CVPR 2022 论文列表

Inverse problems in multi-dimensional imaging, e.g., completion, denoising, and compressive sensing, are challenging owing to the big volume of the data and the inherent ill-posedness. To tackle these issues, this work unsupervisedly learns a hierarchical low-rank tensor factorization (HLRTF) by solely using an observed multi-dimensional image. Specifically, we embed a deep neural network (DNN) into the tensor singular value decomposition framework and develop the HLRTF, which captures the underlying low-rank structures of multi-dimensional images with compact representation abilities. This DNN herein serves as a nonlinear transform from a vector to another to help obtain a better low-rank representation. Our HLRTF infers the parameters of the DNN and the underlying low-rank structure of the original data from its observation via the gradient descent using a non-reference loss function in an unsupervised manner. To address the vanishing gradient in extreme scenarios, e.g., structural missing pixels, we introduce a parametric total variation regularization to constrain the DNN parameters and the tensor factor parameters with theoretical analysis. We apply our HLRTF for typical inverse problems in multi-dimensional imaging including completion, denoising, and snapshot spectral imaging, which demonstrates its generality and wide applicability. Extensive results illustrate the superiority of our method as compared with state-of-the-art methods.

Scene graph generation (SGG) aims to detect objects and predict their pairwise relationships within an image. Current SGG methods typically utilize graph neural networks (GNNs) to acquire context information between objects/relationships. Despite their effectiveness, however, current SGG methods only assume scene graph homophily while ignoring heterophily. Accordingly, in this paper, we propose a novel Heterophily Learning Network (HL-Net) to comprehensively explore the homophily and heterophily between objects/relationships in scene graphs. More specifically, HL-Net comprises the following 1) an adaptive reweighting transformer module, which adaptively integrates the information from different layers to exploit both the heterophily and homophily in objects; 2) a relationship feature propagation module that efficiently explores the connections between relationships by considering heterophily in order to refine the relationship representation; 3) a heterophily-aware message-passing scheme to further distinguish the heterophily and homophily between objects/relationships, thereby facilitating improved message passing in graphs. We conducted extensive experiments on two public datasets: Visual Genome (VG) and Open Images (OI). The experimental results demonstrate the superiority of our proposed HL-Net over existing state-of-the-art approaches. In more detail, HL-Net outperforms the second-best competitors by 2.1% on the VG dataset for scene graph classification and 1.2% on the IO dataset for the final score.

To interpret deep networks, one main approach is to associate neurons with human-understandable concepts. However, existing methods often ignore the inherent connections of different concepts (e.g., dog and cat both belong to animals), and thus lose the chance to explain neurons responsible for higher-level concepts (e.g., animal). In this paper, we study hierarchical concepts inspired by the hierarchical cognition process of human beings. To this end, we propose HIerarchical Neuron concepT explainer (HINT) to effectively build bidirectional associations between neurons and hierarchical concepts in a low-cost and scalable manner. HINT enables us to systematically and quantitatively study whether and how the implicit hierarchical relationships of concepts are embedded into neurons. Specifically, HINT identifies collaborative neurons responsible for one concept and multimodal neurons pertinent to different concepts, at different semantic levels from concrete concepts (e.g., dog) to more abstract ones (e.g., animal). Finally, we verify the faithfulness of the associations using Weakly Supervised Object Localization, and demonstrate its applicability in various tasks, such as discovering saliency regions and explaining adversarial attacks. Code is available on https://github.com/AntonotnaWang/HINT.

This paper presents a novel attention-based neural network for structured reconstruction, which takes a 2D raster image as an input and reconstructs a planar graph depicting an underlying geometric structure. The approach detects corners and classifies edge candidates between corners in an end-to-end manner. Our contribution is a holistic edge classification architecture, which 1) initializes the feature of an edge candidate by a trigonometric positional encoding of its end-points; 2) fuses image feature to each edge candidate by deformable attention; 3) employs two weight-sharing Transformer decoders to learn holistic structural patterns over the graph edge candidates; and 4) is trained with a masked learning strategy. The corner detector is a variant of the edge classification architecture, adapted to operate on pixels as corner candidates. We conduct experiments on two structured reconstruction tasks: outdoor building architecture and indoor floorplan planar graph reconstruction. Extensive qualitative and quantitative evaluations demonstrate the superiority of our approach over the state of the art. Code and pre-trained models are available at https://heat-structured-reconstruction.github.io

