ICCV 2021 Papers

While predicting robot grasps with parallel jaw grippers have been well studied and widely applied in robot manipulation tasks, the study on natural human grasp generation with a multi-finger hand remains a very challenging problem. In this paper, we propose to generate human grasps given a 3D object in the world. Our key observation is that it is crucial to model the consistency between the hand contact points and object contact regions. That is, we encourage the prior hand contact points to be close to the object surface and the object common contact regions to be touched by the hand at the same time. Based on the hand-object contact consistency, we design novel objectives in training the human grasp generation model and also a new self-supervised task which allows the grasp generation network to be adjusted even during test time. Our experiments show significant improvement in human grasp generation over state-of-the-art approaches by a large margin. More interestingly, by optimizing the model during test time with the self-supervised task, it helps achieve larger gain on unseen and out-of-domain objects. Project page: https://hwjiang1510.github.io/GraspTTA/.

Analyzing and understanding hand information from multimedia materials like images or videos is important for many real world applications and remains to be very active in research community. There are various works focusing on recovering hand information from single image, however, they usually solve a single task, for example, hand mask segmentation, 2D/3D hand pose estimation, or hand mesh reconstruction and perform not well in challenging scenarios. To further improve the performance of these tasks, we propose a novel Hand Image Understanding (HIU) framework (HIU-DMTL) to extract comprehensive information of the hand object from a single RGB image, by jointly considering the relationships between these tasks. To achieve this goal, a cascaded multi-task learning (MTL) backbone is designed to estimate the 2D heat maps, to learn the segmentation mask, and to generate the intermediate 3D information encoding, followed by a coarse-to-fine learning paradigm and a self-supervised learning strategy. Qualitative experiments demonstrate that our approach is capable of recovering reasonable mesh representations even in challenging situations. Quantitatively, our method significantly outperforms the state-of-the-art approaches on various widely-used datasets, in terms of diverse evaluation metrics.

Point cloud registration is a fundamental problem in 3D computer vision. Outdoor LiDAR point clouds are typically large-scale and complexly distributed, which makes the registration challenging. In this paper, we propose an efficient hierarchical network named HRegNet for large-scale outdoor LiDAR point cloud registration. Instead of using all points in the point clouds, HRegNet performs registration on hierarchically extracted keypoints and descriptors. The overall framework combines the reliable features in deeper layer and the precise position information in shallower layers to achieve robust and precise registration. We present a correspondence network to generate correct and accurate keypoints correspondences. Moreover, bilateral consensus and neighborhood consensus are introduced for keypoints matching and novel similarity features are designed to incorporate them into the correspondence network, which significantly improves the registration performance. Besides, the whole network is also highly efficient since only a small number of keypoints are used for registration. Extensive experiments are conducted on two large-scale outdoor LiDAR point cloud datasets to demonstrate the high accuracy and efficiency of the proposed HRegNet. The project website is https://ispc-group.github.io/hregnet.

This paper introduces HPNet, a novel deep-learning approach for segmenting a 3D shape represented as a point cloud into primitive patches. The key to deep primitive segmentation is learning a feature representation that can separate points of different primitives. Unlike utilizing a single feature representation, HPNet leverages hybrid representations that combine one learned semantic descriptor, two spectral descriptors derived from predicted geometric parameters, as well as an adjacency matrix that encodes sharp edges. Moreover, instead of merely concatenating the descriptors, HPNet optimally combines hybrid representations by learning combination weights. This weighting module builds on the entropy of input features. The output primitive segmentation is obtained from a mean-shift clustering module. Experimental results on benchmark datasets ANSI and ABCParts show that HPNet leads to significant performance gains from baseline approaches.

Low-latency deep spiking neural networks (SNNs) have become a promising alternative to conventional artificial neural networks (ANNs) because of their potential for increased energy efficiency on event-driven neuromorphic hardware. Neural networks, including SNNs, however, are subject to various adversarial attacks and must be trained to remain resilient against such attacks for many applications. Nevertheless, due to prohibitively high training costs associated with SNNs, analysis, and optimization of deep SNNs under various adversarial attacks have been largely overlooked. In this paper, we first present a detailed analysis of the inherent robustness of low-latency SNNs against popular gradient-based attacks, namely fast gradient sign method (FGSM) and projected gradient descent (PGD). Motivated by this analysis, to harness the model robustness against these attacks we present an SNN training algorithm that uses crafted input noise and incurs no additional training time. To evaluate the merits of our algorithm, we conducted extensive experiments with variants of VGG and ResNet on both CIFAR-10 and CIFAR-100 datasets. Compared to standard trained direct input SNNs, our trained models yield improved classification accuracy of up to 13.7% and 10.1% on FGSM and PGD attack-generated images, respectively, with negligible loss in clean image accuracy. Our models also outperform inherently-robust SNNs trained on rate-coded inputs with improved or similar classification performance on attack-generated im-ages while having up to 25x and 4.6x lower latency and computation energy, respectively.

