CVPR 2021 论文列表

The detection of 3D objects from LiDAR data is a critical component in most autonomous driving systems. Safe, high speed driving needs larger detection ranges, which are enabled by new LiDARs. These larger detection ranges require more efficient and accurate detection models. Towards this goal, we propose Range Sparse Net (RSN) - a simple, efficient, and accurate 3D object detector - in order to tackle real time 3D object detection in this extended detection regime. RSN predicts foreground points from range images and applies sparse convolutions on the selected fore-ground points to detect objects. The lightweight 2D convolutions on dense range images results in significantly fewer selected foreground points, thus enabling the later sparse convolutions in RSN to efficiently operate. Combining features from the range image further enhance detection ac-curacy. RSN runs at more than 60 frames per second on a 150mx150m detection region on Waymo Open Dataset (WOD) while being more accurate than previously published detectors. RSN is ranked first in the WOD leaderboard based on the APH/LEVEL1 metrics for LiDAR-based pedestrian and vehicle detection, while being several times faster than alternatives.

Imbalanced datasets widely exist in practice and are a great challenge for training deep neural models with a good generalization on infrequent classes. In this work, we propose a new rare-class sample generator (RSG) to solve this problem. RSG aims to generate some new samples for rare classes during training, and it has in particular the following advantages: (1) it is convenient to use and highly versatile, because it can be easily integrated into any kind of convolutional neural network, and it works well when combined with different loss functions, and (2) it is only used during the training phase, and therefore, no additional burden is imposed on deep neural networks during the testing phase. In extensive experimental evaluations, we verify the effectiveness of RSG. Furthermore, by leveraging RSG, we obtain competitive results on Imbalanced CIFAR and new state-of-the-art results on Places-LT, ImageNet-LT, and iNaturalist 2018. The source code is available at https://github.com/Jianf-Wang/RSG.

We propose RPSRNet - a novel end-to-end trainable deep neural network for rigid point set registration. For this task, we use a novel 2^D-tree representation for the input point sets and a hierarchical deep feature embedding in the neural network. An iterative transformation refinement module in our network boosts the feature matching accuracy in the intermediate stages. We achieve an inference speed of 12-15ms to register a pair of input point clouds as large as 250K. Extensive evaluation on (i) KITTI LiDAR odometry and (ii) ModelNet-40 datasets shows that our method outperforms prior state-of-the-art methods -- e.g., on the KITTI data set, DCP-v2 by1.3 and 1.5 times, and PointNetLK by 1.8 and 1.9 times better rotational and translational accuracy respectively. Evaluation on ModelNet40 shows that RPSRNet is more robust than other benchmark methods when the samples contain a significant amount of noise and other disturbances. RPSRNet accurately registers point clouds with non-uniform sampling densities, e.g., LiDAR data, which cannot be processed by many existing deep-learning-based registration methods.

Recent years have witnessed great progress in object detection. However, due to the domain shift problem, applying the knowledge of an object detector learned from one specific domain to another one often suffers severe performance degradation. Most existing methods adopt feature alignment either on the backbone network or instance classifier to increase the transferability of object detector. Different from existing methods, we propose to perform feature alignment of foreground and background in the RPN stage such that the foreground and background RPN proposals in target domain can be effectively separated. Specifically, we first construct one set of learnable RPN prototypes, and then enforce the RPN features to align with the prototypes for both source and target domains. It essentially cooperates the learning of RPN prototypes and features to align the source and target RPN features. In this paradigm, the pseudo label of proposals in target domain need be first generated, and we propose a simple yet effective method suitable for RPN feature alignment,i.e., using the filtered detection results to guide the pseudo label generation of RPN proposals by IoU. Furthermore, we adopt Grad CAM to find the discriminative region within a proposal and use it to increase the discriminability of RPN features for alignment by spatially weighting. We conduct extensive experiments on multiple cross-domain detection scenarios. The results show the effectiveness of our proposed method against previous state-of-the-art methods.

Majority of the perception methods in robotics require depth information provided by RGB-D cameras. However, standard 3D sensors fail to capture depth of transparent objects due to refraction and absorption of light. In this paper, we introduce a new approach for depth completion of transparent objects from a single RGB-D image. Key to our approach is a local implicit neural representation built on ray-voxel pairs that allows our method to generalize to unseen objects and achieve fast inference speed. Based on this representation, we present a novel framework that can complete missing depth given noisy RGB-D input. We further improve the depth estimation iteratively using a self-correcting refinement model. To train the whole pipeline, we build a large scale synthetic dataset with transparent objects. Experiments demonstrate that our method performs significantly better than the current state-of-the-art methods on both synthetic and real world data. In addition, our approach improves the inference speed by a factor of 20 compared to the previous best method, ClearGrasp. Code will be released at https://research.nvidia.com/publication/2021-03_RGB-D-Local-Implicit.

