ICCV 2021 Papers

Deep CNN-based methods have so far achieved the state of the art results in multi-view 3D object reconstruction. Despite the considerable progress, the two core modules of these methods - view feature extraction and multi-view fusion, are usually investigated separately, and the relations among multiple input views are rarely explored. Inspired by the recent great success in Transformer models, we reformulate the multi-view 3D reconstruction as a sequence-to-sequence prediction problem and propose a framework named 3D Volume Transformer. Unlike previous CNN-based methods using a separate design, we unify the feature extraction and view fusion in a single Transformer network. A natural advantage of our design lies in the exploration of view-to-view relationships using self-attention among multiple unordered inputs. On ShapeNet - a large-scale 3D reconstruction benchmark, our method achieves a new state-of-the-art accuracy in multi-view reconstruction with fewer parameters (70% less) than CNN-based methods. Experimental results also suggest the strong scaling capability of our method. Our code will be made publicly available.

Multi-view clustering, a long-standing and important research problem, focuses on mining complementary information from diverse views. However, existing works often fuse multiple views' representations or handle clustering in a common feature space, which may result in their entanglement especially for visual representations. To address this issue, we present a novel VAE-based multi-view clustering framework (Multi-VAE) by learning disentangled visual representations. Concretely, we define a view-common variable and multiple view-peculiar variables in the generative model. The prior of view-common variable obeys approximately discrete Gumbel Softmax distribution, which is introduced to extract the common cluster factor of multiple views. Meanwhile, the prior of view-peculiar variable follows continuous Gaussian distribution, which is used to represent each view's peculiar visual factors. By controlling the mutual information capacity to disentangle the view-common and view-peculiar representations, continuous visual information of multiple views can be separated so that their common discrete cluster information can be effectively mined. Experimental results demonstrate that Multi-VAE enjoys the disentangled and explainable visual representations, while obtaining superior clustering performance compared with state-of-the-art methods.

Despite the fast progress in training specialized models for various tasks, learning a single general model that works well for many tasks is still challenging for computer vision. Here we introduce multi-task self-training (MuST), which harnesses the knowledge in independent specialized teacher models (e.g., ImageNet model on classification) to train a single general student model. Our approach has three steps. First, we train specialized teachers independently on labeled datasets. We then use the specialized teachers to label an unlabeled dataset to create a multi-task pseudo labeled dataset. Finally, the dataset, which now contains pseudo labels from teacher models trained on different datasets/tasks, is then used to train a student model with multi-task learning. We evaluate the feature representations of the student model on 6 vision tasks including image recognition (classification, detection, segmentation) and 3D geometry estimation (depth and surface normal estimation). MuST is scalable with unlabeled or partially labeled datasets and outperforms both specialized supervised models and self-supervised models when training on large scale datasets. Lastly, we show MuST can improve upon already strong checkpoints trained with billions of examples. The results suggest self-training is a promising direction to aggregate labeled and unlabeled training data for learning general feature representations.

In this work, we address the task of unsupervised domain adaptation (UDA) for semantic segmentation in presence of multiple target domains: the objective is to train a single model that can handle all these domains at test time. Such a multi-target adaptation is crucial for a variety of scenarios that real-world autonomous systems must handle. It is a challenging set-up since one faces not only the domain gap between the labeled source set and the unlabeled target set, but also the distribution shifts existing within the latter among the different target domains. To this end, we introduce two adversarial frameworks: (i) multi-discriminator, which explicitly aligns each target domain to its counterparts, and (ii) multi-target knowledge transfer, which learns a target-agnostic model thanks to a multi-teacher/single-student distillation mechanism. The evaluation is done on four newly proposed multi-target benchmarks for UDA in semantic segmentation. In all tested scenarios, our approaches consistently outperform baselines, setting competitive standards for the novel task.

To reduce annotation labor associated with object detection, an increasing number of studies focus on transferring the learned knowledge from a labeled source domain to another unlabeled target domain. However, existing methods assume that the labeled data are sampled from a single source domain, which ignores a more generalized scenario, where labeled data are from multiple source domains. For the more challenging task, we propose a unified Faster RCNN based framework, termed Divide-and-Merge Spindle Network (DMSN), which can simultaneously enhance domain invariance and preserve discriminative power. Specifically, the framework contains multiple source subnets and a pseudo target subnet. First, we propose a hierarchical feature alignment strategy to conduct strong and weak alignments for low- and high-level features, respectively, considering their different effects for object detection. Second, we develop a novel pseudo subnet learning algorithm to approximate optimal parameters of pseudo target subset by weighted combination of parameters in different source subnets. Finally, a consistency regularization for region proposal network is proposed to facilitate each subnet to learn more abstract invariances. Extensive experiments on different adaptation scenarios demonstrate the effectiveness of the proposed model.

