ICCV 2021 论文列表

We propose a novel framework for image clustering that incorporates joint representation learning and clustering. Our method consists of two heads that share the same backbone network - a "representation learning" head and a "clustering" head. The "representation learning" head captures fine-grained patterns of objects at the instance level which serve as clues for the "clustering" head to extract coarse-grain information that separates objects into clusters. The whole model is trained in an end-to-end manner by minimizing the weighted sum of two sample-oriented contrastive losses applied to the outputs of the two heads. To ensure that the contrastive loss corresponding to the "clustering" head is optimal, we introduce a novel critic function called "log-of-dot-product". Extensive experimental results demonstrate that our method significantly outperforms state-of-the-art single-stage clustering methods across a variety of image datasets, improving over the best baseline by about 5-7% in accuracy on CIFAR10/20, STL10, and ImageNet-Dogs. Further, the "two-stage" variant of our method also achieves better results than baselines on three challenging ImageNet subsets.

Network quantization, which aims to reduce the bit-lengths of the network weights and activations, has emerged for their deployments to resource-limited devices. Although recent studies have successfully discretized a full-precision network, they still incur large quantization errors after training, thus giving rise to a significant performance gap between a full-precision network and its quantized counterpart. In this work, we propose a novel quantization method for neural networks, Cluster-Promoting Quantization (CPQ) that finds the optimal quantization grids while naturally encouraging the underlying full-precision weights to gather around those quantization grids cohesively during training. This property of CPQ is thanks to our two main ingredients that enable differentiable quantization: i) the use of the categorical distribution designed by a specific probabilistic parametrization in the forward pass and ii) our proposed multi-class straight-through estimator (STE) in the backward pass. Since our second component, multi-class STE, is intrinsically biased, we additionally propose a new bit-drop technique, DropBits, that revises the standard dropout regularization to randomly drop bits instead of neurons. As a natural extension of DropBits, we further introduce the way of learning heterogeneous quantization levels to find proper bit-length for each layer by imposing an additional regularization on DropBits. We experimentally validate our method on various benchmark datasets and network architectures, and also support a new hypothesis for quantization: learning heterogeneous quantization levels outperforms the case using the same but fixed quantization levels from scratch.

We present a new versatile building block for deep point cloud processing architectures that is equally suited for diverse tasks. This building block combines the ideas of spatial transformers and multi-view convolutional networks with the efficiency of standard convolutional layers in two and three-dimensional dense grids. The new block operates via multiple parallel heads, whereas each head differentiably rasterizes feature representations of individual points into a low-dimensional space, and then uses dense convolution to propagate information across points. The results of the processing of individual heads are then combined together resulting in the update of point features. Using the new block, we build architectures for both discriminative (point cloud segmentation, point cloud classification) and generative (point cloud inpainting and image-based point cloud reconstruction) tasks. The resulting architectures achieve state-of-the-art performance for these tasks, demonstrating the versatility of the new block for point cloud processing.

Long-Term person re-identification (LT-reID) exposes extreme challenges because of the longer time gaps between two recording footages where a person is likely to change clothing. There are two types of approaches for LT-reID: biometrics-based approach and data adaptation based approach. The former one is to seek clothing irrelevant biometric features. However, seeking high quality biometric feature is the main concern. The latter one adopts fine-tuning strategy by using data with significant clothing change. However, the performance is compromised when it is applied to cases without clothing change. This work argues that these approaches in fact are not aware of clothing status (i.e., change or no-change) of a pedestrian. Instead, they blindly assume all footages of a pedestrian have different clothes. To tackle this issue, a Regularization via Clothing Status Awareness Network (RCSANet) is proposed to regularize descriptions of a pedestrian by embedding the clothing status awareness. Consequently, the description can be enhanced to maintain the best ID discriminative feature while improving its robustness to real-world LT-reID where both clothing-change case and no-clothing-change case exist. Experiments show that RCSANet performs reasonably well on three LT-reID datasets.

This paper introduces Click to Move (C2M), a novel framework for video generation where the user can control the motion of the synthesized video through mouse clicks specifying simple object trajectories of the key objects in the scene. Our model receives as input an initial frame, its corresponding segmentation map and the sparse motion vectors encoding the input provided by the user. It outputs a plausible video sequence starting from the given frame and with a motion that is consistent with user input. Notably, our proposed deep architecture incorporates a Graph Convolution Network (GCN) modelling the movements of all the objects in the scene in a holistic manner and effectively combining the sparse user motion information and image features. Experimental results show that C2M outperforms existing methods on two publicly available datasets, thus demonstrating the effectiveness of our GCN framework at modelling object interactions. The source code is publicly available at https://github.com/PierfrancescoArdino/C2M.