We address the problem of scene flow: given a pair of stereo or RGB-D video frames, estimate pixelwise 3D motion. We introduce RAFT-3D, a new deep architecture for scene flow. RAFT-3D is based on the RAFT model developed for optical flow but iteratively updates a dense field of pixelwise SE3 motion instead of 2D motion. A key innovation of RAFT-3D is rigid-motion embeddings, which represent a soft grouping of pixels into rigid objects. Integral to rigid-motion embeddings is Dense-SE3, a differentiable layer that enforces geometric consistency of the embeddings. Experiments show that RAFT-3D achieves state-of-the-art performance. On FlyingThings3D, under the two-view evaluation, we improved the best published accuracy (delta < 0.05) from 34.3% to 83.7%. On KITTI, we achieve an error of 5.77, outperforming the best published method (6.31), despite using no object instance supervision.

Similarity learning has been recognized as a crucial step for object tracking. However, existing multiple object tracking methods only use sparse ground truth matching as the training objective, while ignoring the majority of the informative regions on the images. In this paper, we present Quasi-Dense Similarity Learning, which densely samples hundreds of region proposals on a pair of images for contrastive learning. We can directly combine this similarity learning with existing detection methods to build Quasi-Dense Tracking (QDTrack) without turning to displacement regression or motion priors. We also find that the resulting distinctive feature space admits a simple nearest neighbor search at the inference time. Despite its simplicity, QDTrack outperforms all existing methods on MOT, BDD100K, Waymo, and TAO tracking benchmarks. It achieves 68.7 MOTA at 20.3 FPS on MOT17 without using external training data. Compared to methods with similar detectors, it boosts almost 10 points of MOTA and significantly decreases the number of ID switches on BDD100K and Waymo datasets. Our code and trained models are available at https://github.com/SysCV/qdtrack.

We present QuantumSync, the first quantum algorithm for solving a synchronization problem in the context of computer vision. In particular, we focus on permutation synchronization which involves solving a non-convex optimization problem in discrete variables. We start by formulating synchronization into a quadratic unconstrained binary optimization problem (QUBO). While such formulation respects the binary nature of the problem, ensuring that the result is a set of permutations requires extra care. Hence, we: (i) show how to insert permutation constraints into a QUBO problem and (ii) solve the constrained QUBO problem on the current generation of the adiabatic quantum computers D-Wave. Thanks to the quantum annealing, we guarantee global optimality with high probability while sampling the energy landscape to yield confidence estimates. Our proof-of-concepts realization on the adiabatic D-Wave computer demonstrates that quantum machines offer a promising way to solve the prevalent yet difficult synchronization problems.

Explainability of deep learning methods is imperative to facilitate their clinical adoption in digital pathology. However, popular deep learning methods and explainability techniques (explainers) based on pixel-wise processing disregard biological entities' notion, thus complicating comprehension by pathologists. In this work, we address this by adopting biological entity-based graph processing and graph explainers enabling explanations accessible to pathologists. In this context, a major challenge becomes to discern meaningful explainers, particularly in a standardized and quantifiable fashion. To this end, we propose herein a set of novel quantitative metrics based on statistics of class separability using pathologically measurable concepts to characterize graph explainers. We employ the proposed metrics to evaluate three types of graph explainers, namely the layer-wise relevance propagation, gradient-based saliency, and graph pruning approaches, to explain Cell-Graph representations for Breast Cancer Subtyping. The proposed metrics are also applicable in other domains by using domain-specific intuitive concepts. We validate the qualitative and quantitative findings on the BRACS dataset, a large cohort of breast cancer RoIs, by expert pathologists. The code and models will be released upon acceptance.

Despite the remarkable performance of deep models on image recognition tasks, they are known to be susceptible to common corruptions such as blur, noise, and low-resolution. Data augmentation is a conventional way to build a robust model by considering these common corruptions during the training. However, a naive data augmentation scheme may result in a non-specialized model for particular corruptions, as the model tends to learn the averaged distribution among corruptions. To mitigate the issue, we propose a new paradigm of training deep image recognition networks that produce clean-like features from any quality image via an invertible neural architecture. The proposed method consists of two stages. In the first stage, we train an invertible network with only clean images under the recognition objective. In the second stage, its inversion, i.e., the invertible decoder, is attached to a new recognition network and we train this encoder-decoder network using both clean and corrupted images by considering recognition and reconstruction objectives. Our two-stage scheme allows the network to produce clean-like and robust features from any quality images, by reconstructing their clean images via the invertible decoder. We demonstrate the effectiveness of our method on image classification and face recognition tasks.

