ICCV 2021 论文列表

Unsupervised domain adaptation (UDA) involves a supervised loss in a labeled source domain and an unsupervised loss in an unlabeled target domain, which often faces more severe overfitting (than classical supervised learning) as the supervised source loss has clear domain gap and the unsupervised target loss is often noisy due to the lack of annotations. This paper presents RDA, a robust domain adaptation technique that introduces adversarial attacking to mitigate overfitting in UDA. We achieve robust domain adaptation by a novel Fourier adversarial attacking (FAA) method that allows large magnitude of perturbation noises but has minimal modification of image semantics, the former is critical to the effectiveness of its generated adversarial samples due to the existence of domain gaps. Specifically, FAA decomposes images into multiple frequency components (FCs) and generates adversarial samples by just perturbating certain FCs that capture little semantic information. With FAA-generated samples, the training can continue the random walk and drift into an area with a flat loss landscape, leading to more robust domain adaptation. Extensive experiments over multiple domain adaptation tasks show that RDA can work with different computer vision tasks with superior performance.

Predictor-based algorithms have achieved remarkable performance in the Neural Architecture Search (NAS) tasks. However, these methods suffer from high computation costs, as training the performance predictor usually requires training and evaluating hundreds of architectures from scratch. Previous works along this line mainly focus on reducing the number of architectures required to fit the predictor. In this work, we tackle this challenge from a different perspective - improve search efficiency by cutting down the computation budget of architecture training. We propose NOn-uniform Successive Halving (NOSH), a hierarchical scheduling algorithm that terminates the training of underperforming architectures early to avoid wasting budget. To effectively leverage the non-uniform supervision signals produced by NOSH, we formulate predictor-based architecture search as learning to rank with pairwise comparisons. The resulting method - RANK-NOSH, reduces the search budget by 5x while achieving competitive or even better performance than previous state-of-the-art predictor-based methods on various spaces and datasets.

Motion forecasting plays a significant role in various domains (e.g., autonomous driving, human-robot interaction), which aims to predict future motion sequences given a set of historical observations. However, the observed elements may be of different levels of importance. Some information may be irrelevant or even distracting to the forecasting in certain situations. To address this issue, we propose a generic motion forecasting framework (named RAIN) with dynamic key information selection and ranking based on a hybrid attention mechanism. The general framework is instantiated to handle multi-agent trajectory prediction and human motion forecasting tasks, respectively. In the former task, the model learns to recognize the relations between agents with a graph representation and to determine their relative significance. In the latter task, the model learns to capture the temporal proximity and dependency in long-term human motions. We also propose an effective double-stage training pipeline with an alternating training strategy to optimize the parameters in different modules of the framework. We validate the framework on both synthetic simulations and motion forecasting benchmarks in different domains, demonstrating that our method not only achieves state-of-the-art forecasting performance but also provides interpretable and reasonable hybrid attention weights.

We present a method for optimization-based recovery of eye motion from rolling shutter video of the retina. Our approach formulates eye tracking as an optimization problem that jointly estimates the retina's motion and appearance using convex optimization and a constrained version of gradient descent. By incorporating the rolling shutter imaging model into the formulation of our joint optimization, we achieve state-of-the-art accuracy both offline and in real-time. We apply our method to retina video captured with an adaptive optics scanning laser ophthalmoscope (AOSLO), demonstrating eye tracking at 1 kHz with accuracies below one arcminute -- over an order of magnitude higher than conventional eye tracking systems.