We tackle catastrophic forgetting problem in the context of class-incremental learning for video recognition, which has not been explored actively despite the popularity of continual learning. Our framework addresses this challenging task by introducing time-channel importance maps and exploiting the importance maps for learning the representations of incoming examples via knowledge distillation. We also incorporate a regularization scheme in our objective function, which encourages individual features obtained from different time steps in a video to be uncorrelated and eventually improves accuracy by alleviating catastrophic forgetting. We evaluate the proposed approach on brand-new splits of class-incremental action recognition benchmarks constructed upon the UCF101, HMDB51, and Something-Something V2 datasets, and demonstrate the effectiveness of our algorithm in comparison to the existing continual learning methods that are originally designed for image data.

Action localization networks are often structured as a feature encoder sub-network and a localization sub-network, where the feature encoder learns to transform an input video to features that are useful for the localization sub-network to generate reliable action proposals. While some of the encoded features may be more useful for generating action proposals, prior action localization approaches do not include any attention mechanism that enables the localization sub-network to attend more to the more important features. In this paper, we propose a novel attention mechanism, the Class Semantics-based Attention (CSA), that learns from the temporal distribution of semantics of action classes present in an input video to find the importance scores of the encoded features, which are used to provide attention to the more useful encoded features. We demonstrate on two popular action detection datasets that incorporating our novel attention mechanism provides considerable performance gains on competitive action detection models (e.g., around 6.2% improvement over BMN action detection baseline to obtain 47.5% mAP on the THUMOS-14 dataset), and a new state-of-the-art of 36.25% mAP on the ActivityNet v1.3 dataset. Further, the CSA localization model family which includes BMN-CSA, was part of the second-placed submission at the 2021 ActivityNet action localization challenge. Our attention mechanism outperforms prior self-attention modules such as the squeeze-and-excitation in action detection task. We also observe that our attention mechanism is complementary to such self-attention modules in that performance improvements are seen when both are used together.

We propose an end-to-end trainable framework that processes large-scale visual data tensors by looking at a fraction of their entries only. Our method combines a neural network encoder with a tensor train decomposition to learn a low-rank latent encoding, coupled with cross-approximation (CA) to learn the representation through a subset of the original samples. CA is an adaptive sampling algorithm that is native to tensor decompositions and avoids working with the full high-resolution data explicitly. Instead, it actively selects local representative samples that we fetch out-of-core and on demand. The required number of samples grows only logarithmically with the size of the input. Our implicit representation of the tensor in the network enables processing large grids that could not be otherwise tractable in their uncompressed form. The proposed approach is particularly useful for large-scale multidimensional grid data (e.g., 3D tomography), and for tasks that require context over a large receptive field (e.g., predicting the medical condition of entire organs). The code is available at https://github.com/aelphy/c-pic.

Graph convolutional networks (GCNs) have been widely used and achieved remarkable results in skeleton-based action recognition. In GCNs, graph topology dominates feature aggregation and therefore is the key to extracting representative features. In this work, we propose a novel Channel-wise Topology Refinement Graph Convolution (CTR-GC) to dynamically learn different topologies and effectively aggregate joint features in different channels for skeleton-based action recognition. The proposed CTR-GC models channel-wise topologies through learning a shared topology as a generic prior for all channels and refining it with channel-specific correlations for each channel. Our refinement method introduces few extra parameters and significantly reduces the difficulty of modeling channel-wise topologies. Furthermore, via reformulating graph convolutions into a unified form, we find that CTR-GC relaxes strict constraints of graph convolutions, leading to stronger representation capability. Combining CTR-GC with temporal modeling modules, we develop a powerful graph convolutional network named CTR-GCN which notably outperforms state-of-the-art methods on the NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets.

Knowledge distillation (KD) has been proven a simple and effective tool for training compact dense prediction models. Lightweight student networks are trained by extra supervision transferred from large teacher networks. Most previous KD variants for dense prediction tasks align the activation maps from the student and teacher network in the spatial domain, typically by normalizing the activation values on each spatial location and minimizing point-wise and/or pair-wise discrepancy. Different from the previous methods, here we propose to normalize the activation map of each channel to obtain a soft probability map. By simply minimizing the Kullback--Leibler (KL) divergence between the channel-wise probability map of the two networks, the distillation process pays more attention to the most salient regions of each channel, which are valuable for dense prediction tasks. We conduct experiments on a few dense prediction tasks, including semantic segmentation and object detection. Experiments demonstrate that our proposed method outperforms state-of-the-art distillation methods considerably, and can require less computational cost during training. In particular, we improve the RetinaNet detector (ResNet50backbone) by3.4%in mAP on the COCO dataset and spent (ResNet18 backbone) by5.81%in mIoU on the cityscapes dataset. Code is available at: https://git.io/Distiller.

