserve a short trajectory of global model snapshots on the server. Then, we synthesize a small pseudo dataset such that the model trained on it mimics the dynamics of the reserved global model trajectory. Afterward, the synthesized data is used to help aggregate the deflected clients into the global model. We name our method DynaFed, which enjoys the following advantages: 1) we do not rely on any external on-server dataset, which requires no additional cost for data collection; 2) the pseudo data can be synthesized in early communication rounds, which enables DynaFed to take effect early for boosting the convergence and stabilizing training; 3) the pseudo data only needs to be synthesized once and can be directly utilized on the server to help aggregation in subsequent rounds. Experiments across extensive benchmarks are conducted to showcase the effectiveness of DynaFed. We also provide insights and understanding of the underlying mechanism of our method.

Federated learning enables the privacy-preserving training of neural network models using real-world data across distributed clients. FedAvg has become the preferred optimizer for federated learning because of its simplicity and effectiveness. FedAvg uses naïve aggregation to update the server model, interpolating client models based on the number of instances used in their training. However, naïve aggregation suffers from client-drift when the data is heterogenous~(non-IID), leading to unstable and slow convergence. In this work, we propose a novel aggregation approach, elastic aggregation, to overcome these issues. Elastic aggregation interpolates client models adaptively according to parameter sensitivity, which is measured by computing how much the overall prediction function output changes when each parameter is changed. This measurement is performed in an unsupervised and online manner. Elastic aggregation reduces the magnitudes of updates to the more sensitive parameters so as to prevent the server model from drifting to any one client distribution, and conversely boosts updates to the less sensitive parameters to better explore different client distributions. Empirical results on real and synthetic data as well as analytical results show that elastic aggregation leads to efficient training in both convex and non-convex settings, while being fully agnostic to client heterogeneity and robust to large numbers of clients, partial participation, and imbalanced data. Finally, elastic aggregation works well with other federated optimizers and achieves significant improvements across the board.

Federated Learning with Model Distillation (FedMD) is a nascent collaborative learning paradigm, where only output logits of public datasets are transmitted as distilled knowledge, instead of passing on private model parameters that are susceptible to gradient inversion attacks, a known privacy risk in federated learning. In this paper, we found that even though sharing output logits of public datasets is safer than directly sharing gradients, there still exists a substantial risk of data exposure caused by carefully designed malicious attacks. Our study shows that a malicious server can inject a PLI (Paired-Logits Inversion) attack against FedMD and its variants by training an inversion neural network that exploits the confidence gap between the server and client models. Experiments on multiple facial recognition datasets validate that under FedMD-like schemes, by using paired server-client logits of public datasets only, the malicious server is able to reconstruct private images on all tested benchmarks with a high success rate.

For point cloud reconstruction-related tasks, the reconstruction losses to evaluate the shape differences between reconstructed results and the ground truths are typically used to train the task networks. Most existing works measure the training loss with point-to-point distance, which may introduce extra defects as predefined matching rules may deviate from the real shape differences. Although some learning-based works have been proposed to overcome the weaknesses of manually-defined rules, they still measure the shape differences in 3D Euclidean space, which may limit their ability to capture defects in reconstructed shapes. In this work, we propose a learning-based Contrastive Adversarial Loss (CALoss) to measure the point cloud reconstruction loss dynamically in a non-linear representation space by combining the contrastive constraint with the adversarial strategy. Specifically, we use the contrastive constraint to help CALoss learn a representation space with shape similarity, while we introduce the adversarial strategy to help CALoss mine differences between reconstructed results and ground truths. According to experiments on reconstruction-related tasks, CALoss can help task networks improve reconstruction performances and learn more representative representations.

Due to the increasing computational demand of Deep Neural Networks (DNNs), companies and organizations have begun to outsource the training process. However, the externally trained DNNs can potentially be backdoor attacked. It is crucial to defend against such attacks, i.e, to postprocess a suspicious model so that its backdoor behavior is mitigated while its normal prediction power on clean inputs remain uncompromised. To remove the abnormal backdoor behavior, existing methods mostly rely on additional labeled clean samples. However, such requirement may be unrealistic as the training data are often unavailable to end users. In this paper, we investigate the possibility of circumventing such barrier. We propose a novel defense method that does not require training labels. Through a carefully designed layer-wise weight re-initialization and knowledge distillation, our method can effectively cleanse backdoor behaviors of a suspicious network {with negligible compromise in} its normal behavior. In experiments, we show that our method, trained without labels, is on-par with state-of-the-art defense methods trained using labels. We also observe promising defense results even on out-of-distribution data. This makes our method very practical. Code is available at: https://github.com/luluppang/BCU.