Modern deep neural networks (DNNs) cannot be effectively used in mobile and embedded devices due to strict requirements for computational complexity, memory, and power consumption. The quantization of weights and feature maps (activations) is a popular approach to solve this problem. Training-aware quantization often shows excellent results but requires a full dataset, which is not always available. Post-training quantization methods, in turn, are applied without fine-tuning but still work well for many classes of tasks like classification, segmentation, and so on. However, they either imply a big overhead for quantization parameters (QPs) calculation at runtime (dynamic methods) or lead to an accuracy drop if pre-computed static QPs are used (static methods). Moreover, most inference frameworks don't support dynamic quantization. Thus we propose a novel quantization approach called QPP: quantization parameter prediction. With a small subset of a training dataset or unlabeled data from the same domain, we find the predictor that can accurately estimate QPs of activations given only the NN's input data. Such a predictor allows us to avoid complex calculation of precise values of QPs while maintaining the quality of the model. To illustrate our method's efficiency, we added QPP into two dynamic approaches: 1) Dense+Sparse quantization, where the predetermined percentage of activations are not quantized, 2) standard quantization with equal quantization steps. We provide experiments on a wide set of tasks including super-resolution, facial landmark, segmentation, and classification.

We propose a simple, intuitive yet powerful method for human-object interaction (HOI) detection. HOIs are so diverse in spatial distribution in an image that existing CNN-based methods face the following three major drawbacks; they cannot leverage image-wide features due to CNN's locality, they rely on a manually defined location-of-interest for the feature aggregation, which sometimes does not cover contextually important regions, and they cannot help but mix up the features for multiple HOI instances if they are located closely. To overcome these drawbacks, we propose a transformer-based feature extractor, in which an attention mechanism and query-based detection play key roles. The attention mechanism is effective in aggregating contextually important information image-wide, while the queries, which we design in such a way that each query captures at most one human-object pair, can avoid mixing up the features from multiple instances. This transformer-based feature extractor produces so effective embeddings that the subsequent detection heads may be fairly simple and intuitive. The extensive analysis reveals that the proposed method successfully extracts contextually important features, and thus outperforms existing methods by large margins (5.37 mAP on HICO-DET, and 5.6 mAP on V-COCO). The source codes are available at https://github.com/hitachi-rd-cv/qpic.

We study the query-based attack against image retrieval to evaluate its robustness against adversarial examples under the black-box setting, where the adversary only has query access to the top-k ranked unlabeled images from the database. Compared with query attacks in image classification, which produce adversaries according to the returned labels or confidence score, the challenge becomes even more prominent due to the difficulty in quantifying the attack effectiveness on the partial retrieved list. In this paper, we make the first attempt in Query-based Attack against Image Retrieval (QAIR), to completely subvert the top-k retrieval results. Specifically, a new relevance-based loss is designed to quantify the attack effects by measuring the set similarity on the top-k retrieval results before and after attacks and guide the gradient optimization. To further boost the attack efficiency, a recursive model stealing method is proposed to acquire transferable priors on the target model and generate the prior-guided gradients. Comprehensive experiments show that the proposed attack achieves a high attack success rate with few queries against the image retrieval systems under the black-box setting. The attack evaluations on real-world visual search engine show that it successfully deceives a commercial system such as Bing Visual Search with 98% attack success rate by only 33 queries on average.

We have observed significant progress in visual navigation for embodied agents. A common assumption in studying visual navigation is that the environments are static; this is a limiting assumption. Intelligent navigation may involve interacting with the environment beyond just moving forward/backward and turning left/right. Sometimes, the best way to navigate is to push something out of the way. In this paper, we study the problem of interactive navigation where agents learn to change the environment to navigate more efficiently to their goals. To this end, we introduce the Neural Interaction Engine (NIE) to explicitly predict the change in the environment caused by the agent's actions. By modeling the changes while planning, we find that agents exhibit significant improvements in their navigational capabilities. More specifically, we consider two downstream tasks in the physics-enabled, visually rich, AI2-THOR environment: (1) reaching a target while the path to the target is blocked (2) moving an object to a target location by pushing it. For both tasks, agents equipped with an NIE significantly outperform agents without the understanding of the effect of the actions indicating the benefits of our approach. The code and dataset are available at github.com/KuoHaoZeng/Interactive_Visual_Navigation.