This paper introduces a powerful channel augmented joint learning strategy for the visible-infrared recognition problem. For data augmentation, most existing methods directly adopt the standard operations designed for single-modality visible images, and thus do not fully consider the imagery properties in visible to infrared matching. Our basic idea is to homogenously generate color-irrelevant images by randomly exchanging the color channels. It can be seamlessly integrated into existing augmentation operations without modifying the network, consistently improving the robustness against color variations. Incorporated with a random erasing strategy, it further greatly enriches the diversity by simulating random occlusions. For cross-modality metric learning, we design an enhanced channel-mixed learning strategy to simultaneously handle the intra- and cross-modality variations with squared difference for stronger discriminability. Besides, a channel-augmented joint learning strategy is further developed to explicitly optimize the outputs of augmented images. Extensive experiments with insightful analysis on two visible-infrared recognition tasks show that the proposed strategies consistently improve the accuracy. Without auxiliary information, it improves the state-of-the-art Rank-1/mAP by 14.59%/13.00% on the large-scale SYSU-MM01 dataset.

For high spatial resolution (HSR) remote sensing images, bitemporal supervised learning always dominates change detection using many pairwise labeled bitemporal images. However, it is very expensive and time-consuming to pairwise label large-scale bitemporal HSR remote sensing images. In this paper, we propose single-temporal supervised learning (STAR) for change detection from a new perspective of exploiting object changes in unpaired images as supervisory signals. STAR enables us to train a high-accuracy change detector only using unpaired labeled images and generalize to real-world bitemporal images. To evaluate the effectiveness of STAR, we design a simple yet effective change detector called ChangeStar, which can reuse any deep semantic segmentation architecture by the ChangeMixin module. The comprehensive experimental results show that ChangeStar outperforms the baseline with a large margin under single-temporal supervision and achieves superior performance under bitemporal supervision. Code is available at https://github.com/Z-Zheng/ChangeStar.

Attention module does not always help deep models learn causal features that are robust in any confounding context, e.g., a foreground object feature is invariant to different backgrounds. This is because the confounders trick the attention to capture spurious correlations that benefit the prediction when the training and testing data are IID (identical & independent distribution); while harm the prediction when the data are OOD (out-of-distribution). The sole fundamental solution to learn causal attention is by causal intervention, which requires additional annotations of the confounders, e.g., a "dog" model is learned within "grass+dog" and "road+dog" respectively, so the "grass" and "road" contexts will no longer confound the "dog" recognition. However, such annotation is not only prohibitively expensive, but also inherently problematic, as the confounders are elusive in nature. In this paper, we propose a causal attention module (CaaM) that self-annotates the confounders in unsupervised fashion. In particular, multiple CaaMs can be stacked and integrated in conventional attention CNN and self-attention Vision Transformer. In OOD settings, deep models with CaaM outperform those without it significantly; even in IID settings, the attention localization is also improved by CaaM, showing a great potential in applications that require robust visual saliency. Codes are available at https://github.com/Wangt-CN/CaaM.

Image matting refers to the estimation of the opacity of foreground objects. It requires correct contours and fine details of foreground objects for the matting results. To better accomplish human image matting tasks, we propose the Cascade Image Matting Network with Deformable Graph Refinement(CasDGR), which can automatically predict precise alpha mattes from single human images without any additional inputs. We adopt a network cascade architecture to perform matting from low-to-high resolution, which corresponds to coarse-to-fine optimization. We also introduce the Deformable Graph Refinement (DGR) module based on graph neural networks (GNNs) to overcome the limitations of convolutional neural networks (CNNs). The DGR module can effectively capture long-range relations and obtain more global and local information to help produce finer alpha mattes. We also reduce the computation complexity of the DGR module by dynamically predicting the neighbors and apply DGR module to higher-resolution features. Experimental results demonstrate the ability of our CasDGR to achieve state-of-the-art performance on synthetic datasets and produce good results on real human images.

Vector graphic documents present visual elements in a resolution free, compact format and are often seen in creative applications. In this work, we attempt to learn a generative model of vector graphic documents. We define vector graphic documents by a multi-modal set of attributes associated to a canvas and a sequence of visual elements such as shapes, images, or texts, and train variational auto-encoders to learn the representation of the documents. We collect a new dataset of design templates from an online service that features complete document structure including occluded elements. In experiments, we show that our model, named CanvasVAE, constitutes a strong baseline for generative modeling of vector graphic documents.