Deep neural networks (DNNs) are recently shown to be vulnerable to backdoor attacks, where attackers embed hidden backdoors in the DNN model by injecting a few poisoned examples into the training dataset. While extensive efforts have been made to detect and remove backdoors from backdoored DNNs, it is still not clear whether a backdoor-free clean model can be directly obtained from poisoned datasets. In this paper, we first construct a causal graph to model the generation process of poisoned data and find that the backdoor attack acts as the confounder, which brings spurious associations between the input images and target labels, making the model predictions less reliable. Inspired by the causal understanding, we propose the Causality-inspired Backdoor Defense (CBD), to learn deconfounded representations by employing the front-door adjustment. Specifically, a backdoored model is intentionally trained to capture the confounding effects. The other clean model dedicates to capturing the desired causal effects by minimizing the mutual information with the confounding representations from the backdoored model and employing a sample-wise re-weighting scheme. Extensive experiments on multiple benchmark datasets against 6 state-of-the-art attacks verify that our proposed defense method is effective in reducing backdoor threats while maintaining high accuracy in predicting benign samples. Further analysis shows that CBD can also resist potential adaptive attacks.

Recently, self-supervised learning (SSL) was shown to be vulnerable to patch-based data poisoning backdoor attacks. It was shown that an adversary can poison a small part of the unlabeled data so that when a victim trains an SSL model on it, the final model will have a backdoor that the adversary can exploit. This work aims to defend self-supervised learning against such attacks. We use a three-step defense pipeline, where we first train a model on the poisoned data. In the second step, our proposed defense algorithm (PatchSearch) uses the trained model to search the training data for poisoned samples and removes them from the training set. In the third step, a final model is trained on the cleaned-up training set. Our results show that PatchSearch is an effective defense. As an example, it improves a model’s accuracy on images containing the trigger from 38.2% to 63.7% which is very close to the clean model’s accuracy, 64.6%. Moreover, we show that PatchSearch outperforms baselines and state-of-the-art defense approaches including those using additional clean, trusted data. Our code is available at https://github.com/UCDvision/PatchSearch

Recent deep-learning-based compression methods have achieved superior performance compared with traditional approaches. However, deep learning models have proven to be vulnerable to backdoor attacks, where some specific trigger patterns added to the input can lead to malicious behavior of the models. In this paper, we present a novel backdoor attack with multiple triggers against learned image compression models. Motivated by the widely used discrete cosine transform (DCT) in existing compression systems and standards, we propose a frequency-based trigger injection model that adds triggers in the DCT domain. In particular, we design several attack objectives for various attacking scenarios, including: 1) attacking compression quality in terms of bit-rate and reconstruction quality; 2) attacking task-driven measures, such as down-stream face recognition and semantic segmentation. Moreover, a novel simple dynamic loss is designed to balance the influence of different loss terms adaptively, which helps achieve more efficient training. Extensive experiments show that with our trained trigger injection models and simple modification of encoder parameters (of the compression model), the proposed attack can successfully inject several backdoors with corresponding triggers in a single image compression model.

Recently, deep neural networks have shown great success on 3D point cloud classification tasks, which simultaneously raises the concern of adversarial attacks that cause severe damage to real-world applications. Moreover, defending against adversarial examples in point cloud data is extremely difficult due to the emergence of various attack strategies. In this work, with the insight of the fact that the adversarial examples in this task still preserve the same semantic and structural information as the original input, we design a novel defense framework for improving the robustness of existing classification models, which consists of two main modules: the attention-based pooling and the dynamic contrastive learning. In addition, we also develop an algorithm to theoretically certify the robustness of the proposed framework. Extensive empirical results on two datasets and three classification models show the robustness of our approach against various attacks, e.g., the averaged attack success rate of PointNet decreases from 70.2% to 2.7% on the ModelNet40 dataset under 9 common attacks.