In this paper, we propose Recurrent Multi-Scale Feature Modulation (R-MSFM), a new deep network architecture for self-supervised monocular depth estimation. R-MSFM extracts per-pixel features, builds a multi-scale feature modulation module, and iteratively updates an inverse depth through a parameter-shared decoder at the fixed resolution. This architecture enables our R-MSFM to maintain semantically richer while spatially more precise representations and avoid the error propagation caused by the traditional U-Net-like coarse-to-fine architecture widely used in this domain, resulting in strong generalization and efficient parameter count. Experimental results demonstrate the superiority of our proposed R-MSFM both at model size and inference speed, and show the state-of-the-art results on the KITTI benchmark. Code is available at https://github.com/jsczzzk/R-MSFM

Few-shot object detection (FSOD) aims to detect never-seen objects using few examples. This field sees recent improvement owing to the meta-learning techniques by learning how to match between the query image and few-shot class examples, such that the learned model can generalize to few-shot novel classes. However, currently, most of the meta-learning-based methods perform parwise matching between query image regions (usually proposals) and novel classes separately, therefore failing to take into account multiple relationships among them. In this paper, we propose a novel FSOD model using heterogeneous graph convolutional networks. Through efficient message passing among all the proposal and class nodes with three different types of edges, we could obtain context-aware proposal features and query-adaptive, multiclass-enhanced prototype representations for each class, which could help promote the pairwise matching and improve final FSOD accuracy. Extensive experimental results show that our proposed model, denoted as QA-FewDet, outperforms the current state-of-the-art approaches on the PASCAL VOC and MSCOCO FSOD benchmarks under different shots and evaluation metrics.

Finding shape correspondences can be formulated as an NP-hard quadratic assignment problem (QAP) that becomes infeasible for shapes with high sampling density. A promising research direction is to tackle such quadratic optimization problems over binary variables with quantum annealing, which allows for some problems a more efficient search in the solution space. Unfortunately, enforcing the linear equality constraints in QAPs via a penalty significantly limits the success probability of such methods on currently available quantum hardware. To address this limitation, this paper proposes Q-Match, i.e., a new iterative quantum method for QAPs inspired by the alpha-expansion algorithm, which allows solving problems of an order of magnitude larger than current quantum methods. It implicitly enforces the QAP constraints by updating the current estimates in a cyclic fashion. Further, Q-Match can be applied iteratively, on a subset of well-chosen correspondences, allowing us to scale to real-world problems. Using the latest quantum annealer, the D-Wave Advantage, we evaluate the proposed method on a subset of QAPLIB as well as on isometric shape matching problems from the FAUST dataset.

Although convolutional neural networks (CNNs) have achieved great success in computer vision, this work investigates a simpler, convolution-free backbone network useful for many dense prediction tasks. Unlike the recently-proposed Vision Transformer (ViT) that was designed for image classification specifically, we introduce the Pyramid Vision Transformer (PVT), which overcomes the difficulties of porting Transformer to various dense prediction tasks. PVT has several merits compared to current state of the arts. (1) Different from ViT that typically yields low-resolution outputs and incurs high computational and memory costs, PVT not only can be trained on dense partitions of an image to achieve high output resolution, which is important for dense prediction, but also uses a progressive shrinking pyramid to reduce the computations of large feature maps. (2) PVT inherits the advantages of both CNN and Transformer, making it a unified backbone for various vision tasks without convolutions, where it can be used as a direct replacement for CNN backbones. (3) We validate PVT through extensive experiments, showing that it boosts the performance of many downstream tasks, including object detection, instance and semantic segmentation. For example, with a comparable number of parameters, PVT+RetinaNet achieves 40.4 AP on the COCO dataset, surpassing ResNet50+RetinNet (36.3 AP) by 4.1 absolute AP. We hope that PVT could serve as an alternative and useful backbone for pixel-level predictions and facilitate future research.

