CVPR 2022 论文列表

We consider the problem of recovering a single person's 3D human mesh from in-the-wild crowded scenes. While much progress has been in 3D human mesh estimation, existing methods struggle when test input has crowded scenes. The first reason for the failure is a domain gap between training and testing data. A motion capture dataset, which provides accurate 3D labels for training, lacks crowd data and impedes a network from learning crowded scene-robust image features of a target person. The second reason is a feature processing that spatially averages the feature map of a localized bounding box containing multiple people. Averaging the whole feature map makes a target person's feature indistinguishable from others. We present 3DCrowdNet that firstly explicitly targets in-the-wild crowded scenes and estimates a robust 3D human mesh by addressing the above issues. First, we leverage 2D human pose estimation that does not require a motion capture dataset with 3D labels for training and does not suffer from the domain gap. Second, we propose a joint-based regressor that distinguishes a target person's feature from others. Our joint-based regressor preserves the spatial activation of a target by sampling features from the target's joint locations and regresses human model parameters. As a result, 3DCrowdNet learns target-focused features and effectively excludes the irrelevant features of nearby persons. We conduct experiments on various benchmarks and prove the robustness of 3DCrowdNet to the in-the-wild crowded scenes both quantitatively and qualitatively. Codes are available here: https://github.com/hongsukchoi/3DCrowdNet_RELEASE

Current 3D object detectors for autonomous driving are almost entirely trained on human-annotated data. Although of high quality, the generation of such data is laborious and costly, restricting them to a few specific locations and object types. This paper proposes an alternative approach entirely based on unlabeled data, which can be collected cheaply and in abundance almost everywhere on earth. Our approach leverages several simple common sense heuristics to create an initial set of approximate seed labels. For example, relevant traffic participants are generally not persistent across multiple traversals of the same route, do not fly, and are never under ground. We demonstrate that these seed labels are highly effective to bootstrap a surprisingly accurate detector through repeated self-training without a single human annotated label. Code is available at https://github.com/YurongYou/MODEST.

Learning personalized models for user-customized computer-vision tasks is challenging due to the limited private-data and computation available on each edge device. Decentralized learning (DL) can exploit the images distributed over devices on a network topology to train a global model but is not designed to train personalized models for different tasks or optimize the topology. Moreover, the mixing weights used to aggregate neighbors' gradient messages in DL can be sub-optimal for personalization since they are not adaptive to different nodes/tasks and learning stages. In this paper, we dynamically update the mixing-weights to improve the personalized model for each node's task and meanwhile learn a sparse topology to reduce communication costs. Our first approach, "learning to collaborate (L2C)", directly optimizes the mixing weights to minimize the local validation loss per node for a pre-defined set of nodes/tasks. In order to produce mixing weights for new nodes or tasks, we further develop "meta-L2C", which learns an attention mechanism to automatically assign mixing weights by comparing two nodes' model updates. We evaluate both methods on diverse benchmarks and experimental settings for image classification. Thorough comparisons to both classical and recent methods for IID/non-IID decentralized and federated learning demonstrate our method's advantages in identifying collaborators among nodes, learning sparse topology, and producing better personalized models with low communication and computational cost.

Anticipating future events is an essential feature for intelligent systems and embodied AI. However, compared to the traditional recognition task, the uncertainty of future and reasoning ability requirement make the anticipation task very challenging and far beyond solved. In this filed, previous methods usually care more about the model architecture design or but few attention has been put on how to train an anticipation model with a proper learning policy. To this end, in this work, we propose a novel training scheme called Dynamic Context Removal (DCR), which dynamically schedule the visibility of observed future in the learning procedure. It follows the human-like curriculum learning process, i.e., gradually removing the event context to increase the anticipation difficulty till satisfying the final anticipation target. Our learning scheme is plug-and-play and easy to integrate any reasoning model including transformer and LSTM, with advantages in both effectiveness and efficiency. In extensive experiments, the proposed method achieves state-of-the-art on four widely-used benchmarks. Our code and models are publicly released at https://github.com/AllenXuuu/DCR.

State-of-the-art methods for self-supervised sequential action alignment rely on deep networks that find correspondences across videos in time. They either learn frame-to-frame mapping across sequences, which does not leverage temporal information, or assume monotonic alignment between each video pair, which ignores variations in the order of actions. As such, these methods are not able to deal with common real-world scenarios that involve background frames or videos that contain non-monotonic sequence of actions. In this paper, we propose an approach to align sequential actions in the wild that involve diverse temporal variations. To this end, we propose an approach to enforce temporal priors on the optimal transport matrix, which leverages temporal consistency, while allowing for variations in the order of actions. Our model accounts for both monotonic and non-monotonic sequences and handles background frames that should not be aligned. We demonstrate that our approach consistently outperforms the state-of-the-art in self-supervised sequential action representation learning on four different benchmark datasets.

