CVPR 2021 Papers

Traditional classifiers are deployed under closed-set setting, with both training and test classes belong to the same set. However, real-world applications probably face the input of unknown categories, and the model will recognize them as known ones. Under such circumstances, open-set recognition is proposed to maintain classification performance on known classes and reject unknowns. The closed-set models make overconfident predictions over familiar known class instances, so that calibration and thresholding across categories become essential issues when extending to an open-set environment. To this end, we proposed to learn PlaceholdeRs for Open-SEt Recognition (Proser), which prepares for the unknown classes by allocating placeholders for both data and classifier. In detail, learning data placeholders tries to anticipate open-set class data, thus transforms closed-set training into open-set training. Besides, to learn the invariant information between target and non-target classes, we reserve classifier placeholders as the class-specific boundary between known and unknown. The proposed Proser efficiently generates novel class by manifold mixup, and adaptively sets the value of reserved open-set classifier during training. Experiments on various datasets validate the effectiveness of our proposed method.

This paper focuses on the task of 4D shape reconstruction from a sequence of point clouds. Despite the recent success achieved by extending deep implicit representations into 4D space, it is still a great challenge in two respects, i.e. how to design a flexible framework for learning robust spatio-temporal shape representations from 4D point clouds, and develop an efficient mechanism for capturing shape dynamics. In this work, we present a novel pipeline to learn a temporal evolution of the 3D human shape through spatially continuous transformation functions among cross-frame occupancy fields. The key idea is to parallelly establish the dense correspondence between predicted occupancy fields at different time steps via explicitly learning continuous displacement vector fields from robust spatio-temporal shape representations. Extensive comparisons against previous state-of-the-arts show the superior accuracy of our approach for 4D human reconstruction in the problems of 4D shape auto-encoding and completion, and a much faster network inference with about 8 times speedup demonstrates the significant efficiency of our approach.

State-of-the-art neural network models for optical flow estimation require a dense correlation volume at high resolutions for representing per-pixel displacement. Although the dense correlation volume is informative for accurate estimation, its heavy computation and memory usage hinders the efficient training and deployment of the models. In this paper, we show that the dense correlation volume representation is redundant and accurate flow estimation can be achieved with only a fraction of elements in it. Based on this observation, we propose an alternative displacement representation, named Sparse Correlation Volume, which is constructed directly by computing the k closest matches in one feature map for each feature vector in the other feature map and stored in a sparse data structure. Experiments show that our method can reduce computational cost and memory use significantly and produce fine-structure motion, while maintaining high accuracy compared to previous approaches with dense correlation volumes.

This paper deals with the scarcity of data for training optical flow networks, highlighting the limitations of existing sources such as labeled synthetic datasets or unlabeled real videos. Specifically, we introduce a framework to generate accurate ground-truth optical flow annotations quickly and in large amounts from any readily available single real picture. Given an image, we use an off-the-shelf monocular depth estimation network to build a plausible point cloud for the observed scene. Then, we virtually move the camera in the reconstructed environment with known motion vectors and rotation angles, allowing us to synthesize both a novel view and the corresponding optical flow field connecting each pixel in the input image to the one in the new frame. When trained with our data, state-of-the-art optical flow networks achieve superior generalization to unseen real data compared to the same models trained either on annotated synthetic datasets or unlabeled videos, and better specialization if combined with synthetic images.

Frame reconstruction (current or future frames) based on Auto-Encoder (AE) is a popular method for video anomaly detection. With models trained on the normal data, the reconstruction errors of anomalous scenes are usually much larger than those of normal ones. Previous methods introduced the memory bank into AE, for encoding diverse normal patterns across the training videos. However, they are memory-consuming and cannot cope with unseen new scenarios in the training data. In this work, we propose a dynamic prototype unit (DPU) to encode the normal dynamics as prototypes in real time, free from extra memory cost. In addition, we introduce meta-learning to our DPU to form a novel few-shot normalcy learner, namely Meta-Prototype Unit (MPU). It enables the fast adaption capability on new scenes by only consuming a few iterations of update. Extensive experiments are conducted on various benchmarks. The superior performance over the state-of-the-art demonstrates the effectiveness of our method.