Video-based person re-identification aims to associate the video clips of the same person across multiple non-overlapping cameras. Spatial-temporal representations can provide richer and complementary information between frames, which are crucial to distinguish the target person when occlusion occurs. This paper proposes a novel Pyramid Spatial-Temporal Aggregation (PSTA) framework to aggregate the frame-level features progressively and fuse the hierarchical temporal features into a final video-level representation. Thus, short-term and long-term temporal information could be well exploited by different hierarchies. Furthermore, a Spatial-Temporal Aggregation Module (STAM) is proposed to enhance the aggregation capability of PSTA. It mainly consists of two novel attention blocks: Spatial Reference Attention (SRA) and Temporal Reference Attention (TRA). SRA explores the spatial correlations within a frame to determine the attention weight of each location. While TRA extends SRA with the correlations between adjacent frames, temporal consistency information can be fully explored to suppress the interference features and strengthen the discriminative ones. Extensive experiments on several challenging benchmarks demonstrate the effectiveness of the proposed PSTA, and our full model reaches 91.5% and 98.3% Rank-1 accuracy on MARS and DukeMTMC-VID benchmarks.

We present a flexible and high-performance framework, named Pyramid R-CNN, for two-stage 3D object detection from point clouds. Current approaches generally rely on the points or voxels of interest for RoI feature extraction on the second stage, but cannot effectively handle the sparsity and non-uniform distribution of those points, and this may result in failures in detecting objects that are far away. To resolve the problems, we propose a novel second-stage module, named pyramid RoI head, to adaptively learn the features from the sparse points of interest. The pyramid RoI head consists of three key components. Firstly, we propose the RoI-grid Pyramid, which addresses the sparsity problem by extensively collecting points of interest for each RoI in a pyramid manner. Secondly, we propose RoI-grid Attention, a new operation that can encode richer information from sparse points by incorporating conventional attention-based and graph-based point operators into a unified formulation. Thirdly, we propose the Density-Aware Radius Prediction (DARP) module, which can adapt to different point density levels by dynamically adjusting the focusing range of RoIs. Combining the three components, our pyramid RoI head is robust to the sparse and imbalanced circumstances, and can be applied upon various 3D backbones to consistently boost the detection performance. Extensive experiments show that Pyramid R-CNN outperforms the state-of-the-art 3D detection models by a large margin on both the KITTI dataset and the Waymo Open dataset.

Recently, deep learning based point cloud descriptors have achieved impressive results in the place recognition task. Nonetheless, due to the sparsity of point clouds, how to extract discriminative local features of point clouds to efficiently form a global descriptor is still a challenging problem. In this paper, we propose a pyramid point cloud transformer network (PPT-Net) to learn the discriminative global descriptors from point clouds for efficient retrieval. Specifically, we first develop a pyramid point transformer module that adaptively learns the spatial relationship of the different local k-NN graphs of point clouds, where the grouped self-attention is proposed to extract discriminative local features of the point clouds. Furthermore, the grouped self-attention not only enhances long-term dependencies of the point clouds, but also reduces the computational cost. In order to obtain discriminative global descriptors, we construct a pyramid VLAD module to aggregate the multi-scale feature maps of point clouds into the global descriptors. By applying VLAD pooling on multi-scale feature maps, we utilize the context gating mechanism on the multiple global descriptors to adaptively weight the multi-scale global context information into the final global descriptor. Experimental results on the Oxford dataset and three in-house datasets show that our method achieves the state-of-the-art on the point cloud based place recognition task. Code is available at https://github.com/fpthink/PPT-Net.

Multi-scale and multi-patch deep models have been shown effective in removing blurs of dynamic scenes. However, these methods still have one major obstacle: manually designing a lightweight and high-efficiency network is challenging and time-consuming. To tackle this problem, we propose a novel deblurring method, dubbed PyNAS (pyramid neural architecture search network), towards automatically designing hyper-parameters including the scales, patches, and standard cell operators. The proposed PyNAS adopts gradient-based search strategies and innovatively searches the hierarchy patch and scale scheme not limited to the cell searching. Specifically, we introduce a hierarchical search strategy tailored for the multi-scale and multi-patch deblurring task. The strategy follows the principle that the first distinguishes between the top-level (pyramid-scales and pyramid-patches) and bottom-level variables (cell operators) and then searches multi-scale variables using the top-to-bottom principle. During the search stage, PyNAS employs an early stopping strategy to avoid the collapse and computational issue. Furthermore, we use a path-level binarization mechanism for multi-scale cell searching to save memory consumption. Our model is a real-time deblurring algorithm (around 58 fps) for 720p images while achieves state-of-the-art deblurring performance on the GoPro and Video Deblurring dataset.

