ICCV 2023 Papers

In instance image retrieval, considering local spatial information within an image has proven effective to boost retrieval performance, as demonstrated by local visual descriptor based geometric verification. Nevertheless, it will be highly valuable to make ordinary global image representations spatial-context-aware because global representation based image retrieval is appealing thanks to its algorithmic simplicity, low memory cost, and being friendly to sophisticated data structures. To this end, we propose a novel feature learning framework for instance image retrieval, which embeds local spatial context information into the learned global feature representations. Specifically, in parallel to the visual feature branch in a CNN backbone, we design a spatial context branch that consists of two modules called online token learning and distance encoding. For each local descriptor learned in CNN, the former module is used to indicate the types of its surrounding descriptors, while their spatial distribution information is captured by the latter module. After that, the visual feature branch and the spatial context branch are fused to produce a single global feature representation per image. As experimentally demonstrated, with the spatial-context-aware characteristic, we can well improve the performance of global representation based image retrieval while maintaining all of its appealing properties. Our code is available at https://github.com/Zy-Zhang/SpCa

Shape primitives decomposition has been an important and long-standing task in 3D shape analysis. Prior arts heavily rely on 3D point clouds or voxel data for shape primitives extraction, which are less practical in real-world scenarios. This paper proposes to learn shape primitives from multi-view images by introducing implicit surface rendering. It is challenging since implicit shapes have a high degree of freedom, which violates the simplicity property of shape primitives. In this work, a novel regularization term named Implicit Convexity Regularization (ICR) imposed on implicit primitive learning is proposed to tackle this problem. We start with the convexity definition of general 3D shapes, and then derive the equivalent expression for implicit shapes represented by signed distance functions (SDFs). Further, instead of directly constraining the output SDF values which cause unstable optimization, we alternatively impose constraint on second order directional derivatives on line segments inside the shapes, which proves to be a tighter condition for 3D convexity. Implicit primitives constrained by the proposed ICR are combined into a whole object via softmax-weighted-sum operation over all primitive SDFs. Experiments on synthetic and real-world datasets show that our method is able to decompose objects into simple and reasonable shape primitives without the need of segmentation labels or 3D data. Code and data is publicly available in https://github.com/seanywang0408/ICR.

In this paper, we address the problem of generalized category discovery (GCD), i.e., given a set of images where part of them are labelled and the rest are not, the task is to automatically cluster the images in the unlabelled data, leveraging the information from the labelled data, while the unlabelled data contain images from the labelled classes and also new ones. GCD is similar to semi-supervised learning (SSL) but is more realistic and challenging, as SSL assumes all the unlabelled images are from the same classes as the labelled ones. We also do not assume the class number in the unlabelled data is known a-priori, making the GCD problem even harder. To tackle the problem of GCD without knowing the class number, we propose an EM-like framework that alternates between representation learning and class number estimation. We propose a semi-supervised variant of the Gaussian Mixture Model (GMM) with a stochastic splitting and merging mechanism to dynamically determine the prototypes by examining the cluster compactness and separability. With these prototypes, we leverage prototypical contrastive learning for representation learning on the partially labelled data subject to the constraints imposed by the labelled data. Our framework alternates between these two steps until convergence. The cluster assignment for an unlabelled instance can then be retrieved by identifying its nearest prototype. We comprehensively evaluate our framework on both generic image classification datasets and challenging fine-grained object recognition datasets, achieving state-of-the-art performance.

Although previous RGB-D-based gesture recognition methods have shown promising performance, researchers often overlook the interference of task-irrelevant cues like illumination and background. These unnecessary factors are learned together with the predictive ones by the network and hinder accurate recognition. In this paper, we propose a convenient and analytical framework to learn a robust feature representation that is impervious to gesture-irrelevant factors. Based on the Information Bottleneck theory, two rules of Sufficiency and Compactness are derived to develop a new information-theoretic loss function, which cultivates a more sufficient and compact representation from the feature encoding and mitigates the impact of gesture-irrelevant information. To highlight the predictive information, we further integrate a memory network. Using our proposed content-based and contextual memory addressing scheme, we weaken the nuisances while preserving the task-relevant information, providing guidance for refining the feature representation. Experiments conducted on three public datasets demonstrate that our approach leads to a better feature representation and achieves better performance than state-of-the-art methods.