How to efficiently represent camera pose is an essential problem in 3D computer vision, especially in tasks like camera pose regression and novel view synthesis. Traditionally, 3D position of the camera is represented by Cartesian coordinate and the orientation is represented by Euler angle or quaternions. These representations are manually designed, which may not be the most efficient representation for downstream tasks. In this work, we propose an approach to learn neural representations of camera poses and 3D scenes, coupled with neural representations of local camera movements. Specifically, the camera pose and 3D scene are represented as vectors and the local camera movement is represented as a matrix operating on the vector of the camera pose. We demonstrate that the camera movement can further be parametrized as a matrix Lie algebra that underlies a rotation system in the neural space. The vector representations are then concatenated and generate the posed 2D image through a decoder network. The model is learned from only posed 2D images and corresponding camera poses, without access to depth or shape. We conduct extensive experiments on synthetic and real datasets. The results show that compared with other camera pose representations, our learned representation is more robust to noise in novel view synthesis and more effective in camera pose regression.

Capturing photographs with wrong exposures remains a major source of errors in camera-based imaging. Exposure problems are categorized as either: (i) overexposed, where the camera exposure was too long, resulting in bright and washed-out image regions, or (ii) underexposed, where the exposure was too short, resulting in dark regions. Both under- and overexposure greatly reduce the contrast and visual appeal of an image. Prior work mainly focuses on underexposed images or general image enhancement. In contrast, our proposed method targets both over- and under-exposure errors in photographs. We formulate the exposure correction problem as two main sub-problems: (i) color enhancement and (ii) detail enhancement. Accordingly, we propose a coarse-to-fine deep neural network (DNN) model, trainable in an end-to-end manner, that addresses each sub-problem separately. A key aspect of our solution is a new dataset of over 24,000 images exhibiting the broadest range of exposure values to date with a corresponding properly exposed image. Our method achieves results on par with existing state-of-the-art methods on underexposed images and yields significant improvements for images suffering from overexposure errors.

Monocular 3D reconstruction of articulated object categories is challenging due to the lack of training data and the inherent ill-posedness of the problem. In this work we use video self-supervision, forcing the consistency of consecutive 3D reconstructions by a motion-based cycle loss. This largely improves both optimization-based and learning-based 3D mesh reconstruction. We further introduce an interpretable model of 3D template deformations that controls a 3D surface through the displacement of a small number of local, learnable handles. We formulate this operation as a structured layer relying on mesh-laplacian regularization and show that it can be trained in an end-to-end manner. We finally introduce a per-sample numerical optimisation approach that jointly optimises over mesh displacements and cameras within a video, boosting accuracy both for training and also as test time post-processing. While relying exclusively on a small set of videos collected per category for supervision, we obtain state-of-the-art reconstructions with diverse shapes, viewpoints and textures for multiple articulated object categories.

The success of supervised learning crucially hinges on the assumption that training data matches test data, which rarely holds in practice due to potential distribution shift. In light of this, most existing methods for unsupervised domain adaptation focus on achieving domain-invariant representations and small source domain error. However, recent works have shown that this is not sufficient to guarantee good generalization on target domain and in fact is provably detrimental under label distribution shift. Furthermore, in many real-world applications it is often feasible to obtain a small amount of labeled data from the target domain and use them to facilitate model training with source data. Inspired by the above observations, in this paper we propose the first method that aims to simultaneously learn invariant representations and risks under the setting of semi-supervised domain adaptation (Semi-DA). To start with, we first give a finite sample bound for both classification and regression problems under Semi-DA. The bound suggests a principled way for target generalization by aligning both the marginal and conditional distributions across domains in feature space. Motivated by this, we then introduce our LIRR algorithm for jointly Learning Invariant Representations and Risks. Finally, we conduct extensive experiments on both classification and regression tasks to demonstrate the effectiveness of LIRR. Compared with methods that only learn invariant representations or invariant risks, LIRR achieves significant improvements.

A key challenge of learning the geometry of dressed humans lies in the limited availability of the ground truth data (e.g., 3D scanned models), which results in the performance degradation of 3D human reconstruction when applying to real world imagery. We address this challenge by leveraging a new data resource: a number of social media dance videos that span diverse appearance, clothing styles, performances, and identities. Each video depicts dynamic movements of the body and clothes of a single person while lacking the 3D ground truth geometry. To utilize these videos, we present a new method to use the local transformation that warps the predicted local geometry of the person from an image to that of the other image at a different time instant. With the transformation, the predicted geometry can be self-supervised by the warped geometry from the other image. In addition, we jointly learn the depth along with the surface normals, which are highly responsive to local texture, wrinkle, and shade by maximizing their geometric consistency. Our method is end-to-end trainable, resulting in high fidelity depth estimation that predicts fine geometry faithful to the input real image. We demonstrate that our method outperforms the state-of-the-art human depth estimation and human shape recovery approaches on both real and rendered images.