Few-shot learning (FSL) aims to learn a classifier that can be easily adapted to accommodate new tasks, given only a few examples. To handle the limited-data in few-shot regimes, recent methods tend to collectively use a set of local features to densely represent an image instead of using a mixed global feature. They generally explore a unidirectional paradigm, e.g., find the nearest support feature for every query feature and aggregate these local matches for a joint classification. In this paper, we propose a novel Mutual Centralized Learning (MCL) to fully affiliate these two disjoint dense features sets in a bidirectional paradigm. We first associate each local feature with a particle that can bidirectionally random walk in a discrete feature space. To estimate the class probability, we propose the dense features' accessibility that measures the expected number of visits to the dense features of that class in a Markov process. We relate our method to learning a centrality on an affiliation network and demonstrate its capability to be plugged in existing methods by highlighting centralized local features. Experiments show that our method achieves the new state-of-the-art.

This paper proposes using a sparse-structured multivariate Gaussian to provide a closed-form approximator for the output of probabilistic ensemble models used for dense image prediction tasks. This is achieved through a convolutional neural network that predicts the mean and covariance of the distribution, where the inverse covariance is parameterised by a sparsely structured Cholesky matrix. Similarly to distillation approaches, our single network is trained to maximise the probability of samples from pre-trained probabilistic models, in this work we use a fixed ensemble of networks. Once trained, our compact representation can be used to efficiently draw spatially correlated samples from the approximated output distribution. Importantly, this approach captures the uncertainty and structured correlations in the predictions explicitly in a formal distribution, rather than implicitly through sampling alone. This allows direct introspection of the model, enabling visualisation of the learned structure. Moreover, this formulation provides two further benefits: estimation of a sample probability, and the introduction of arbitrary spatial conditioning at test time. We demonstrate the merits of our approach on monocular depth estimation and show that the advantages of our approach are obtained with comparable quantitative performance.

Estimating keypoint scale and orientation is crucial to extracting invariant features under significant geometric changes. Recently, the estimators based on self-supervised learning have been designed to adapt to complex imaging conditions. Such learning-based estimators generally predict a single scalar for the keypoint scale or orientation, called hard estimators. However, hard estimators are difficult to handle the local patches containing structures of different objects or multiple edges. In this paper, a Soft Self-Supervised Estimator (S3Esti) is proposed to overcome this problem by learning to predict multiple scales and orientations. S3Esti involves three core factors. First, the estimator is constructed to predict the discrete distributions of scales and orientations. The elements with high confidence will be kept as the final scales and orientations. Second, a probabilistic covariant loss is proposed to improve the consistency of the scale and orientation distributions under different transformations. Third, an optimization algorithm is designed to minimize the loss function, whose convergence is proved in theory. When combined with different keypoint extraction models, S3Esti generally improves over 50% accuracy in image matching tasks under significant viewpoint changes. In the 3D reconstruction task, S3Esti decreases more than 10% reprojection error and improves the number of registered images.

Mirrors generally lack a consistent visual appearance, making mirror detection very challenging. Although recent works that are based on exploiting contextual contrasts and corresponding relations have achieved good results, heavily relying on contextual contrasts and corresponding relations to discover mirrors tend to fail in complex real-world scenes, where a lot of objects, e.g., doorways, may have similar features as mirrors. We observe that humans tend to place mirrors in relation to certain objects for specific functional purposes, e.g., a mirror above the sink. Inspired by this observation, we propose a model to exploit the semantic associations between the mirror and its surrounding objects for a reliable mirror localization. Our model first acquires class-specific knowledge of the surrounding objects via a semantic side-path. It then uses two novel modules to exploit semantic associations: 1) an Associations Exploration (AE) Module to extract the associations of the scene objects based on fully connected graph models, and 2) a Quadruple-Graph (QG) Module to facilitate the diffusion and aggregation of semantic association knowledge using graph convolutions. Extensive experiments show that our method outperforms the existing methods and sets the new state-of-the-art on both PMD dataset (f-measure: 0.844) and MSD dataset (f-measure: 0.889).