Rain can significantly degrade image quality and visibility, making deraining a critical area of research in computer vision. Despite recent progress in learning-based deraining methods, there is a lack of focus on nighttime deraining due to the unique challenges posed by non-uniform local illuminations from artificial light sources. Rain streaks in these scenes have diverse appearances that are tightly related to their relative positions to light sources, making it difficult for existing deraining methods to effectively handle them. In this paper, we highlight the importance of rain streak location information in nighttime deraining. Specifically, we propose a Rain Location Prior (RLP) that is learned implicitly from rainy images using a recurrent residual model. This learned prior contains location information of rain streaks and, when injected into deraining models, can significantly improve their performance. To further improve the effectiveness of the learned prior, we also propose a Rain Prior Injection Module (RPIM) to modulate the prior before injection, increasing the importance of features within rain streak areas. Experimental results demonstrate that our approach outperforms existing state-of-the-art methods by about 1dB and effectively improves the performance of deraining models. We also evaluate our method on real night rainy images to show the capability to handle real scenes with fully synthetic data for training. Our method represents a significant step forward in the area of nighttime deraining and highlights the importance of location information in this challenging problem. The code is publicly available at https://github.com/zkawfanx/RLP.

Domain adaptive semantic segmentation aims to adapt a model trained on labeled source domain to the unlabeled target domain. Self-training shows competitive potential in this field. Existing methods along this stream mainly focus on selecting reliable predictions on target data as pseudo-labels for category learning, while ignoring the useful relations between pixels for relation learning. In this paper, we propose a pseudo-relation learning framework, Relation Teacher (RTea), which can exploitable pixel relations to efficiently use unreliable pixels and learn generalized representations. In this framework, we build reasonable pseudo-relations on local grids and fuse them with low-level relations in the image space, which are motivated by the reliable local relations prior and available low-level relations prior. Then, we design a pseudo-relation learning strategy and optimize the class probability to meet the relation consistency by finding the optimal sub-graph division. In this way, the model's certainty and consistency of prediction are enhanced on the target domain, and the cross-domain inadaptation is further eliminated. Extensive experiments on three datasets demonstrate the effectiveness of the proposed method.

Point cloud completion is to restore complete 3D scenes and objects from incomplete observations or limited sensor data. Existing fully-supervised methods rely on paired datasets of incomplete and complete point clouds, which are labor-intensive to obtain. Unpaired methods have been proposed, but still require a set of complete point clouds as a reference. As a remedy, in this paper, we propose a novel point cloud completion framework without using any complete point cloud at all. Our main idea is to generate multiple incomplete point clouds of various poses and integrate them into a complete point cloud. We train our framework based on cycle consistency, to generate an incomplete point cloud such that 1) shares the same object as the input incomplete point cloud and 2) corresponds to an arbitrarily given pose. In addition, we devise a novel projection method conditioned by pose to gather visible features, from a volumetric feature extracted by an encoder. Extensive experiments demonstrate that the proposed method achieves comparable or better results than existing unpaired methods. Further, we show that our method also can be applied to real incomplete point clouds.