Regression-based methods have recently shown promising results in reconstructing human meshes from monocular images. By directly mapping raw pixels to model parameters, these methods can produce parametric models in a feed-forward manner via neural networks. However, minor deviation in parameters may lead to noticeable misalignment between the estimated meshes and image evidences. To address this issue, we propose a Pyramidal Mesh Alignment Feedback (PyMAF) loop to leverage a feature pyramid and rectify the predicted parameters explicitly based on the mesh-image alignment status in our deep regressor. In PyMAF, given the currently predicted parameters, mesh-aligned evidences will be extracted from finer-resolution features accordingly and fed back for parameter rectification. To reduce noise and enhance the reliability of these evidences, an auxiliary pixel-wise supervision is imposed on the feature encoder, which provides mesh-image correspondence guidance for our network to preserve the most related information in spatial features. The efficacy of our approach is validated on several benchmarks, including Human3.6M, 3DPW, LSP, and COCO, where experimental results show that our approach consistently improves the mesh-image alignment of the reconstruction. The project page with code and video results can be found at https://hongwenzhang.github.io/pymaf.

We present DietNeRF, a 3D neural scene representation estimated from a few images. Neural Radiance Fields (NeRF) learn a continuous volumetric representation of a scene through multi-view consistency, and can be rendered from novel viewpoints by ray casting. While NeRF has an impressive ability to reconstruct geometry and fine details given many images, up to 100 for challenging 360 degree scenes, it often finds a degenerate solution to its image reconstruction objective when only a few input views are available. To improve few-shot quality, we propose DietNeRF. We introduce an auxiliary semantic consistency loss that encourages realistic renderings at novel poses. DietNeRF is trained on individual scenes to (1) correctly render given input views from the same pose, and (2) match high-level semantic attributes across different, random poses. Our semantic loss allows us to supervise DietNeRF from arbitrary poses. We extract these semantics using a pre-trained visual encoder such as CLIP, a Vision Transformer trained on hundreds of millions of diverse single-view, 2D photographs mined from the web with natural language supervision. In experiments, DietNeRF improves the perceptual quality of few-shot view synthesis when learned from scratch, can render novel views with as few as one observed image when pre-trained on a multi-view dataset, and produces plausible completions of completely unobserved regions. Our project website is available at https://www.ajayj.com/dietnerf.

Most weakly supervised semantic segmentation (WSSS) methods follow the pipeline that generates pseudo-masks initially and trains the segmentation model with the pseudo-masks in fully supervised manner after. However, we find some matters related to the pseudo-masks, including high quality pseudo-masks generation from class activation maps (CAMs), and training with noisy pseudo-mask supervision. For these matters, we propose the following designs to push the performance to new state-of-art: (i) Coefficient of Variation Smoothing to smooth the CAMs adaptively; (ii) Proportional Pseudo-mask Generation to project the expanded CAMs to pseudo-mask based on a new metric indicating the importance of each class on each location, instead of the scores trained from binary classifiers. (iii) Pretended Under-Fitting strategy to suppress the influence of noise in pseudo-mask; (iv) Cyclic Pseudo-mask to boost the pseudo-masks during training of fully supervised semantic segmentation (FSSS). Experiments based on our methods achieve new state-of-art results on two changeling weakly supervised semantic segmentation datasets, pushing the mIoU to 70.0% and 40.2% on PAS-CAL VOC 2012 and MS COCO 2014 respectively. Codes including segmentation framework are released at https://github.com/Eli-YiLi/PMM