In this work, we propose a novel method to improve the generalization ability of CNN-based face forgery detectors. Our method considers the feature anomalies of forged faces caused by the prevalent blending operations in face forgery algorithms. Specifically, we propose a weakly supervised Second Order Local Anomaly (SOLA) learning module to mine anomalies in local regions using deep feature maps. SOLA first decomposes the neighborhood of local features by different directions and distances and then calculates the first and second order local anomaly maps which provide more general forgery traces for the classifier. We also propose a Local Enhancement Module (LEM) to improve the discrimination between local features of real and forged regions, so as to ensure accuracy in calculating anomalies. Besides, an improved Adaptive Spatial Rich Model (ASRM) is introduced to help mine subtle noise features via learnable high pass filters. With neither pixel level annotations nor external synthetic data, our method using a simple ResNet18 backbone achieves competitive performances compared with state-of-the-art works when evaluated on unseen forgeries.

Recent image-based rendering (IBR) methods usually adopt plenty of views to reconstruct dense scene geometry. However, the number of available views is limited in practice. When only few views are provided, the performance of these methods drops off significantly, as the scene geometry becomes sparse as well. Therefore, in this paper, we propose Sparse-IBRNet (SIBRNet) to perform robust IBR on sparse scene geometry by depth completion. The SIBRNet has two stages, geometry recovery (GR) stage and light blending (LB) stage. Specifically, GR stage takes sparse depth map and RGB as input to predict dense depth map by exploiting the correlation between two modals. As inaccuracy of the complete depth map may cause projection biases in the warping process, LB stage first uses a bias-corrected module (BCM) to rectify deviations, and then aggregates modified features from different views to render a novel view. Extensive experimental results demonstrate that our method performs best on sparse scene geometry than recent IBR methods, and it can generate better or comparable results as well when the geometric information is dense.

The goal of video highlight detection is to select the most attractive segments from a long video to depict the most interesting parts of the video. Existing methods typically focus on modeling relationship between different video segments in order to learning a model that can assign highlight scores to these segments; however, these approaches do not explicitly consider the contextual dependency within individual segments. To this end, we propose to learn pixel-level distinctions to improve the video highlight detection. This pixel-level distinction indicates whether or not each pixel in one video belongs to an interesting section. The advantages of modeling such fine-level distinctions are two-fold. First, it allows us to exploit the temporal and spatial relations of the content in one video, since the distinction of a pixel in one frame is highly dependent on both the content before this frame and the content around this pixel in this frame. Second, learning the pixel-level distinction also gives a good explanation to the video highlight task regarding what contents in a highlight segment will be attractive to people. We design an encoder-decoder network to estimate the pixel-level distinction, in which we leverage the 3D convolutional neural networks to exploit the temporal context information, and further take advantage of the visual saliency to model the spatial distinction. State-of-the-art performance on three public benchmarks clearly validates the effectiveness of our framework for video highlight detection.

A range of video modeling tasks, from optical flow to multiple object tracking, share the same fundamental challenge: establishing space-time correspondence. Yet, approaches that dominate each space differ. We take a step towards bridging this gap by extending the recent contrastive random walk formulation to much more dense, pixel-level space-time graphs. The main contribution is introducing hierarchy into the search problem by computing the transition matrix in a coarse-to-fine manner, forming a multiscale contrastive random walk. This establishes a unified technique for self-supervised learning of optical flow, keypoint tracking, and video object segmentation. Experiments demonstrate that, for each of these tasks, our unified model achieves performance competitive with strong self-supervised approaches specific to that task.

The inherent slow imaging speed of Magnetic Resonance Image (MRI) has spurred the development of various acceleration methods, typically through heuristically undersampling of the associated measurement domain known as k-space. Recently, deep neural networks have been applied to reconstruct undersampled k-space and shown improved reconstruction performance. While most methods focus on designing novel reconstruction networks or new training strategies for a given undersampling pattern, e.g., random Cartesian undersampling or standard non-Cartesian sampling, to date, there is limited research that aims to learn and optimize k-space sampling strategies using deep neural networks. In this work, we propose a novel framework to learn optimized k-space sampling trajectories using deep learning by considering it as an Ordinary Differential Equation (ODE) problem that can be solved using neural ODE. In particular, the sampling of k-space data is framed as a dynamic system, in which the control points serve as an initial state and a physical-conditioned neural ODE is formulated to approximate the system. Moreover, we also enforce additional constraints on gradient slew rate and amplitude in trajectory learning, so that severe gradient-indued artifacts can be minimized. Furthermore, we have also demonstrated that sampling trajectory optimization and MRI reconstruction can be jointly trained, such that the optimized trajectory is task-oriented and can enhance overall image reconstruction performance. Experiments were conducted on different in-vivo dataset (e.g., Brain and Knee) with different contrast. Initial results have shown that our proposed method is able to generate better image quality in accelerated MRI compared to conventional undersampling schemes in both Cartesian and non-Cartesian acquisitions.