We present High Dynamic Range Neural Radiance Fields (HDR-NeRF) to recover an HDR radiance field from a set of low dynamic range (LDR) views with different exposures. Using the HDR-NeRF, we are able to generate both novel HDR views and novel LDR views under different exposures. The key to our method is to model the physical imaging process, which dictates that the radiance of a scene point transforms to a pixel value in the LDR image with two implicit functions: a radiance field and a tone mapper. The radiance field encodes the scene radiance (values vary from 0 to +infty), which outputs the density and radiance of a ray by giving corresponding ray origin and ray direction. The tone mapper models the mapping process that a ray hitting on the camera sensor becomes a pixel value. The color of the ray is predicted by feeding the radiance and the corresponding exposure time into the tone mapper. We use the classic volume rendering technique to project the output radiance, colors, and densities into HDR and LDR images, while only the input LDR images are used as the supervision. We collect a new forward-facing HDR dataset to evaluate the proposed method. Experimental results on synthetic and real-world scenes validate that our method can not only accurately control the exposures of synthesized views but also render views with a high dynamic range.

The rapid development of deep learning provides a better solution for the end-to-end reconstruction of hyperspectral image (HSI). However, existing learning-based methods have two major defects. Firstly, networks with self-attention usually sacrifice internal resolution to balance model performance against complexity, losing fine-grained high-resolution (HR) features. Secondly, even if the optimization focusing on spatial-spectral domain learning (SDL) converges to the ideal solution, there is still a significant visual difference between the reconstructed HSI and the truth. So we propose a high-resolution dual-domain learning network (HDNet) for HSI reconstruction. On the one hand, the proposed HR spatial-spectral attention module with its efficient feature fusion provides continuous and fine pixel-level features. On the other hand, frequency domain learning (FDL) is introduced for HSI reconstruction to narrow the frequency domain discrepancy. Dynamic FDL supervision forces the model to reconstruct fine-grained frequencies and compensate for excessive smoothing and distortion caused by pixel-level losses. The HR pixel-level attention and frequency-level refinement in our HDNet mutually promote HSI perceptual quality. Extensive quantitative and qualitative experiments show that our method achieves SOTA performance on simulated and real HSI datasets. https://github.com/Huxiaowan/HDNet

Hierarchical semantic structures naturally exist in an image dataset, in which several semantically relevant image clusters can be further integrated into a larger cluster with coarser-grained semantics. Capturing such structures with image representations can greatly benefit the semantic understanding on various downstream tasks. Existing contrastive representation learning methods lack such an important model capability. In addition, the negative pairs used in these methods are not guaranteed to be semantically distinct, which could further hamper the structural correctness of learned image representations. To tackle these limitations, we propose a novel contrastive learning framework called Hierarchical Contrastive Selective Coding (HCSC). In this framework, a set of hierarchical prototypes are constructed and also dynamically updated to represent the hierarchical semantic structures underlying the data in the latent space. To make image representations better fit such semantic structures, we employ and further improve conventional instance-wise and prototypical contrastive learning via an elaborate pair selection scheme. This scheme seeks to select more diverse positive pairs with similar semantics and more precise negative pairs with truly distinct semantics. On extensive downstream tasks, we verify the superior performance of HCSC over state-of-the-art contrastive methods, and the effectiveness of major model components is proved by plentiful analytical studies. We are continually building a comprehensive model zoo (see supplementary material). Our source code and model weights are available at https://github.com/gyfastas/HCSC.

We propose a novel hierarchical approach for multiple rotation averaging, dubbed HARA. Our method incrementally initializes the rotation graph based on a hierarchy of triplet support. The key idea is to build a spanning tree by prioritizing the edges with many strong triplet supports and gradually adding those with weaker and fewer supports. This reduces the risk of adding outliers in the spanning tree. As a result, we obtain a robust initial solution that enables us to filter outliers prior to nonlinear optimization. With minimal modification, our approach can also integrate the knowledge of the number of valid 2D-2D correspondences. We perform extensive evaluations on both synthetic and real datasets, demonstrating state-of-the-art results.