We study the problem of estimating optical flow from event cameras. One important issue is how to build a high-quality event-flow dataset with accurate event values and flow labels. Previous datasets are created by either capturing real scenes by event cameras or synthesizing from images with pasted foreground objects. The former case can produce real event values but with calculated flow labels, which are sparse and inaccurate. The later case can generate dense flow labels but the interpolated events are prone to errors. In this work, we propose to render a physically correct event-flow dataset using computer graphics models. In particular, we first create indoor and outdoor 3D scenes by Blender with rich scene content variations. Second, diverse camera motions are included for the virtual capturing, producing images and accurate flow labels. Third, we render high-framerate videos between images for accurate events. The rendered dataset can adjust the density of events, based on which we further introduce an adaptive density module (ADM). Experiments show that our proposed dataset can facilitate event-flow learning, whereas previous approaches when trained on our dataset can improve their performances constantly by a relatively large margin. In addition, event-flow pipelines when equipped with our ADM can further improve performances. Our code and dataset will be publicly available.

Exploiting spatial-angular correlation is crucial to light field (LF) image super-resolution (SR), but is highly challenging due to its non-local property caused by the disparities among LF images. Although many deep neural networks (DNNs) have been developed for LF image SR and achieved continuously improved performance, existing methods cannot well leverage the long-range spatial-angular correlation and thus suffer a significant performance drop when handling scenes with large disparity variations. In this paper, we propose a simple yet effective method to learn the non-local spatial-angular correlation for LF image SR. In our method, we adopt the epipolar plane image (EPI) representation to project the 4D spatial-angular correlation onto multiple 2D EPI planes, and then develop a Transformer network with repetitive self-attention operations to learn the spatial-angular correlation by modeling the dependencies between each pair of EPI pixels. Our method can fully incorporate the information from all angular views while achieving a global receptive field along the epipolar line. We conduct extensive experiments with insightful visualizations to validate the effectiveness of our method. Comparative results on five public datasets show that our method not only achieves state-of-the-art SR performance, but also performs robust to disparity variations.

Recent progress on multi-view 3D object reconstruction has featured neural implicit surfaces via learning high-fidelity radiance fields. However, most approaches hinge on the visual hull derived from cost-expensive silhouette masks to obtain object surfaces. In this paper, we propose a novel Object-aware Radiance Fields (ORF) to automatically learn an object-aware geometry reconstruction. The geometric correspondences between multi-view 2D object regions and 3D implicit/explicit object surfaces are additionally exploited to boost the learning of object surfaces. Technically, a critical transparency discriminator is designed to distinguish the object-intersected and object-bypassed rays based on the estimated 2D object regions, leading to 3D implicit object surfaces. Such implicit surfaces can be directly converted into explicit object surfaces (e.g., meshes) via marching cubes. Then, we build the geometric correspondence between 2D planes and 3D meshes by rasterization, and project the estimated object regions into 3D explicit object surfaces by aggregating the object information across multiple views. The aggregated object information in 3D explicit object surfaces is further reprojected back to 2D planes, aiming to update 2D object regions and enforce them to be multi-view consistent. Extensive experiments on DTU and BlendedMVS verify the capability of ORF to produce comparable surfaces against the state-of-the-art models that demand silhouette masks.

Unsupervised semantic segmentation is a long-standing challenge in computer vision with great significance. Spectral clustering is a theoretically grounded solution to it where the spectral embeddings for pixels are computed to construct distinct clusters. Despite recent progress in enhancing spectral clustering with powerful pre-trained models, current approaches still suffer from inefficiencies in spectral decomposition and inflexibility in applying them to the test data. This work addresses these issues by casting spectral clustering as a parametric approach that employs neural network-based eigenfunctions to produce spectral embeddings. The outputs of the neural eigenfunctions are further restricted to discrete vectors that indicate clustering assignments directly. As a result, an end-to-end NN-based paradigm of spectral clustering emerges. In practice, the neural eigenfunctions are lightweight and take the features from pre-trained models as inputs, improving training efficiency and unleashing the potential of pre-trained models for dense prediction. We conduct extensive empirical studies to validate the effectiveness of our approach and observe significant performance gains over competitive baselines on Pascal Context, Cityscapes, and ADE20K benchmarks. The code is available at https://github.com/thudzj/NeuralEigenfunctionSegmentor.

