ICCV 2021 论文列表

Generalization on out-of-distribution (OOD) test data is an essential but underexplored topic in visual question answering. Current state-of-the-art VQA models often exploit the biased correlation between data and labels, which results in a large performance drop when the test and training data have different distributions. Inspired by the fact that humans can recognize novel concepts by composing existed concepts and capsule network's ability of representing part-whole hierarchies, we propose to use capsules to represent parts and introduce "Linguistically Routing" to merge parts with human-prior hierarchies. Specifically, we first fuse visual features with a single question word as atomic parts. Then we introduce the "Linguistically Routing" to reweight the capsule connections between two layers such that: 1) the lower layer capsules can transfer their outputs to the most compatible higher capsules, and 2) two capsules can be merged if their corresponding words are merged in the question parse tree. The routing process maximizes the above unary and binary potentials across multiple layers and finally carves a tree structure inside the capsule network. We evaluate our proposed routing method on the CLEVR compositional generation test, the VQA-CP2 dataset and the VQAv2 dataset. The experimental results show that our proposed method can improve current VQA models on OOD split without losing performance on the in-domain test data.

Forecasting complex vehicle and pedestrian multi-modal distributions requires powerful probabilistic approaches. Normalizing flows (NF) have recently emerged as an attractive tool to model such distributions. However, a key drawback is that independent samples drawn from a flow model often do not adequately capture all the modes in the underlying distribution. We propose Likelihood-Based Diverse Sampling (LDS), a method for improving the quality and the diversity of trajectory samples from a pre-trained flow model. Rather than producing individual samples, LDS produces a set of trajectories in one shot. Given a pre-trained forecasting flow model, we train LDS using gradients from the model, to optimize an objective function that rewards high likelihood for individual trajectories in the predicted set, together with high spatial separation among trajectories. LDS outperforms state-of-art post-hoc neural diverse forecasting methods for various pre-trained flow models as well as conditional variational autoencoder (CVAE) models. Crucially, it can also be used for transductive trajectory forecasting, where the diverse forecasts are trained on-the-fly on unlabeled test examples. LDS is easy to implement, and we show that it offers a simple plug-in improvement over baselines on two challenging benchmarks. Code is at: https://github.com/JasonMa2016/LDS

Multi-person total motion capture is extremely challenging when it comes to handle severe occlusions, different reconstruction granularities from body to face and hands, drastically changing observation scales and fast body movements. To overcome these challenges above, we contribute a lightweight total motion capture system for multi-person interactive scenarios using only sparse multi-view cameras. By contributing a novel hand and face bootstrapping algorithm, our method is capable of efficient localization and accurate association of the hands and faces even on severe occluded occasions. We leverage both pose regression and keypoints detection methods and further propose a unified two-stage parametric fitting method for achieving pixel-aligned accuracy. Moreover, for extremely self-occluded poses and close interactions, a novel feedback mechanism is proposed to propagate the pixel-aligned reconstructions into the next frame for more accurate association. Overall, we propose the first light-weight total capture system and achieves fast, robust and accurate multi-person total motion capture performance. The results and experiments show that our method achieves more accurate results than existing methods under sparse-view setups.

Causally-taken images often suffer from flare artifacts, due to the unintended reflections and scattering of light inside the camera. However, as flares may appear in a variety of shapes, positions, and colors, detecting and removing them entirely from an image is very challenging. Existing methods rely on predefined intensity and geometry priors of flares, and may fail to distinguish the difference between light sources and flare artifacts. We observe that the conditions of the light source in the image play an important role in the resulting flares. In this paper, we present a deep framework with light source aware guidance for single-image flare removal (SIFR). In particular, we first detect the light source regions and the flare regions separately, and then remove the flare artifacts based on the light source aware guidance. By learning the underlying relationships between the two types of regions, our approach can remove different kinds of flares from the image. In addition, instead of using paired training data which are difficult to collect, we propose the first unpaired flare removal dataset and new cycle-consistency constraints to obtain more diverse examples and avoid manual annotations. Extensive experiments demonstrate that our method outperforms the baselines qualitatively and quantitatively. We also show that our model can be applied to flare effect manipulation (e.g., adding or changing image flares).