This paper presents a new Vision Transformer (ViT) architecture Multi-Scale Vision Longformer, which significantly enhances the ViT of [??] for encoding high-resolution images using two techniques. The first is the multi-scale model structure, which provides image encodings at multiple scales with manageable computational cost. The second is the attention mechanism of Vision Longformer, which is a variant of Longformer [??], originally developed for natural language processing, and achieves a linear complexity w.r.t. the number of input tokens. A comprehensive empirical study shows that the new ViT significantly outperforms several strong baselines, including the existing ViT models and their ResNet counterparts, and the Pyramid Vision Transformer from a concurrent work [??], on a range of vision tasks, including image classification, object detection, and segmentation. The models and source code are released at https://github.com/microsoft/vision-longformer.

Although recent research has witnessed a significant progress on the video deblurring task, these methods struggle to reconcile inference efficiency and visual quality simultaneously, especially on ultra-high-definition (UHD) videos (e.g., 4K resolution). To address the problem, we propose a novel deep model for fast and accurate UHD Video Deblurring (UHDVD). The proposed UHDVD is achieved by a separable-patch architecture, which collaborates with a multi-scale integration scheme to achieve a large receptive field without adding the number of generic convolutional layers and kernels. Additionally, we design a residual channel-spatial attention (RCSA) module to improve accuracy and reduce the depth of the network appropriately. The proposed UHDVD is the first real-time deblurring model for 4K videos at 35 fps. To train the proposed model, we build a new dataset comprised of 4K blurry videos and corresponding sharp frames using three different smartphones. Comprehensive experimental results show that our network performs favorably against the state-ofthe-art methods on both the 4K dataset and public benchmarks in terms of accuracy, speed, and model size.

Deep features have been proven powerful in building accurate dense semantic correspondences in various previous works. However, the multi-scale and pyramidal hierarchy of convolutional neural networks has not been well studied to learn discriminative pixel-level features for semantic correspondence. In this paper, we propose a multiscale matching network that is sensitive to tiny semantic differences between neighboring pixels. We follow the coarse-to-fine matching strategy, and build a top-down feature and matching enhancement scheme that is coupled with the multi-scale hierarchy of deep convolutional neural networks. During feature enhancement, intra-scale enhancement fuses same-resolution feature maps from multiple layers together via local self-attention, and cross-scale enhancement hallucinates higher resolution feature maps along the top-down hierarchy. Besides, we learn complementary matching details at different scales, and thus the overall matching score is refined by features at different semantic levels gradually. Our multi-scale matching network can be trained end-to-end easily with few additional learnable parameters. Experimental results demonstrate the proposed method achieves state-of-the-art performance on three popular benchmarks with high computational efficiency.

In this paper, we introduce a novel audio-visual multi-modal bridging framework that can utilize both audio and visual information, even with uni-modal inputs. We exploit a memory network that stores source (i.e., visual) and target (i.e., audio) modal representations, where source modal representation is what we are given, and target modal representations are what we want to obtain from the memory network. We then construct an associative bridge between source and target memories that considers the interrelationship between the two memories. By learning the interrelationship through the associative bridge, the proposed bridging framework is able to obtain the target modal representations inside the memory network, even with the source modal input only, and it provides rich information for its downstream tasks. We apply the proposed framework to two tasks: lip reading and speech reconstruction from silent video. Through the proposed associative bridge and modality-specific memories, each task knowledge is enriched with the recalled audio context, achieving state-of-the-art performance. We also verify that the associative bridge properly relates the source and target memories.

Multi-action video recognition is much more challenging due to the requirement to recognize multiple actions co-occurring simultaneously or sequentially. Modeling multi-action relations is beneficial and crucial to understand videos with multiple actions, and actions in a video are usually presented in multiple modalities. In this paper, we propose a novel multi-action relation model for videos, by leveraging both relational graph convolutional networks (GCNs) and video multi-modality. We first build multi-modal GCNs to explore modality-aware multi-action relations, fed by modality-specific action representation as node features, e.g., spatiotemporal features learned by 3D convolutional neural network (CNN), audio and textual embeddings queried from respective feature lexicons. We then joint both multi-modal CNN-GCN models and multi-modal feature representations for learning better relational action predictions. Ablation study, multi-action relation visualization, and boosts analysis, all show efficacy of our multi-modal multi-action relation modeling. Also our method achieves state-of-the-art performance on large-scale multi-action M-MiT benchmark. Our code is made publicly available at https://github.com/zhenglab/multi-action-video.