In recent years, as various realistic face forgery techniques known as DeepFake improves by leaps and bounds, more and more DeepFake detection techniques have been proposed. These methods typically rely on detecting statistical differences between natural (i.e., real) and DeepFake-generated images in both spatial and frequency domains. In this work, we propose to explicitly minimize the statistical differences to evade state-of-the-art DeepFake detectors. To this end, we propose a statistical consistency attack (StatAttack) against DeepFake detectors, which contains two main parts. First, we select several statistical-sensitive natural degradations (i.e., exposure, blur, and noise) and add them to the fake images in an adversarial way. Second, we find that the statistical differences between natural and DeepFake images are positively associated with the distribution shifting between the two kinds of images, and we propose to use a distribution-aware loss to guide the optimization of different degradations. As a result, the feature distributions of generated adversarial examples is close to the natural images. Furthermore, we extend the StatAttack to a more powerful version, MStatAttack, where we extend the single-layer degradation to multi-layer degradations sequentially and use the loss to tune the combination weights jointly. Comprehensive experimental results on four spatial-based detectors and two frequency-based detectors with four datasets demonstrate the effectiveness of our proposed attack method in both white-box and black-box settings.

In this paper, we propose a novel transfer-based targeted attack method that optimizes the adversarial perturbations without any extra training efforts for auxiliary networks on training data. Our new attack method is proposed based on the observation that highly universal adversarial perturbations tend to be more transferable for targeted attacks. Therefore, we propose to make the perturbation to be agnostic to different local regions within one image, which we called as self-universality. Instead of optimizing the perturbations on different images, optimizing on different regions to achieve self-universality can get rid of using extra data. Specifically, we introduce a feature similarity loss that encourages the learned perturbations to be universal by maximizing the feature similarity between adversarial perturbed global images and randomly cropped local regions. With the feature similarity loss, our method makes the features from adversarial perturbations to be more dominant than that of benign images, hence improving targeted transferability. We name the proposed attack method as Self-Universality (SU) attack. Extensive experiments demonstrate that SU can achieve high success rates for transfer-based targeted attacks. On ImageNet-compatible dataset, SU yields an improvement of 12% compared with existing state-of-the-art methods. Code is available at https://github.com/zhipeng-wei/Self-Universality.

Deep neural networks (DNNs) are susceptible to adversarial images, raising concerns about their reliability in safety-critical tasks. Sparse adversarial attacks, which limit the number of modified pixels, have shown to be highly effective in causing DNNs to misclassify. However, existing methods often struggle to simultaneously minimize the number of modified pixels and the size of the modifications, often requiring a large number of queries and assuming unrestricted access to the targeted DNN. In contrast, other methods that limit the number of modified pixels often permit unbounded modifications, making them easily detectable. To address these limitations, we propose a novel multi-objective sparse attack algorithm that efficiently minimizes the number of modified pixels and their size during the attack process. Our algorithm draws inspiration from evolutionary computation and incorporates a mechanism for prioritizing objectives that aligns with an attacker’s goals. Our approach outperforms existing sparse attacks on CIFAR-10 and ImageNet trained DNN classifiers while requiring only a small query budget, attaining competitive attack success rates while perturbing fewer pixels. Overall, our proposed attack algorithm provides a solution to the limitations of current sparse attack methods by jointly minimizing the number of modified pixels and their size. Our results demonstrate the effectiveness of our approach in restricted scenarios, highlighting its potential to enhance DNN security.

The origin of adversarial examples is still inexplicable in research fields, and it arouses arguments from various viewpoints, albeit comprehensive investigations. In this paper, we propose a way of delving into the unexpected vulnerability in adversarially trained networks from a causal perspective, namely adversarial instrumental variable (IV) regression. By deploying it, we estimate the causal relation of adversarial prediction under an unbiased environment dissociated from unknown confounders. Our approach aims to demystify inherent causal features on adversarial examples by leveraging a zero-sum optimization game between a casual feature estimator (i.e., hypothesis model) and worst-case counterfactuals (i.e., test function) disturbing to find causal features. Through extensive analyses, we demonstrate that the estimated causal features are highly related to the correct prediction for adversarial robustness, and the counterfactuals exhibit extreme features significantly deviating from the correct prediction. In addition, we present how to effectively inoculate CAusal FEatures (CAFE) into defense networks for improving adversarial robustness.