Semi-supervised few-shot learning is developed to train a classifier that can adapt to new tasks with limited labeled data and a fixed quantity of unlabeled data. Most semi-supervised few-shot learning methods select pseudo-labeled data of unlabeled set by task-specific confidence estimation. This work presents a task-unified confidence estimation approach for semi-supervised few-shot learning, named pseudo-loss confidence metric (PLCM). It measures the data credibility by the loss distribution of pseudo-labels, which is synthetical considered multi-tasks. Specifically, pseudo-labeled data of different tasks are mapped to a unified metric space by mean of the pseudo-loss model, making it possible to learn the prior pseudo-loss distribution. Then, confidence of pseudo-labeled data is estimated according to the distribution component confidence of its pseudo-loss. Thus highly reliable pseudo-labeled data are selected to strengthen the classifier. Moreover, to overcome the pseudo-loss distribution shift and improve the effectiveness of classifier, we advance the multi-step training strategy coordinated with the class balance measures of class-apart selection and class weight. Experimental results on four popular benchmark datasets demonstrate that the proposed approach can effectively select pseudo-labeled data and achieve the state-of-the-art performance.

The goal of the alignment problem is to align a (given) point cloud P = \ p_1,\cdots,p_n\ to another (observed) point cloud Q = \ q_1,\cdots,q_n\ . That is, to compute a rotation matrix R \in \mathbb R ^ 3 x3 and a translation vector t \in \mathbb R ^ 3 that minimize the sum of paired distances between every transformed point Rp_i-t, to its corresponding point q_i, over every i\in \br 1,\cdots,n . A harder version is the registration problem, where the correspondence is unknown, and the minimum is also over all possible correspondence functions from P to Q. Algorithms such as the Iterative Closest Point (ICP) and its variants were suggested for these problems, but none yield a provable non-trivial approximation for the global optimum. We prove that there always exists a "witness" set of 3 pairs in P xQ that, via novel alignment algorithm, defines a constant factor approximation (in the worst case) to this global optimum. We then provide algorithms that recover this witness set and yield the first provable constant factor approximation for the: (i) alignment problem in O(n) expected time, and (ii) registration problem in polynomial time. Such small witness sets exist for many variants including points in d-dimensional space, outlier-resistant cost functions, and different correspondence types. Extensive experimental results on real and synthetic datasets show that, in practice, our approximation constants are close to 1 and our error is up to x10 times smaller than state-of-the-art algorithms.

The deep learning methods in addressing semantic segmentation typically demand vast amount of pixel-wise annotated training samples. In this work, we present zero-shot semantic segmentation, which aims to identify not only the seen classes contained in training but also the novel classes that have never been seen. We adopt a stringent inductive setting in which only the instances of seen classes are accessible during training. We propose an open-aware prototypical matching approach to accomplish the segmentation. The prototypical way extracts the visual representations by a set of prototypes, making it convenient and flexible to add new unseen classes. A prototype projection is trained to map the semantic representations towards prototypes based on seen instances, and will generate prototypes for unseen classes. Moreover, an open-set rejection is utilized to detect the objects that do not belong to any seen classes, which greatly reduces the misclassifications of unseen objects as seen classes caused by the lack of unseen training instances. We apply the framework on two segmentation datasets, Pascal VOC 2012 and Pascal Context, and achieve impressively state-of-the-art performance.

With the development of 3D scanning technologies, 3D vision tasks have become a popular research area. Owing to the large amount of data acquired by sensors, unsupervised learning is essential for understanding and utilizing point clouds without an expensive annotation process. In this paper, we propose a novel framework and an effective auto-encoder architecture named "PSG-Net" for reconstruction-based learning of point clouds. Unlike existing studies that used fixed or random 2D points, our framework generates input-dependent point-wise features for the latent point set. PSG-Net uses the encoded input to produce point-wise features through the seed generation module and extracts richer features in multiple stages with gradually increasing resolution by applying the seed feature propagation module progressively. We prove the effectiveness of PSG-Net experimentally; PSG-Net shows state-of-the-art performances in point cloud reconstruction and unsupervised classification, and achieves comparable performance to counterpart methods in supervised completion.