Barrel distortion rectification aims at removing the radial distortion in a distorted image captured by a wide-angle lens. Previous deep learning methods mainly solve this problem by learning the implicit distortion parameters or the nonlinear rectified mapping function in a direct manner. However, this type of manner results in an indistinct learning process of rectification and thus limits the deep perception of distortion. In this paper, inspired by the curriculum learning, we analyze the barrel distortion rectification task in a progressive and meaningful manner. By considering the relationship among different construction levels in an image, we design a multi-level curriculum that disassembles the rectification task into three levels, structure recovery, semantics embedding, and texture rendering. With the guidance of the curriculum that corresponds to the construction of images, the proposed hierarchical architecture enables a progressive rectification and achieves more accurate results. Moreover, we present a novel distortion-aware pre-training strategy to facilitate the initial learning of neural networks, promoting the model to converge faster and better. Experimental results on the synthesized and real-world distorted image datasets show that the proposed approach significantly outperforms other learning methods, both qualitatively and quantitatively.

A key assumption of top-down human pose estimation approaches is their expectation of having a single person/instance present in the input bounding box. This often leads to failures in crowded scenes with occlusions. We propose a novel solution to overcome the limitations of this fundamental assumption. Our Multi-Instance Pose Network (MIPNet) allows for predicting multiple 2D pose instances within a given bounding box. We introduce a Multi-Instance Modulation Block (MIMB) that can adaptively modulate channel-wise feature responses for each instance and is parameter efficient. We demonstrate the efficacy of our approach by evaluating on COCO, CrowdPose, and OCHuman datasets. Specifically, we achieve 70.0 AP on CrowdPose and 42.5 AP on OCHuman test sets, a significant improvement of 2.4 AP and 6.5 AP over the prior art, respectively. When using ground truth bounding boxes for inference, MIPNet achieves an improvement of 0.7 AP on COCO, 0.9 AP on CrowdPose, and 9.1 AP on OCHuman validation sets compared to HRNet. Interestingly, when fewer, high confidence bounding boxes are used, HRNet's performance degrades (by 5 AP) on OCHuman, whereas MIPNet maintains a relatively stable performance (drop of 1 AP) for the same inputs.

The success of deep neural networks (DNNs) has promoted the widespread applications of person re-identification (ReID). However, ReID systems inherit the vulnerability of DNNs to malicious attacks of visually inconspicuous adversarial perturbations. Detection of adversarial attacks is, therefore, a fundamental requirement for robust ReID systems. In this work, we propose a Multi-Expert Adversarial Attack Detection (MEAAD) approach to achieve this goal by checking context inconsistency, which is suitable for any DNNs-based ReID systems. Specifically, three kinds of context inconsistencies caused by adversarial attacks are employed to learn a detector for detecting adversarial attacks, i.e., a) the embedding distances between a perturbed query person image and its top-K retrievals are generally larger than those between a benign query image and its top-K retrievals, b) the embedding distances among the top-K retrievals of a perturbed query image are larger than those of a benign query image, c) the top-K retrievals of a benign query image obtained with multiple expert ReID models tend to be consistent, which is not preserved when attacks are present. Extensive experiments on the Market1501 and DukeMTMC-ReID datasets show that, as the first adversarial attack detection approach for ReID, MEAAD effectively detects various adversarial attacks and achieves high ROC-AUC (over 97.5%).

LiDAR sensors can be used to obtain a wide range of measurement signals other than a simple 3D point cloud, and those signals can be leveraged to improve perception tasks like 3D object detection. A single laser pulse can be partially reflected by multiple objects along its path, resulting in multiple measurements called echoes. Multi-echo measurement can provide information about object contours and semi-transparent surfaces which can be used to better identify and locate objects. LiDAR can also measure surface reflectance (intensity of laser pulse return), as well as ambient light of the scene (sunlight reflected by objects). These signals are already available in commercial LiDAR devices but have not been used in most LiDAR-based detection models. We present a 3D object detection model which leverages the full spectrum of measurement signals provided by LiDAR. First, we propose a multi-signal fusion (MSF) module to combine (1) the reflectance and ambient features extracted with a 2D CNN, and (2) point cloud features extracted using a 3D graph neural network (GNN). Second, we propose a multi-echo aggregation (MEA) module to combine the information encoded in different set of echo points. Compared with traditional single echo point cloud methods, our proposed multi-signal LiDAR Detector (MSLiD) extracts richer context information from a wider range of sensing measurements and achieves more accurate 3D object detection. Experiments show that by incorporating the multi-modality of LiDAR, our method outperforms the state-of-the-art by up to relatively 9.1%.