Being able to perceive the semantics and the spatial structure of the environment is essential for visual navigation of a household robot. However, most existing works only employ visual backbones pre-trained either with independent images for classification or with self-supervised learning methods to adapt to the indoor navigation domain, both neglecting the spatial relationships that are essential to the learning of navigation. Inspired by the behavior that human naturally build semantically and spatially meaningful cognitive maps in their brain during navigation, in this paper, we propose a novel navigational-specific visual representation learning method by contrasting the agent's egocentric views and semantic maps (Ego^2-Map). We apply the visual transformer as the backbone encoder and train the model with data collected from the large-scale Habitat-Matterport3D environments. Ego^2-Map learning transfers the compact and rich information from a map, such as objects, structure and transition, to the agent's egocentric representations for navigation. Experiments show that agents using our learned representations on object-goal navigation outperforms recent visual pre-training methods. Moreover, our representations lead to a significant improvement in vision-and-language navigation in continuous environments for both high-level and low-level action spaces, achieving new state-of-the-art results of 47% SR and 41% SPL on the test server.

Due to the limited resolution of 3D sensors and the inevitable mutual occlusion between objects, 3D scans of real scenes are commonly incomplete. Previous scene completion methods struggle to capture long-range spatial feature, resulting in unsatisfactory completion results. To alleviate the problem, we propose a novel Dual-Scale Transformer Network (DST-Net) that efficiently utilizes both long-range and short-range spatial context information to improve the quality of 3D scene completion. To reduce the heavy computation cost of extracting long-range features via transformers, DST-Net adopts a self-supervised two-stage completion strategy. In the first stage, we split the input scene into blocks, and perform completion on individual blocks. In the second stage, the blocks are merged together as a whole and then further refined to improve completeness. More importantly, we propose a contrastive attention training strategy to encourage the transformers to learn distinguishable features for better scene completion. Experiments on datasets of Matterport3D, ScanNet, and ICL-NUIM demonstrate that our method can generate better completion results, and our method outperforms the state-of-the-art methods quantitatively and qualitatively.

Recent work of discrete generative priors, in the form of codebooks, has shown exciting performance for image reconstruction and restoration, since the discrete prior space spanned by the codebooks increases the robustness against diverse image degradations. Nevertheless, these methods require separate training of codebooks for different image categories, which limits their use to specific image categories only (e.g. face, architecture, etc.), and fail to handle arbitrary natural images. In this paper, we propose AdaCode for learning image-adaptive codebooks for class-agnostic image restoration. Instead of learning a single codebook for all categories of images, we learn a set of basis codebooks. For a given input image, AdaCode learns a weight map with which we compute a weighted combination of these basis codebooks for adaptive image restoration. Intuitively, AdaCode is a more flexible and expressive discrete generative prior than previous work. Experimental results show that AdaCode achieves state-of-the-art performance on image reconstruction and restoration tasks, including image super-resolution and inpainting.

Harmonizing cut-and-paste images into perceptually realistic ones is challenging, as it requires a full understanding of the discrepancies between the background of the target image and the inserted object. Existing methods mainly adjust the appearances of the inserted object via pixel-level manipulations. They are not effective in correcting color discrepancy caused by different scene illuminations and the image formation processes. We note that image colors are essentially camera ISP projection of the scene radiance. If we can trace the image colors back to the radiance field, we may be able to model the scene illumination and harmonize the discrepancy better. In this paper, we propose a novel neural approach to harmonize the image colors in a camera-independent color space, in which color values are proportional to the scene radiance. To this end, we propose a novel image unprocessing module to estimate an intermediate high dynamic range version of the object to be inserted. We then propose a novel color harmonization module that harmonizes the colors of the inserted object by querying the estimated scene radiance and re-rendering the harmonized object in the output color space. Extensive experiments demonstrate that our method outperforms the state-of-the-art approaches.