Despite the impressive results achieved by deep learning based 3D reconstruction, the techniques of directly learning to model 4D human captures with detailed geometry have been less studied. This work presents a novel framework that can effectively learn a compact and compositional representation for dynamic human by exploiting the human body prior from the widely used SMPL parametric model. Particularly, our representation, named H4D, represents a dynamic 3D human over a temporal span with the SMPL parameters of shape and initial pose, and latent codes encoding motion and auxiliary information. A simple yet effective linear motion model is proposed to provide a rough and regularized motion estimation, followed by per-frame compensation for pose and geometry details with the residual encoded in the auxiliary code. Technically, we introduce novel GRU-based architectures to facilitate learning and improve the representation capability. Extensive experiments demonstrate our method is not only efficacy in recovering dynamic human with accurate motion and detailed geometry, but also amenable to various 4D human related tasks, including motion retargeting, motion completion and future prediction.

Cross-domain weakly supervised object detection (CDWSOD) aims to adapt the detection model to a novel target domain with easily acquired image-level annotations. How to align the source and target domains is critical to the CDWSOD accuracy. Existing methods usually focus on partial detection components for domain alignment. In contrast, this paper considers that all the detection components are important and proposes a Holistic and Hierarchical Feature Alignment (H^2FA) R-CNN. H^2FA R-CNN enforces two image-level alignments for the backbone features, as well as two instance-level alignments for the RPN and detection head. This coarse-to-fine aligning hierarchy is in pace with the detection pipeline, i.e., processing the image-level feature and the instance-level features from bottom to top. Importantly, we devise a novel hybrid supervision method for learning two instance-level alignments. It enables the RPN and detection head to simultaneously receive weak/full supervision from the target/source domains. Combining all these feature alignments, H^2FA R-CNN effectively mitigates the gap between the source and target domains. Experimental results show that H^2FA R-CNN significantly improves cross-domain object detection accuracy and sets new state of the art on popular benchmarks. Code and pre-trained models are available at https://github.com/XuYunqiu/H2FA_R-CNN.

Depth completion has been widely studied to predict a dense depth image from its sparse measurement and a single color image. However, most state-of-the-art methods rely on static convolutional neural networks (CNNs) which are not flexible enough for capturing the dynamic nature of input contexts. In this paper, we propose GuideFormer, a fully transformer-based architecture for dense depth completion. We first process sparse depth and color guidance images with separate transformer branches to extract hierarchical and complementary token representations. Each branch consists of a stack of self-attention blocks and has key design features to make our model suitable for the task. We also devise an effective token fusion method based on guided-attention mechanism. It explicitly models information flow between the two branches and captures inter-modal dependencies that cannot be obtained from depth or color image alone. These properties allow GuideFormer to enjoy various visual dependencies and recover precise depth values while preserving fine details. We evaluate GuideFormer on the KITTI dataset containing real-world driving scenes and provide extensive ablation studies. Experimental results demonstrate that our approach significantly outperforms the state-of-the-art methods.

Grouping and recognition are important components of visual scene understanding, e.g., for object detection and semantic segmentation. With end-to-end deep learning systems, grouping of image regions usually happens implicitly via top-down supervision from pixel-level recognition labels. Instead, in this paper, we propose to bring back the grouping mechanism into deep networks, which allows semantic segments to emerge automatically with only text supervision. We propose a hierarchical Grouping Vision Transformer (GroupViT), which goes beyond the regular grid structure representation and learns to group image regions into progressively larger arbitrary-shaped segments. We train GroupViT jointly with a text encoder on a large-scale image-text dataset via contrastive losses. With only text supervision and without any pixel-level annotations, GroupViT learns to group together semantic regions and successfully transfers to the task of semantic segmentation in a zero-shot manner, i.e., without any further fine-tuning. It achieves a zero-shot accuracy of 52.3% mIoU on the PASCAL VOC 2012 and 22.4% mIoU on PASCAL Context datasets, and performs competitively to state-of-the-art transfer-learning methods requiring greater levels of supervision. We open-source our code at https://github.com/NVlabs/GroupViT