Fixed input graphs are a mainstay in approaches that utilize Graph Convolution Networks (GCNs) for knowledge transfer. The standard paradigm is to utilize relationships in the input graph to transfer information using GCNs from training to testing nodes in the graph; for example, the semi-supervised, zero-shot, and few-shot learning setups. We propose a generalized framework for learning and improving the input graph as part of the standard GCN-based learning setup. Moreover, we use additional constraints between similar and dissimilar neighbors for each node in the graph by applying triplet loss on the intermediate layer output. We present results of semi-supervised learning on Citeseer, Cora, and Pubmed benchmarking datasets, and zero/few-shot action recognition on UCF101 and HMDB51 datasets, significantly outperforming current approaches. We also present qualitative results visualizing the graph connections that our approach learns to update.

In compositional zero-shot learning, the goal is to recognize unseen compositions (e.g. old dog) of observed visual primitives states (e.g. old, cute) and objects (e.g. car, dog)in the training set. This is challenging because the same state can for example alter the visual appearance of a dog drastically differently from a car. As a solution, we propose a novel graph formulation called Compositional Graph Embedding (CGE) that learns image features, compositional classifiers, and latent representations of visual primitives in an end-to-end manner. The key to our approach is exploit-ing the dependency between states, objects, and their compositions within a graph structure to enforce the relevant knowledge transfer from seen to unseen compositions. By learning a joint compatibility that encodes semantics between concepts, our model allows for generalization to un-seen compositions without relying on an external knowledgebase like WordNet. We show that in the challenging generalized compositional zero-shot setting our CGE significantly outperforms the state of the art on MIT-States and UT-Zappos. We also propose a new benchmark for this task based on the recent GQA dataset.

We introduce a framework that predicts the goals behind observable human action in video. Motivated by evidence in developmental psychology, we leverage video of unintentional action to learn video representations of goals without direct supervision. Our approach models videos as contextual trajectories that represent both low-level motion and high-level action features. Experiments and visualizations show our trained model is able to predict the underlying goals in video of unintentional action. We also propose a method to "automatically correct" unintentional action by leveraging gradient signals of our model to adjust latent trajectories. Although the model is trained with minimal supervision, it is competitive with or outperforms baselines trained on large (supervised) datasets of successfully executed goals, showing that observing unintentional action is crucial to learning about goals in video.

Lip reading aims to predict the spoken sentences from silent lip videos. Due to the fact that such a vision task usually performs worse than its counterpart speech recognition, one potential scheme is to distill knowledge from a teacher pretrained by audio signals. However, the latent domain gap between the cross-modal data could lead to an learning ambiguity and thus limits the performance of lip reading. In this paper, we propose a novel collaborative framework for lip reading, and two aspects of issues are considered: 1) the teacher should understand bi-modal knowledge to possibly bridge the inherent cross-modal gap; 2) the teacher should adjust teaching contents adaptively with the evolution of the student. To these ends, we introduce a trainable "master" network which ingests both audio signals and silent lip videos instead of a pretrained teacher. The master produces logits from three modalities of features: audio modality, video modality, and their combination. To further provide an interactive strategy to fuse these knowledge organically, we regularize the master with the task-specific feedback from the student, in which the requirement of the student is implicitly embedded. Meanwhile we involve a couple of "tutor" networks into our system as guidance for emphasizing the fruitful knowledge flexibly. In addition, we incorporate a curriculum learning design to ensure a better convergence. Extensive experiments demonstrate that the proposed network outperforms the state-of-the-art methods on several benchmarks, including in both word-level and sentence-level scenarios.

Existing learning-based approaches to 3D shape segmentation usually formulate it as a semantic labeling problem, assuming that all parts of training shapes are annotated with a given set of labels. This assumption, however, is unrealistic for training fine-grained segmentation on large datasets since the annotation of fine-grained parts is extremely tedious. In this paper, we approach the problem with deep clustering, where the key idea is to learn part priors from a dataset with fine-grained segmentation but no part annotations. Given point sampled 3D shapes, we model the clustering priors of points with a similarity matrix and achieve part-based segmentation through minimizing a novel low rank loss. Further, since fine-grained parts can be very tiny, a 3D shape has to be densely sampled to ensure the tiny parts are well captured and segmented. To handle densely sampled point sets, we adopt a divide-and-conquer scheme. We first partition the large point set into a number of blocks. Each block is segmented using a deep-clustering-based part prior network (PriorNet) trained in a category-agnostic manner. We then train MergeNet, a graph convolution network, to merge the segments of all blocks to form the final segmentation result. Our method is evaluated with a challenging benchmark of fine-grained segmentation, showing significant advantage over the state-of-the-art ones.