High dynamic range (HDR) video reconstruction from sequences captured with alternating exposures is a very challenging problem. Existing methods often align low dynamic range (LDR) input sequence in the image space using optical flow, and then merge the aligned images to produce HDR output. However, accurate alignment and fusion in the image space are difficult due to the missing details in the over-exposed regions and noise in the under-exposed regions, resulting in unpleasing ghosting artifacts. To enable more accurate alignment and HDR fusion, we introduce a coarse-to-fine deep learning framework for HDR video reconstruction. Firstly, we perform coarse alignment and pixel blending in the image space to estimate the coarse HDR video. Secondly, we conduct more sophisticated alignment and temporal fusion in the feature space of the coarse HDR video to produce better reconstruction. Considering the fact that there is no publicly available dataset for quantitative and comprehensive evaluation of HDR video reconstruction methods, we collect such a benchmark dataset, which contains 97 sequences of static scenes and 184 testing pairs of dynamic scenes. Extensive experiments show that our method outperforms previous state-of-the-art methods. Our dataset, code and model will be made publicly available.

Relational reasoning is at the heart of video question answering. However, existing approaches suffer from several common limitations: (1) they only focus on either object-level or frame-level relational reasoning, and fail to integrate the both; and (2) they neglect to leverage semantic knowledge for relational reasoning. In this work, we propose a Hierarchical VisuAl-Semantic RelatIonal Reasoning (HAIR) framework to address these limitations. Specifically, we present a novel graph memory mechanism to perform relational reasoning, and further develop two types of graph memory: a) visual graph memory that leverages visual information of video for relational reasoning; b) semantic graph memory that is specifically designed to explicitly leverage semantic knowledge contained in the classes and attributes of video objects, and perform relational reasoning in the semantic space. Taking advantage of both graph memory mechanisms, we build a hierarchical framework to enable visual-semantic relational reasoning from object level to frame level. Experiments on four challenging benchmark datasets show that the proposed framework leads to state-of-the-art performance, with fewer parameters and faster inference speed. Besides, our approach also shows superior performance on other video+language task.

We contribute HAA500, a manually annotated human-centric atomic action dataset for action recognition on 500 classes with over 591K labeled frames. To minimize ambiguities in action classification, HAA500 consists of highly diversified classes of fine-grained atomic actions, where only consistent actions fall under the same label, e.g., ""Baseball Pitching"" vs ""Free Throw in Basketball"". Thus HAA500 is different from existing atomic action datasets, where coarse-grained atomic actions were labeled with coarse action-verbs such as ""Throw"". HAA500 has been carefully curated to capture the precise movement of human figures with little class-irrelevant motions or spatio-temporal label noises. The advantages of HAA500 are fourfold: 1) human-centric actions with a high average of 69.7% detectable joints for the relevant human poses; 2) high scalability since adding a new class can be done under 20-60 minutes; 3) curated videos capturing essential elements of an atomic action without irrelevant frames; 4) fine-grained atomic action classes. Our extensive experiments including cross-data validation using datasets collected in the wild demonstrate the clear benefits of human-centric and atomic characteristics of HAA500, which enable training even a baseline deep learning model to improve prediction by attending to atomic human poses. We detail the HAA500 dataset statistics and collection methodology and compare quantitatively with existing action recognition datasets.

Recent learning approaches that implicitly represent surface geometry using coordinate-based neural representations have shown impressive results in the problem of multi-view 3D reconstruction. The effectiveness of these techniques is, however, subject to the availability of a large number (several tens) of input views of the scene, and computationally demanding optimizations. In this paper, we tackle these limitations for the specific problem of few-shot full 3D head reconstruction, by endowing coordinate-based representations with a probabilistic shape prior that enables faster convergence and better generalization when using few input images (down to three). First, we learn a shape model of 3D heads from thousands of incomplete raw scans using implicit representations. At test time, we jointly overfit two coordinate-based neural networks to the scene, one modeling the geometry and another estimating the surface radiance, using implicit differentiable rendering. We devise a two-stage optimization strategy in which the learned prior is used to initialize and constrain the geometry during an initial optimization phase. Then, the prior is unfrozen and fine-tuned to the scene. By doing this, we achieve high-fidelity head reconstructions, including hair and shoulders, and with a high level of detail that consistently outperforms both state-of-the-art 3D Morphable Models methods in the few-shot scenario, and non-parametric methods when large sets of views are available.