We propose Pulsar, an efficient sphere-based differentiable rendering module that is orders of magnitude faster than competing techniques, modular, and easy-to-use due to its tight integration with PyTorch. Differentiable rendering is the foundation for modern neural rendering approaches, since it enables end-to-end training of 3D scene representations from image observations. However, gradient-based optimization of neural mesh, voxel, or function representations suffers from multiple challenges, i.e., topological in-consistencies, high memory footprints, or slow rendering speeds. To alleviate these problems, Pulsar employs: 1) asphere-based scene representation, 2) a modular, efficient differentiable projection operation, and 3) (optional) neural shading. Pulsar executes orders of magnitude faster than existing techniques and allows real-time rendering and optimization of representations with millions of spheres. Using spheres for the scene representation, unprecedented speed is obtained while avoiding topology problems. Pulsar is fully differentiable and thus enables a plethora of applications, ranging from 3D reconstruction to neural rendering.

Face recognition has achieved a great success in recent years, it is still challenging to recognize those facial images with extreme poses. Traditional methods consider it as a domain gap problem. Many of them settle it by generating fake frontal faces from extreme ones, whereas they are tough to maintain the identity information with high computational consumption and uncontrolled disturbances. Our experimental analysis shows a dramatic precision drop with extreme poses. Meanwhile, those extreme poses just exist minor visual differences after small rotations. Derived from this insight, we attempt to relieve such a huge precision drop by making minor changes to the input images without modifying existing discriminators. A novel lightweight pseudo facial generation is proposed to relieve the problem of extreme poses without generating any frontal facial image. It can depict the facial contour information and make appropriate modifications to preserve the critical identity information. Specifically, the proposed method reconstructs pseudo profile faces by minimizing the pixel-wise differences with original profile faces and maintaining the identity consistent information from their corresponding frontal faces simultaneously. The proposed framework can improve existing discriminators and obtain a great promotion on several benchmark datasets.

The extraction of auto-correlation in images has shown great potential in deep learning networks, such as the self-attention mechanism in the channel domain and the self-similarity mechanism in the spatial domain. However, the realization of the above mechanisms mostly requires complicated module stacking and a large number of convolution calculations, which inevitably increases model complexity and memory cost. Therefore, we propose a pseudo 3D auto-correlation network (P3AN) to explore a more efficient way of capturing contextual information in image denoising. On the one hand, P3AN uses fast 1D convolution instead of dense connections to realize criss-cross interaction, which requires less computational resources. On the other hand, the operation does not change the feature size and makes it easy to expand. It means that only a simple adaptive fusion is needed to obtain contextual information that includes both the channel domain and the spatial domain. Our method built a pseudo 3D auto-correlation attention block through 1D convolutions and a lightweight 2D structure for more discriminative features. Extensive experiments have been conducted on three synthetic and four real noisy datasets. According to quantitative metrics and visual quality evaluation, the P3AN shows great superiority and surpasses state-of-the-art image denoising methods.

Self-training is a competitive approach in domain adaptive segmentation, which trains the network with the pseudo labels on the target domain. However inevitably, the pseudo labels are noisy and the target features are dispersed due to the discrepancy between source and target domains. In this paper, we rely on representative prototypes, the feature centroids of classes, to address the two issues for unsupervised domain adaptation. In particular, we take one step further and exploit the feature distances from prototypes that provide richer information than mere prototypes. Specifically, we use it to estimate the likelihood of pseudo labels to facilitate online correction in the course of training. Meanwhile, we align the prototypical assignments based on relative feature distances for two different views of the same target, producing a more compact target feature space. Moreover, we find that distilling the already learned knowledge to a self-supervised pretrained model further boosts the performance. Our method shows tremendous performance advantage over state-of-the-art methods.

Unsupervised Domain Adaptation (UDA) transfers predictive models from a fully-labeled source domain to an unlabeled target domain. In some applications, however, it is expensive even to collect labels in the source domain, making most previous works impractical. To cope with this problem, recent work performed instance-wise cross-domain self-supervised learning, followed by an additional fine-tuning stage. However, the instance-wise self-supervised learning only learns and aligns low-level discriminative features. In this paper, we propose an end-to-end Prototypical Cross-domain Self-Supervised Learning (PCS) framework for Few-shot Unsupervised Domain Adaptation (FUDA). PCS not only performs cross-domain low-level feature alignment, but it also encodes and aligns semantic structures in the shared embedding space across domains. Our framework captures category-wise semantic structures of the data by in-domain prototypical contrastive learning; and performs feature alignment through cross-domain prototypical self-supervision. Compared with state-of-the-art methods, PCS improves the mean classification accuracy over different domain pairs on FUDA by 10.5%, 3.5%, 9.0%, and 13.2% on Office, Office-Home, VisDA-2017, and DomainNet, respectively.