Image generation has rapidly evolved in recent years. Modern architectures for adversarial training allow to generate even high resolution images with remarkable quality. At the same time, more and more effort is dedicated towards controlling the content of generated images. In this paper, we take one further step in this direction and propose a conditional generative adversarial network (GAN) that generates images with a defined number of objects from given classes. This entails two fundamental abilities (1) being able to generate high-quality images given a complex constraint and (2) being able to count object instances per class in a given image. Our proposed model modularly extends the successful StyleGAN2 architecture with a count-based conditioning as well as with a regression sub-network to count the number of generated objects per class during training. In experiments on three different datasets, we show that the proposed model learns to generate images according to the given multiple-class count condition even in the presence of complex backgrounds. In particular, we propose a new dataset, CityCount, which is derived from the Cityscapes street scenes dataset, to evaluate our approach in a challenging and practically relevant scenario. An implementation is available at https://github.com/boschresearch/MCCGAN.

In digital pathology, both detection and classification of cells are important for automatic diagnostic and prognostic tasks. Classifying cells into subtypes, such as tumor cells, lymphocytes or stromal cells is particularly challenging. Existing methods focus on morphological appearance of individual cells, whereas in practice pathologists often infer cell classes through their spatial context. In this paper, we propose a novel method for both detection and classification that explicitly incorporates spatial contextual information. We use the spatial statistical function to describe local density in both a multi-class and a multi-scale manner. Through representation learning and deep clustering techniques, we learn advanced cell representation with both appearance and spatial context. On various benchmarks, our method achieves better performance than state-of-the-arts, especially on the classification task.

Unsupervised domain adaption has proven to be an effective approach for alleviating the intensive workload of manual annotation by aligning the synthetic source-domain data and the real-world target-domain samples. Unfortunately, mapping the target-domain distribution to the source-domain unconditionally may distort the essential structural information of the target-domain data. To this end, we firstly propose to introduce a novel multi-anchor based active learning strategy to assist domain adaptation regarding the semantic segmentation task. By innovatively adopting multiple anchors instead of a single centroid, the source domain can be better characterized as a multimodal distribution, thus more representative and complimentary samples are selected from the target domain. With little workload to manually annotate these active samples, the distortion of the target-domain distribution can be effectively alleviated, resulting in a large performance gain. The multi-anchor strategy is additionally employed to model the target-distribution. By regularizing the latent representation of the unlabeled target samples compact around multiple anchors through a novel soft alignment loss, more precise segmentation can be achieved. Extensive experiments are conducted on public datasets to demonstrate that the proposed approach outperforms state-of-the-art methods significantly, along with thorough ablation study to verify the effectiveness of each component.

We introduce the active audio-visual source separation problem, where an agent must move intelligently in order to better isolate the sounds coming from an object of interest in its environment. The agent hears multiple audio sources simultaneously (e.g., a person speaking down the hall in a noisy household) and it must use its eyes and ears to automatically separate out the sounds originating from a target object within a limited time budget. Towards this goal, we introduce a reinforcement learning approach that trains movement policies controlling the agent's camera and microphone placement over time, guided by the improvement in predicted audio separation quality. We demonstrate our approach in scenarios motivated by both augmented reality (system is already co-located with the target object) and mobile robotics (agent begins arbitrarily far from the target object). Using state-of-the-art realistic audio-visual simulations in 3D environments, we demonstrate our model's ability to find minimal movement sequences with maximal payoff for audio source separation. Project: http://vision.cs.utexas.edu/projects/move2hear

Motion, as the most distinct phenomenon in a video to involve the changes over time, has been unique and critical to the development of video representation learning. In this paper, we ask the question: how important is the motion particularly for self-supervised video representation learning. To this end, we compose a duet of exploiting the motion for data augmentation and feature learning in the regime of contrastive learning. Specifically, we present a Motion-focused Contrastive Learning (MCL) method that regards such duet as the foundation. On one hand, MCL capitalizes on optical flow of each frame in a video to temporally and spatially sample the tubelets (i.e., sequences of associated frame patches across time) as data augmentations. On the other hand, MCL further aligns gradient maps of the convolutional layers to optical flow maps from spatial, temporal and spatio-temporal perspectives, in order to ground motion information in feature learning. Extensive experiments conducted on R(2+1)D backbone demonstrate the effectiveness of our MCL. On UCF101, the linear classifier trained on the representations learnt by MCL achieves 81.91% top-1 accuracy, outperforming ImageNet supervised pre-training by 6.78%. On Kinetics-400, MCL achieves 66.62% top-1 accuracy under the linear protocol.