The application of light field data in salient object detection is becoming increasingly popular in recent years. The difficulty lies in how to effectively fuse the features within the focal stack and how to cooperate them with the feature of the all-focus image. Previous methods usually fuse focal stack features via convolution or ConvLSTM, which are both less effective and ill-posed. In this paper, we model the information fusion within focal stack via graph networks. They introduce powerful context propagation from neighbouring nodes and also avoid ill-posed implementations. On the one hand, we construct local graph connections thus avoiding prohibitive computational costs of traditional graph networks. On the other hand, instead of processing the two kinds of data separately, we build a novel dual graph model to guide the focal stack fusion process using all-focus patterns. To handle the second difficulty, previous methods usually implement one-shot fusion for focal stack and all-focus features, hence lacking a thorough exploration of their supplements. We introduce a reciprocative guidance scheme and enable mutual guidance between these two kinds of information at multiple steps. As such, both kinds of features can be enhanced iteratively, finally benefiting the saliency prediction. Extensive experimental results show that the proposed models are all beneficial and we achieve significantly better results than state-of-the-art methods.

Recent research efforts in lifelong learning propose to grow a mixture of models to adapt to an increasing number of tasks. The proposed methodology shows promising results in overcoming catastrophic forgetting. However, the theory behind these successful models is still not well understood. In this paper, we perform the theoretical analysis for lifelong learning models by deriving the risk bounds based on the discrepancy distance between the probabilistic representation of data generated by the model and that corresponding to the target dataset. Inspired by the theoretical analysis, we introduce a new lifelong learning approach, namely the Lifelong Infinite Mixture (LIMix) model, which can automatically expand its network architectures or choose an appropriate component to adapt its parameters for learning a new task, while preserving its previously learnt information. We propose to incorporate the knowledge by means of Dirichlet processes by using a gating mechanism which computes the dependence between the knowledge learnt previously and stored in each component, and a new set of data. Besides, we train a compact Student model which can accumulate cross-domain representations over time and make quick inferences. The code is available at https://github.com/dtuzi123/Lifelong-infinite-mixture-model.

Semantic segmentation is a challenging task in the absence of densely labelled data. Only relying on class activation maps (CAM) with image-level labels provides deficient segmentation supervision. Prior works thus consider pre-trained models to produce coarse saliency maps to guide the generation of pseudo segmentation labels. However, the commonly used off-line heuristic generation process cannot fully exploit the benefits of these coarse saliency maps. Motivated by the significant inter-task correlation, we propose a novel weakly supervised multi-task framework termed as AuxSegNet, to leverage saliency detection and multi-label image classification as auxiliary tasks to improve the primary task of semantic segmentation using only image-level ground-truth labels. Inspired by their similar structured semantics, we also propose to learn a cross-task global pixel-level affinity map from the saliency and segmentation representations. The learned cross-task affinity can be used to refine saliency predictions and propagate CAM maps to provide improved pseudo labels for both tasks. The mutual boost between pseudo label updating and cross-task affinity learning enables iterative improvements on segmentation performance. Extensive experiments demonstrate the effectiveness of the proposed auxiliary learning network structure and the cross-task affinity learning method. The proposed approach achieves state-of-the-art weakly supervised segmentation performance on the challenging PASCAL VOC 2012 and MS COCO benchmarks.