We present a comprehensive framework for egocentric interaction recognition using markerless 3D annotations of two hands manipulating objects. To this end, we propose a method to create a unified dataset for egocentric 3D interaction recognition. Our method produces annotations of the 3D pose of two hands and the 6D pose of the manipulated objects, along with their interaction labels for each frame. Our dataset, called H2O (2 Hands and Objects), provides synchronized multi-view RGB-D images, interaction labels, object classes, ground-truth 3D poses for left & right hands, 6D object poses, ground-truth camera poses, object meshes and scene point clouds. To the best of our knowledge, this is the first benchmark that enables the study of first-person actions with the use of the pose of both left and right hands manipulating objects and presents an unprecedented level of detail for egocentric 3D interaction recognition. We further propose the method to predict interaction classes by estimating the 3D pose of two hands and the 6D pose of the manipulated objects, jointly from RGB images. Our method models both inter- and intra-dependencies between both hands and objects by learning the topology of a graph convolutional network that predicts interactions. We show that our method facilitated by this dataset establishes a strong baseline for joint hand-object pose estimation and achieves state-of-the-art accuracy for first person interaction recognition.

Object handover is a common human collaboration behavior that attracts attention from researchers in Robotics and Cognitive Science. Though visual perception plays an important role in the object handover task, the whole handover process has been specifically explored. In this work, we propose a novel rich-annotated dataset, H2O, for visual analysis of human-human object handovers. The H2O, which contains 18K video clips involving 15 people who hand over 30 objects to each other, is a multi-purpose benchmark. It can support several vision-based tasks, from which, we specifically provide a baseline method, RGPNet, for a less-explored task named Receiver Grasp Prediction. Extensive experiments show that the RGPNet can produce plausible grasps based on the giver's hand-object states in the pre-handover phase. Besides, we also report the hand and object pose errors with existing baselines and show that the dataset can serve as the video demonstrations for robot imitation learning on the handover task.

Existing optical flow methods are erroneous in challenging scenes, such as fog, rain, and night because the basic optical flow assumptions such as brightness and gradient constancy are broken. To address this problem, we present an unsupervised learning approach that fuses gyroscope into optical flow learning. Specifically, we first convert gyroscope readings into motion fields named gyro field. Second, we design a self-guided fusion module to fuse the background motion extracted from the gyro field with the optical flow and guide the network to focus on motion details. To the best of our knowledge, this is the first deep learning-based framework that fuses gyroscope data and image content for optical flow learning. To validate our method, we propose a new dataset that covers regular and challenging scenes. Experiments show that our method outperforms the state-of-art methods in both regular and challenging scenes. Code and dataset are available at https://github.com/megvii-research/GyroFlow.

Rapid progress in 3D semantic segmentation is inseparable from the advances of deep network models, which highly rely on large-scale annotated data for training. To address the high cost and challenges of 3D point-level labeling, we present a method for semi-supervised point cloud semantic segmentation to adopt unlabeled point clouds in training to boost the model performance. Inspired by the recent contrastive loss in self-supervised tasks, we propose the guided point contrastive loss to enhance the feature representation and model generalization ability in semi-supervised setting. Semantic predictions on unlabeled point clouds serve as pseudo-label guidance in our loss to avoid negative pairs in the same category. Also, we design the confidence guidance to ensure high-quality feature learning. Besides, a category-balanced sampling strategy is proposed to collect positive and negative samples to mitigate the class imbalance problem. Extensive experiments on three datasets (ScanNet V2, S3DIS, and SemanticKITTI) show the effectiveness of our semi-supervised method to improve the prediction quality with unlabeled data.

Group activity recognition is a crucial yet challenging problem, whose core lies in fully exploring spatial-temporal interactions among individuals and generating reasonable group representations. However, previous methods either model spatial and temporal information separately, or directly aggregate individual features to form group features. To address these issues, we propose a novel group activity recognition network termed GroupFormer. It captures spatial-temporal contextual information jointly to augment the individual and group representations effectively with a clustered spatial-temporal transformer. Specifically, our GroupFormer has three appealing advantages: (1) A tailor-modified Transformer, Clustered Spatial-Temporal Transformer, is proposed to enhance the individual and group representation. (2) It models the spatial and temporal dependencies integrally and utilizes decoders to build the bridge between the spatial and temporal information. (3) A clustered attention mechanism is utilized to dynamically divide individuals into multiple clusters for better learning activity-aware semantic representations. Moreover, experimental results show that the proposed framework outperforms state-of-the-art methods on the Volleyball dataset and Collective Activity dataset.