Demystifying the interactions among multiple agents from their past trajectories is fundamental to precise and interpretable trajectory prediction. However, previous works only consider pair-wise interactions with limited relational reasoning. To promote more comprehensive interaction modeling for relational reasoning, we propose GroupNet, a multiscale hypergraph neural network, which is novel in terms of both interaction capturing and representation learning. From the aspect of interaction capturing, we propose a trainable multiscale hypergraph to capture both pair-wise and group-wise interactions at multiple group sizes. From the aspect of interaction representation learning, we propose a three-element format that can be learnt end-to-end and explicitly reason some relational factors including the interaction strength and category. We apply GroupNet into both CVAE-based prediction system and previous state-of-the-art prediction systems for predicting socially plausible trajectories with relational reasoning. To validate the ability of relational reasoning, we experiment with synthetic physics simulations to reflect the ability to capture group behaviors, reason interaction strength and interaction category. To validate the effectiveness of prediction, we conduct extensive experiments on three real-world trajectory prediction datasets, including NBA, SDD and ETH-UCY; and we show that with GroupNet, the CVAE-based prediction system outperforms state-of-the-art methods. We also show that adding GroupNet will further improve the performance of previous state-of-the-art prediction systems.

We study the problem of weakly semi-supervised object detection with points (WSSOD-P), where the training data is combined by a small set of fully annotated images with bounding boxes and a large set of weakly-labeled images with only a single point annotated for each instance. The core of this task is to train a point-to-box regressor on well-labeled images that can be used to predict credible bounding boxes for each point annotation. We challenge the prior belief that existing CNN-based detectors are not compatible with this task. Based on the classic R-CNN architecture, we propose an effective point-to-box regressor: Group R-CNN. Group R-CNN first uses instance-level proposal grouping to generate a group of proposals for each point annotation and thus can obtain a high recall rate. To better distinguish different instances and improve precision, we propose instance-level proposal assignment to replace the vanilla assignment strategy adopted in original R-CNN methods. As naive instance-level assignment brings converging difficulty, we propose instance-aware representation learning which consists of instance-aware feature enhancement and instance-aware parameter generation to overcome this issue. Comprehensive experiments on the MS-COCO benchmark demonstrate the effectiveness of our method. Specifically, Group R-CNN significantly outperforms the prior method Point DETR by 3.9 mAP with 5% well-labeled images, which is the most challenging scenario. The source code can be found at https://github.com/jshilong/GroupRCNN.

Learning discriminative representation from the complex spatio-temporal dynamic space is essential for video recognition. On top of those stylized spatio-temporal computational units, further refining the learnt feature with axial contexts is demonstrated to be promising in achieving this goal. However, previous works generally focus on utilizing a single kind of contexts to calibrate entire feature channels and could hardly apply to deal with diverse video activities. The problem can be tackled by using pair-wise spatio-temporal attentions to recompute feature response with cross-axis contexts at the expense of heavy computations. In this paper, we propose an efficient feature refinement method that decomposes the feature channels into several groups and separately refines them with different axial contexts in parallel. We refer this lightweight feature calibration as group contextualization (GC). Specifically, we design a family of efficient element-wise calibrators, i.e., ECal-G/S/T/L, where their axial contexts are information dynamics aggregated from other axes either globally or locally, to contextualize feature channel groups. The GC module can be densely plugged into each residual layer of the off-the-shelf video networks. With little computational overhead, consistent improvement is observed when plugging in GC on different networks. By utilizing calibrators to embed feature with four different kinds of contexts in parallel, the learnt representation is expected to be more resilient to diverse types of activities. On videos with rich temporal variations, empirically GC can boost the performance of 2D-CNN (e.g., TSN and TSM) to a level comparable to the state-of-the-art video networks. Code is available at https://github.com/haoyanbin918/Group-Contextualization.