Optical flow is a fundamental method used for quantitative motion estimation on the image plane. In the deep learning era, most works treat it as a task of 'matching of features', learning to pull matched pixels as close as possible in feature space and vice versa. However, spatial affinity (smoothness constraint), another important component for motion understanding, has been largely overlooked. In this paper, we introduce a novel approach, called kernel patch attention (KPA), to better resolve the ambiguity in dense matching by explicitly taking the local context relations into consideration. Our KPA operates on each local patch, and learns to mine the context affinities for better inferring the flow fields. It can be plugged into contemporary optical flow architecture and empower the model to conduct comprehensive motion analysis with both feature similarities and spatial relations. On Sintel dataset, the proposed KPA-Flow achieves the best performance with EPE of 1.35 on clean pass and 2.36 on final pass, and it sets a new record of 4.60% in F1-all on KITTI-15 benchmark.

Novel-view prediction of a scene has many applications. Existing works mainly focus on generating novel-view images via pixel-wise prediction in the image space, often resulting in severe ghosting and blurry artifacts. In this paper, we make the first attempt to explore novel-view prediction in the layout space, and introduce the new problem of novel-view scene layout generation. Given a single scene layout and the camera transformation as inputs, our goal is to generate a plausible scene layout for a specified viewpoint. Such a problem is challenging as it involves accurate understanding of the 3D geometry and semantics of the scene from as little as a single 2D scene layout. To tackle this challenging problem, we propose a deep model to capture contextualized object representation by explicitly modeling the object context transformation in the scene. The contextualized object representation is essential in generating geometrically and semantically consistent scene layouts of different views. Experiments show that our model outperforms several strong baselines on many indoor and outdoor scenes, both qualitatively and quantitatively. We also show that our model enables a wide range of applications, including novel-view image synthesis, novel-view image editing, and amodal object estimation.

Existing studies in few-shot semantic segmentation only focus on mining the target object information, however, often are hard to tell ambiguous regions, especially in non-target regions, which include background (BG) and Distracting Objects (DOs). To alleviate this problem, we propose a novel framework, namely Non-Target Region Eliminating (NTRE) network, to explicitly mine and eliminate BG and DO regions in the query. First, a BG Mining Module (BGMM) is proposed to extract the BG region via learning a general BG prototype. To this end, we design a BG loss to supervise the learning of BGMM only using the known target object segmentation ground truth. Then, a BG Eliminating Module and a DO Eliminating Module are proposed to successively filter out the BG and DO information from the query feature, based on which we can obtain a BG and DO-free target object segmentation result. Furthermore, we propose a prototypical contrastive learning algorithm to improve the model ability of distinguishing the target object from DOs. Extensive experiments on both PASCAL- 5^ i and COCO- 20^ i datasets show that our approach is effective despite its simplicity. Code is available at https://github.com/LIUYUANWEI98/NERTNet

Neural radiance fields (NeRFs) produce state-of-the-art view synthesis results, but are slow to render, requiring hundreds of network evaluations per pixel to approximate a volume rendering integral. Baking NeRFs into explicit data structures enables efficient rendering, but results in large memory footprints and, in some cases, quality reduction. Additionally, volumetric representations for view synthesis often struggle to represent challenging view dependent effects such as distorted reflections and refractions. We present a novel neural light field representation that, in contrast to prior work, is fast, memory efficient, and excels at modeling complicated view dependence. Our method supports rendering with a single network evaluation per pixel for small baseline light fields and with only a few evaluations per pixel for light fields with larger baselines. At the core of our approach is a ray-space embedding network that maps 4D ray-space into an intermediate, interpolable latent space. Our method achieves state-of-the-art quality on dense forward-facing datasets such as the Stanford Light Field dataset. In addition, for forward-facing scenes with sparser inputs we achieve results that are competitive with NeRF-based approaches while providing a better speed/quality/memory trade-off with far fewer network evaluations.

In this paper, an ill-posed problem of multiple adverse weather removal is investigated. Our goal is to train a model with a 'unified' architecture and only one set of pretrained weights that can tackle multiple types of adverse weathers such as haze, snow, and rain simultaneously. To this end, a two-stage knowledge learning mechanism including knowledge collation (KC) and knowledge examination (KE) based on a multi-teacher and student architecture is proposed. At the KC, the student network aims to learn the comprehensive bad weather removal problem from multiple well-trained teacher networks where each of them is specialized in a specific bad weather removal problem. To accomplish this process, a novel collaborative knowledge transfer is proposed. At the KE, the student model is trained without the teacher networks and examined by challenging pixel loss derived by the ground truth. Moreover, to improve the performance of our training framework, a novel loss function called multi-contrastive knowledge regularization (MCR) loss is proposed. Experiments on several datasets show that our student model can achieve promising performance on different bad weather removal tasks simultaneously.