The convolutional neural network (CNN) is vulnerable to degraded images with even very small variations (e.g. corrupted and adversarial samples). One of the possible reasons is that CNN pays more attention to the most discriminative regions, but ignores the auxiliary features when learning, leading to the lack of feature diversity for final judgment. In our method, we propose to dynamically suppress significant activation values of CNN by group-wise inhibition, but not fixedly or randomly handle them when training. The feature maps with different activation distribution are then processed separately to take the feature independence into account. CNN is finally guided to learn richer discriminative features hierarchically for robust classification according to the proposed regularization. Our method is comprehensively evaluated under multiple settings, including classification against corruptions, adversarial attacks and low data regime. Extensive experimental results show that the proposed method can achieve significant improvements in terms of both robustness and generalization performances, when compared with the state-of-the-art methods. Code is available at https://github. com/LinusWu/TENET_Training.

Recently, directly detecting 3D objects from 3D point clouds has received increasing attention. To extract object representation from an irregular point cloud, existing methods usually take a point grouping step to assign the points to an object candidate so that a PointNet-like network could be used to derive object features from the grouped points. However, the inaccurate point assignments caused by the hand-crafted grouping scheme decrease the performance of 3D object detection. In this paper, we present a simple yet effective method for directly detecting 3D objects from the 3D point cloud. Instead of grouping local points to each object candidate, our method computes the feature of an object from all the points in the point cloud with the help of an attention mechanism in the Transformers, where the contribution of each point is automatically learned in the network training. With an improved attention stacking scheme, our method fuses object features in different stages and generates more accurate object detection results. With few bells and whistles, the proposed method achieves state-of-the-art 3D object detection performance on two widely used benchmarks, ScanNet V2 and SUN RGB-D.

Assessing action quality is challenging due to the subtle differences between videos and large variations in scores. Most existing approaches tackle this problem by regressing a quality score from a single video, suffering a lot from the large inter-video score variations. In this paper, we show that the relations among videos can provide important clues for more accurate action quality assessment during both training and inference. Specifically, we reformulate the problem of action quality assessment as regressing the relative scores with reference to another video that has shared attributes (e.g. category and difficulty), instead of learning unreferenced scores. Following this formulation, we propose a new contrastive regression (CoRe) framework to learn the relative scores by pair-wise comparison, which highlights the differences between videos and guides the models to learn the key hints for assessment. In order to further exploit the relative information between two videos, we devise a group-aware regression tree to convert the conventional score regression into two easier sub-problems: coarse-to-fine classification and regression in small intervals. To demonstrate the effectiveness of CoRe, we conduct extensive experiments on three mainstream AQA datasets including AQA-7, MTL-AQA, and JIGSAWS. Our approaches outperform previous methods by a large margin and establish new state-of-the-art on all three benchmarks.

Scene Graph Generators (SGGs) are models that, given an image, build a directed graph where each edge represents a predicted subject predicate object triplet. Most SGGs silently exploit datasets' bias on relationships' context, i.e. its subject and object, to improve recall and neglect spatial and visual evidence, e.g. having seen a glut of data for person wearing shirt, they are overconfident that every person is wearing every shirt. Such imprecise predictions are mainly ascribed to the lack of negative examples for most relationships, fact that obstructs models from meaningfully learning predicates, even those which have ample positive examples. We first present an in-depth investigation of the context bias issue to showcase that all examined state-of-the-art SGGs share the above vulnerabilities. In response, we propose a semi-supervised scheme that forces predicted triplets to be grounded consistently back to the image, in a closed-loop manner. The developed spatial common sense can be then distilled to a student SGG and substantially enhance its spatial reasoning ability. This Grounding Consistency Distillation (GCD) approach is model-agnostic and profits from the superfluous unlabeled samples to retain the valuable context information and avert memorization of annotations. Furthermore, we ascertain that current metrics disregard unlabeled samples, rendering themselves incapable of reflecting context bias, then we mine and incorporate during evaluation hard-negatives to reformulate precision as a reliable metric. Extensive experimental comparisons exhibit large quantitative - up to 70% relative precision boost on VG200 dataset - and qualitative improvements to prove the significance of our GCD method and our metrics towards refocusing graph generation as a core aspect of scene understanding. Code available at https://github.com/deeplab-ai/grounding-consistent-vrd.

Many vision tasks use secondary information at inference time---a seed---to assist a computer vision model in solving a problem. For example, an initial bounding box is needed to initialize visual object tracking. To date, all such work makes the assumption that the seed is a good one. However, in practice, from crowdsourcing to noisy automated seeds, this is often not the case. We hence propose the problem of seed rejection---determining whether to reject a seed based on the expected performance degradation when it is provided in place of a gold-standard seed. We provide a formal definition to this problem, and focus on two meaningful subgoals: understanding causes of error and understanding the model's response to noisy seeds conditioned on the primary input. With these goals in mind, we propose a novel training method and evaluation metrics for the seed rejection problem. We then use seeded versions of the viewpoint estimation and fine-grained classification tasks to evaluate these contributions. In these experiments, we show our method can reduce the number of seeds that need to be reviewed for a target performance by over 23% compared to strong baselines.