Recent studies show that convolutional neural networks (CNNs) are vulnerable under various settings, including adversarial attacks, common corruptions, and backdoor attacks. Motivated by the findings that human visual system pays more attention to global structure (e.g., shapes) for recognition while CNNs are biased towards local texture features in images, in this work we aim to analyze whether "edge features" could improve the recognition robustness in these scenarios, and if so, to what extent? To answer these questions and systematically evaluate the global structure features, we focus on shape features and pro-pose two edge-enabled pipelines EdgeNetRob and Edge-GANRob, forcing the CNNs to rely more on edge features. Specifically, EdgeNetRob and EdgeGANRob first explicitly extract shape structure features from a given image via an edge detection algorithm. Then EdgeNetRob trains down-stream learning tasks directly on the extracted edge features, while EdgeGANRob reconstructs a new image by re-filling the texture information with a trained generative adversarial network (GANs). To reduce the sensitivity of edge detection algorithms to perturbations, we additionally propose a robust edge detection approach Robust Canny based on vanilla Canny. Based on our evaluation, we find that EdgeNetRob can help boost model robustness under differ-ent attack scenarios at the cost of the clean model accuracy. EdgeGANRob, on the other hand, is able to improve the clean model accuracy compared to EdgeNetRob while preserving robustness. This shows that given such edge features, how to leverage them matters for robustness, and it also depends on data properties. Our systematic studies on edge structure features under different settings will shed light on future robust feature exploration and optimization.

Geometric constraints are shown to enforce scale consistency and remedy the scale ambiguity issue in self-supervised monocular depth estimation. Meanwhile, scale-invariant losses focus on learning relative depth, leading to accurate relative depth prediction. To combine the best of both worlds, we learn scale-consistent self-supervised depth in a scale-invariant manner. Towards this goal, we present a scale-aware geometric (SAG) loss, which enforces scale consistency through point cloud alignment. Compared to prior arts, SAG loss takes relative scale into consideration during relative motion estimation, enabling more precise alignment and explicit supervision for scale inference. In addition, a novel two-stream architecture for depth estimation is designed, which disentangles scale from depth estimation and allows depth to be learned in a scale-invariant manner. The integration of SAG loss and two-stream network enables more consistent scale inference and more accurate relative depth estimation. Our method achieves state-of-the-art performance under both scale-invariant and scale-dependent evaluation settings.

Existing 3D human pose estimation algorithms trained on distortion-free datasets suffer performance drop when applied to new scenarios with a specific camera distortion. In this paper, we propose a simple yet effective model for 3D human pose estimation in video that can quickly adapt to any distortion environment by utilizing MAML, a representative optimization-based meta-learning algorithm. We consider a sequence of 2D keypoints in a particular distortion as a single task of MAML. However, due to the absence of a large-scale dataset in a distorted environment, we propose an efficient method to generate synthetic distorted data from undistorted 2D keypoints. For the evaluation, we assume two practical testing situations depending on whether a motion capture sensor is available or not. In particular, we propose Inference Stage Optimization using bone-length symmetry and consistency. Extensive evaluation shows that our proposed method successfully adapts to various degrees of distortion in the testing phase and outperforms the existing state-of-the-art approaches. The proposed method is useful in practice because it does not require camera calibration and additional computations in a testing set-up.

While neural symbolic methods demonstrate impressive performance in visual question answering on synthetic images, their performance suffers on real images. We identify that the long-tail distribution of visual concepts and unequal importance of reasoning steps in real data are the two key obstacles that limit the models' real-world potentials. To address these challenges, we propose a new paradigm, Calibrating Concepts and Operations (CCO), which enables neural symbolic models to capture underlying data characteristics and to reason with hierarchical importance. Specifically, we introduce an executor with learnable concept embedding magnitudes for handling distribution imbalance, and an operation calibrator for highlighting important operations and suppressing redundant ones. Our experiments show CCO substantially boosts the performance of neural symbolic methods on real images. By evaluating models on the real world dataset GQA, CCO helps the neural symbolic method NSCL outperforms its vanilla counterpart by 9.1% (from 47.0% to 56.1%); this result also largely reduces the performance gap between symbolic and non-symbolic methods. Additionally, we create a perturbed test set for better understanding and analyzing model performance on real images. Code is available at https://lizw14.github.io/project/ccosr.