Contaminants such as dust, dirt and moisture adhering to the camera lens can greatly affect the quality and clarity of the resulting image or video. In this paper, we propose a video restoration method to automatically remove these contaminants and produce a clean video. Our approach first seeks to detect attention maps that indicate the regions that need to be restored. In order to leverage the corresponding clean pixels from adjacent frames, we propose a flow completion module to hallucinate the flow of the background scene to the attention regions degraded by the contaminants. Guided by the attention maps and completed flows, we propose a recurrent technique to restore the input frame by fetching clean pixels from adjacent frames. Finally, a multi-frame processing stage is used to further process the entire video sequence in order to enforce temporal consistency. The entire network is trained on a synthetic dataset that approximates the physical lighting properties of contaminant artifacts. This new dataset and our novel framework lead to our method that is able to address different contaminants and outperforms competitive restoration approaches both qualitatively and quantitatively.

Significant effort has been recently devoted to modeling visual relations. This has mostly addressed the design of architectures, typically by adding parameters and increasing model complexity. However, visual relation learning is a long-tailed problem, due to the combinatorial nature of joint reasoning about groups of objects. Increasing model complexity is, in general, ill-suited for long-tailed problems due to their tendency to overfit. In this paper, we explore an alternative hypothesis, denoted the Devil is in the Tails. Under this hypothesis, better performance is achieved by keeping the model simple but improving its ability to cope with long-tailed distributions. To test this hypothesis, we devise a new approach for training visual relationships models, which is inspired by state-of-the-art long-tailed recognition literature. This is based on an iterative decoupled training scheme, denoted Decoupled Training for Devil in the Tails (DT2). DT2 employs a novel sampling approach, Alternating Class-Balanced Sampling (ACBS), to capture the interplay between the long-tailed entity and predicate distributions of visual relations. Results show that, with an extremely simple architecture, DT2-ACBS significantly outperforms much more complex state-of-the-art methods on scene graph generation tasks. This suggests that the development of sophisticated models must be considered in tandem with the long-tailed nature of the problem.

We address the problem of visible-infrared person re-identification (VI-reID), that is, retrieving a set of person images, captured by visible or infrared cameras, in a cross-modal setting. Two main challenges in VI-reID are intra-class variations across person images, and cross-modal discrepancies between visible and infrared images. Assuming that the person images are roughly aligned, previous approaches attempt to learn coarse image- or rigid part-level person representations that are discriminative and generalizable across different modalities. However, the person images, typically cropped by off-the-shelf object detectors, are not necessarily well-aligned, which distract discriminative person representation learning. In this paper, we introduce a novel feature learning framework that addresses these problems in a unified way. To this end, we propose to exploit dense correspondences between cross-modal person images. This allows to address the cross-modal discrepancies in a pixel-level, suppressing modality-related features from person representations more effectively. This also encourages pixel-wise associations between cross-modal local features, further facilitating discriminative feature learning for VI-reID. Extensive experiments and analyses on standard VI-reID benchmarks demonstrate the effectiveness of our approach, which significantly outperforms the state of the art.

In video understanding, most cross-modal knowledge distillation (KD) methods are tailored for classification tasks, focusing on the discriminative representation of the trimmed videos. However, action detection requires not only categorizing actions, but also localizing them in untrimmed videos. Therefore, transferring knowledge pertaining to temporal relations is critical for this task which is missing in the previous cross-modal KD frameworks. To this end, we aim at learning an augmented RGB representation for action detection, taking advantage of additional modalities at training time through KD. We propose a KD framework consisting of two levels of distillation. On one hand, atomic-level distillation encourages the RGB student to learn the sub-representation of the actions from the teacher in a contrastive manner. On the other hand, sequence-level distillation encourages the student to learn the temporal knowledge from the teacher, which consists of transferring the Global Contextual Relations and the action Boundary Saliency. The result is an Augmented-RGB stream that can achieve competitive performance as the two-stream network while using only RGB at inference time. Extensive experimental analysis shows that our proposed distillation framework is generic and outperforms other popular cross-modal distillation methods in the action detection task.