Interactions between humans are diverse and context-dependent, but previous works have treated them as categorical, disregarding the heavy tail of possible interactions. We propose a new paradigm of learning human-human interactions as free text from a single still image, allowing for flexibility in modeling the unlimited space of situations and relationships between people. To overcome the absence of data labelled specifically for this task, we use knowledge distillation applied to synthetic caption data produced by a large language model without explicit supervision. We show that the pseudo-labels produced by this procedure can be used to train a captioning model to effectively understand human-human interactions in images, as measured by a variety of metrics that measure textual and semantic faithfulness and factual groundedness of our predictions. We further show that our approach outperforms SOTA image captioning and situation recognition models on this task. We will release our code and pseudo-labels along with Waldo and Wenda, a manually-curated test set for still image human-human interaction understanding.

Simulation has emerged as an indispensable tool for scaling and accelerating the development of self-driving systems. A critical aspect of this is simulating realistic and diverse human behavior and intent. In this work, we propose a holistic framework for learning physically plausible human dynamics from real driving scenarios, narrowing the gap between real and simulated human behavior in safety-critical applications. We show that state-of-the-art methods underperform in driving scenarios where video data is recorded from moving vehicles, and humans are frequently partially or fully occluded. Furthermore, existing methods often disregard the global scene where humans are situated, resulting in various motion artifacts like foot sliding, floating, or ground penetration. Therefore, the primary technical challenge of this work is to infer physically plausible human dynamics for the occluded body parts on uneven terrain, based on visible motions. To address this challenge, we propose an approach that incorporates physics with a reinforcement learning-based motion controller to learn human dynamics for driving scenarios. Our framework can simulate physically plausible human dynamics that accurately match observed human motions and infill motions for occluded body parts, while improving the physical plausibility of the entire motion sequence. We evaluate our method on the challenging driving scenarios in the Waymo Open Dataset. Experiments on the challenging Waymo Open Dataset show that our method outperforms state-of-the-art motion capture approaches significantly in recovering high-quality, physically plausible, and scene-aware human dynamics.

This paper studies the fundamental problem of multi-layer generator models in learning hierarchical representations. The multi-layer generator model that consists of multiple layers of latent variables organized in a top-down architecture tends to learn multiple levels of data abstraction. However, such multi-layer latent variables are typically parameterized to be Gaussian, which can be less informative in capturing complex abstractions, resulting in limited success in hierarchical representation learning. On the other hand, the energy-based (EBM) prior is known to be expressive in capturing the data regularities, but it often lacks the hierarchical structure to capture different levels of hierarchical representations. In this paper, we propose a joint latent space EBM prior model with multi-layer latent variables for effective hierarchical representation learning. We develop a variational joint learning scheme that seamlessly integrates an inference model for efficient inference. Our experiments demonstrate that the proposed joint EBM prior is effective and expressive in capturing hierarchical representations and modelling data distribution.

Image harmonization aims to solve the visual inconsistency problem in composited images by adaptively adjusting the foreground pixels with the background as references. Existing methods employ local color transformation or region matching between foreground and background, which neglects powerful proximity prior and independently distinguishes fore-/back-ground as a whole part for harmonization. As a result, they still show a limited performance across varied foreground objects and scenes. To address this issue, we propose a novel Global-aware Kernel Network (GKNet) to harmonize local regions with comprehensive consideration of long-distance background references. Specifically, GKNet includes two parts, i.e., harmony kernel prediction and harmony kernel modulation branches. The former includes a Long-distance Reference Extractor (LRE) to obtain long-distance context and Kernel Prediction Blocks (KPB) to predict multi-level harmony kernels by fusing global information with local features. To achieve this goal, a novel Selective Correlation Fusion (SCF) module is proposed to better select relevant long-distance background references for local harmonization. The latter employs the predicted kernels to harmonize foreground regions with both local and global awareness. Abundant experiments demonstrate the superiority of our method for image harmonization over state-of-the-art methods, e.g., achieving 39.53dB PSNR that surpasses the best counterpart by +0.78dB; decreasing fMSE by 11.5% and MSE by 6.7% compared with the SoTA method.