Training a good supernet in one-shot NAS methods is difficult since the search space is usually considerably huge (e.g., 13^ 21 ). In order to enhance the supernet's evaluation ability, one greedy strategy is to sample good paths, and let the supernet lean towards the good ones and ease its evaluation burden as a result. However, in practice the search can be still quite inefficient since the identification of good paths is not accurate enough and sampled paths still scatter around the whole search space. In this paper, we leverage an explicit path filter to capture the characteristics of paths and directly filter those weak ones, so that the search can be thus implemented on the shrunk space more greedily and efficiently. Concretely, based on the fact that good paths are way much less than the weak ones in the space, we argue that the label of "weak paths" will be more confident and reliable than that of "good paths" in multi-path sampling. In this way, we thus cast the training of path filter in the positive and unlabeled (PU) learning paradigm, and also encourage a path embedding as better path/operation representation to enhance the identification capacity of the learned filter. By dint of this embedding, we can further shrink the search space by aggregating similar operations with similar embeddings, and the search can be more efficient and accurate. Extensive experiments validate the effectiveness of the proposed method GreedyNASv2. For example, our obtained GreedyNASv2-L achieves 81.1% Top-1 accuracy on ImageNet dataset, significantly outperforming the ResNet-50 strong baselines.

Measurements from the Event Horizon Telescope enabled the visualization of light emission around a black hole for the first time. So far, these measurements have been used to recover a 2D image under the assumption that the emission field is static over the period of acquisition. In this work, we propose BH-NeRF, a novel tomography approach that leverages gravitational lensing to recover the continuous 3D emission field near a black hole. Compared to other 3D reconstruction or tomography settings, this task poses two significant challenges: first, rays near black holes follow curved paths dictated by general relativity, and second, we only observe measurements from a single viewpoint. Our method captures the unknown emission field using a continuous volumetric function parameterized by a coordinate-based neural network, and uses knowledge of Keplerian orbital dynamics to establish correspondence between 3D points over time. Together, these enable BH-NeRF to recover accurate 3D emission fields, even in challenging situations with sparse measurements and uncertain orbital dynamics. This work takes the first steps in showing how future measurements from the Event Horizon Telescope could be used to recover evolving 3D emission around the supermassive black hole in our Galactic center.

Deep learning for graph matching has received growing interest and developed rapidly in the past decade. Although recent deep graph matching methods have shown excellent performance on matching between graphs of equal size in the computer vision area, the size-varied graph matching problem, where the number of keypoints in the images of the same category may vary due to occlusion, is still an open and challenging problem. To tackle this, we firstly propose to formulate the combinatorial problem of graph matching as an Integer Linear Programming (ILP) problem, which is more flexible and efficient to facilitate comparing graphs of varied sizes. A novel Graph-context Attention Network (GCAN), which jointly capture intrinsic graph structure and cross-graph information for improving the discrimination of node features, is then proposed and trained to resolve this ILP problem with node correspondence supervision. We further show that the proposed GCAN model is efficient to resolve the graph-level matching problem and is able to automatically learn node-to-node similarity via graph-level matching. The proposed approach is evaluated on three public keypoint-matching datasets and one graph-matching dataset for blood vessel patterns, with experimental results showing its superior performance over existing state-of-the-art algorithms on the keypoint and graph-level matching tasks.

Pedestrian trajectory prediction is an essential and challenging task for a variety of real-life applications such as autonomous driving and robotic motion planning. Besides generating a single future path, predicting multiple plausible future paths is becoming popular in some recent work on trajectory prediction. However, existing methods typically emphasize spatial interactions between pedestrians and surrounding areas but ignore the smoothness and temporal consistency of predictions. Our model aims to forecast multiple paths based on a historical trajectory by modeling multi-scale graph-based spatial transformers combined with a trajectory smoothing algorithm named "Memory Replay" utilizing a memory graph. Our method can comprehensively exploit the spatial information as well as correct the temporally inconsistent trajectories (e.g., sharp turns). We also propose a new evaluation metric named "Percentage of Trajectory Usage" to evaluate the comprehensiveness of diverse multi-future predictions. Our extensive experiments show that the proposed model achieves state-of-the-art performance on multi-future prediction and competitive results for single-future prediction. Code released at https://github.com/Jacobieee/ST-MR.