3D morphable models are widely used for the shape representation of an object class in computer vision and graphics applications. In this work, we focus on deep 3D morphable models that directly apply deep learning on 3D mesh data with a hierarchical structure to capture information at multiple scales. While great efforts have been made to design the convolution operator, how to best aggregate vertex features across hierarchical levels deserves further attention. In contrast to resorting to mesh decimation, we propose an attention based module to learn mapping matrices for better feature aggregation across hierarchical levels. Specifically, the mapping matrices are generated by a compatibility function of the keys and queries. The keys and queries are trainable variables, learned by optimizing the target objective, and shared by all data samples of the same object class. Our proposed module can be used as a train-only drop-in replacement for the feature aggregation in existing architectures for both downsampling and upsampling. Our experiments show that through the end-to-end training of the mapping matrices, we achieve state-of-the-art results on a variety of 3D shape datasets in comparison to existing morphable models.

Multi-person pose estimation and tracking serve as crucial steps for video understanding. Most state-of-the-art approaches rely on first estimating poses in each frame and only then implementing data association and refinement. Despite the promising results achieved, such a strategy is inevitably prone to missed detections especially in heavily-cluttered scenes, since this tracking-by-detection paradigm is, by nature, largely dependent on visual evidences that are absent in the case of occlusion. In this paper, we propose a novel online approach to learning the pose dynamics, which are independent of pose detections in current fame, and hence may serve as a robust estimation even in challenging scenarios including occlusion. Specifically, we derive this prediction of dynamics through a graph neural network (GNN) that explicitly accounts for both spatial-temporal and visual information. It takes as input the historical pose tracklets and directly predicts the corresponding poses in the following frame for each tracklet. The predicted poses will then be aggregated with the detected poses, if any, at the same frame so as to produce the final pose, potentially recovering the occluded joints missed by the estimator. Experiments on PoseTrack 2017 and PoseTrack 2018 datasets demonstrate that the proposed method achieves results superior to the state of the art on both human pose estimation and tracking tasks.

Current state-of-the-art approaches for Semi-supervised Video Object Segmentation (Semi-VOS) propagates information from previous frames to generate segmentation mask for the current frame. This results in high-quality segmentation across challenging scenarios such as changes in appearance and occlusion. But it also leads to unnecessary computations for stationary or slow-moving objects where the change across frames is minimal. In this work, we exploit this observation by using temporal information to quickly identify frames with minimal change and skip the heavyweight mask generation step. To realize this efficiency, we propose a novel dynamic network that estimates change across frames and decides which path -- computing a full network or reusing previous frame's feature -- to choose depending on the expected similarity. Experimental results show that our approach significantly improves inference speed without much accuracy degradation on challenging Semi-VOS datasets -- DAVIS 16, DAVIS 17, and YouTube-VOS. Furthermore, our approach can be applied to multiple Semi-VOS methods demonstrating its generality. The code is available in https://github.com/HYOJINPARK/Reuse VOS.

Few-shot learning (FSL), which aims to recognise new classes by adapting the learned knowledge with extremely limited few-shot (support) examples, remains an important open problem in computer vision. Most of the existing methods for feature alignment in few-shot learning only consider image-level or spatial-level alignment while omitting the channel disparity. Our insight is that these methods would lead to poor adaptation with redundant matching, and leveraging channel-wise adjustment is the key to well adapting the learned knowledge to new classes. Therefore, in this paper, we propose to learn a dynamic alignment, which can effectively highlight both query regions and channels according to different local support information. Specifically, this is achieved by first dynamically sampling the neighbourhood of the feature position conditioned on the input few shot, based on which we further predict a both position-dependent and channel-dependent Dynamic Meta-filter. The filter is used to align the query feature with position-specific and channel-specific knowledge. Moreover, we adopt Neural Ordinary Differential Equation (ODE) to enable a more accurate control of the alignment. In such a sense our model is able to better capture fine-grained semantic context of the few-shot example and thus facilitates dynamical knowledge adaptation for few-shot learning. The resulting framework establishes the new state-of-the-arts on major few-shot visual recognition benchmarks, including miniImageNet and tieredImageNet.