Adversarial training is widely used to make classifiers robust to a specific threat or adversary, such as lp-norm bounded perturbations of a given p-norm. However, existing methods for training classifiers robust to multiple threats require knowledge of all attacks during training and remain vulnerable to unseen distribution shifts. In this work, we describe how to obtain adversarially-robust model soups (i.e., linear combinations of parameters) that smoothly trade-off robustness to different lp-norm bounded adversaries. We demonstrate that such soups allow us to control the type and level of robustness, and can achieve robustness to all threats without jointly training on all of them. In some cases, the resulting model soups are more robust to a given l_p-norm adversary than the constituent model specialized against that same adversary. Finally, we show that adversarially-robust model soups can be a viable tool to adapt to distribution shifts from a few examples.

Physical world adversarial attack is a highly practical and threatening attack, which fools real world deep learning systems by generating conspicuous and maliciously crafted real world artifacts. In physical world attacks, evaluating naturalness is highly emphasized since human can easily detect and remove unnatural attacks. However, current studies evaluate naturalness in a case-by-case fashion, which suffers from errors, bias and inconsistencies. In this paper, we take the first step to benchmark and assess visual naturalness of physical world attacks, taking autonomous driving scenario as the first attempt. First, to benchmark attack naturalness, we contribute the first Physical Attack Naturalness (PAN) dataset with human rating and gaze. PAN verifies several insights for the first time: naturalness is (disparately) affected by contextual features (i.e., environmental and semantic variations) and correlates with behavioral feature (i.e., gaze signal). Second, to automatically assess attack naturalness that aligns with human ratings, we further introduce Dual Prior Alignment (DPA) network, which aims to embed human knowledge into model reasoning process. Specifically, DPA imitates human reasoning in naturalness assessment by rating prior alignment and mimics human gaze behavior by attentive prior alignment. We hope our work fosters researches to improve and automatically assess naturalness of physical world attacks. Our code and exemplar data can be found at https://github.com/zhangsn-19/PAN.

Owing to the extensive application of infrared object detectors in the safety-critical tasks, it is necessary to evaluate their robustness against adversarial examples in the real world. However, current few physical infrared attacks are complicated to implement in practical application because of their complex transformation from digital world to physical world. To address this issue, in this paper, we propose a physically feasible infrared attack method called “adversarial infrared patches”. Considering the imaging mechanism of infrared cameras by capturing objects’ thermal radiation, adversarial infrared patches conduct attacks by attaching a patch of thermal insulation materials on the target object to manipulate its thermal distribution. To enhance adversarial attacks, we present a novel aggregation regularization to guide the simultaneous learning for the patch’ shape and location on the target object. Thus, a simple gradient-based optimization can be adapted to solve for them. We verify adversarial infrared patches in different object detection tasks with various object detectors. Experimental results show that our method achieves more than 90% Attack Success Rate (ASR) versus the pedestrian detector and vehicle detector in the physical environment, where the objects are captured in different angles, distances, postures, and scenes. More importantly, adversarial infrared patch is easy to implement, and it only needs 0.5 hour to be constructed in the physical world, which verifies its effectiveness and efficiency.

Advancements in the generation quality of various Generative Models (GMs) has made it necessary to not only perform binary manipulation detection but also localize the modified pixels in an image. However, prior works termed as passive for manipulation localization exhibit poor generalization performance over unseen GMs and attribute modifications. To combat this issue, we propose a proactive scheme for manipulation localization, termed MaLP. We encrypt the real images by adding a learned template. If the image is manipulated by any GM, this added protection from the template not only aids binary detection but also helps in identifying the pixels modified by the GM. The template is learned by leveraging local and global-level features estimated by a two-branch architecture. We show that MaLP performs better than prior passive works. We also show the generalizability of MaLP by testing on 22 different GMs, providing a benchmark for future research on manipulation localization. Finally, we show that MaLP can be used as a discriminator for improving the generation quality of GMs. Our models/codes are available at www.github.com/vishal3477/pro_loc.