This paper studies the task of inpainting man-made scenes. It is very challenging due to the difficulty in preserving the visual patterns of images, such as edges, lines, and junctions. Especially, most previous works are failed to restore the object/building structures for images of man-made scenes. To this end, this paper proposes learning a Sketch Tensor (ST) space for inpainting man-made scenes. Such a space is learned to restore the edges, lines, and junctions in images, and thus makes reliable predictions of the holistic image structures. To facilitate the structure refinement, we propose a Multi-scale Sketch Tensor inpainting (MST) network, with a novel encoder-decoder structure. The encoder extracts lines and edges from the input images to project them into an ST space. From this space, the decoder is learned to restore the input images. Extensive experiments validate the efficacy of our model. Furthermore, our model can also achieve competitive performance in inpainting general nature images over the competitors.

Recently, the performance of single image super-resolution (SR) has been significantly improved with powerful networks. However, these networks are developed for image SR with specific integer scale factors (e.g., x2/3/4), and cannot handle non-integer and asymmetric SR. In this paper, we propose to learn a scale-arbitrary image SR network from scale-specific networks. Specifically, we develop a plug-in module for existing SR networks to perform scale-arbitrary SR, which consists of multiple scale-aware feature adaption blocks and a scale-aware upsampling layer. Moreover, conditional convolution is used in our plug-in module to generate dynamic scale-aware filters, which enables our network to adapt to arbitrary scale factors. Our plug-in module can be easily adapted to existing networks to realize scale-arbitrary SR with a single model. These networks plugged with our module can produce promising results for non-integer and asymmetric SR while maintaining state-of-the-art performance for SR with integer scale factors. Besides, the additional computational and memory cost of our module is very small.

Multi-objective multi-task learning aims to boost the performance of all tasks by leveraging their correlation and conflict appropriately. Nevertheless, in real practice, users may have preference for certain tasks, and other tasks simply serve as privileged or auxiliary tasks to assist the training of target tasks. The privileged tasks thus possess less or even no priority in the final task assessment by users. Motivated by this, we propose a privileged multiple descent algorithm to arbitrate the learning of target tasks and privileged tasks. Concretely, we introduce a privileged parameter so that the optimization direction does not necessarily follow the gradient from the privileged tasks, but concentrates more on the target tasks. Besides, we also encourage a priority parameter for the target tasks to control the potential distraction of optimization direction from the privileged tasks. In this way, the optimization direction can be more aggressively determined by weighting the gradients among target and privileged tasks, and thus highlight more the performance of target tasks under the unified multi-task learning context. Extensive experiments on synthetic and real-world datasets indicate that our method can achieve versatile Pareto solutions under varying preference for the target tasks.

Learning with noisy labels is an important and challenging task for training accurate deep neural networks. However, some commonly-used loss functions, such as Cross Entropy (CE), always suffer from severe overfitting to noisy labels. Although robust loss functions have been designed, they often encounter underfitting. In this paper, we theoretically prove that any loss will be robust to noisy labels when restricting the output of a network to the set of permutations over any fixed vector. When the fixed vector is one-hot, we only need to constrain the output to be one-hot, which means a discrete image and thus zero gradients almost everywhere. This prohibits gradient-based learning of models. In this work, we introduce two sparse regularization strategies to approximate the one-hot constraint: output sharpening and l_p-norm (p\le 1). Output sharpening directly modifies the output distribution of a network to be sharp by adjusting the "temperature" parameter. l_p-norm plays the role of a regularization term to make the output to be sparse. These two simple strategies guarantee the robustness of arbitrary loss functions while not hindering the fitting ability of networks. Experiments on baseline and real-world datasets demonstrate that the sparse regularization can significantly improve the performance of commonly-used loss functions in the presence of noisy labels, and outperform state-of-the-art methods.

Point cloud segmentation is a fundamental task in 3D. Despite recent progress on point cloud segmentation with the power of deep networks, current deep learning methods based on the clean label assumptions may fail with noisy labels. Yet, object class labels are often mislabeled in real-world point cloud datasets. In this work, we take the lead in solving this issue by proposing a novel Point Noise-Adaptive Learning (PNAL) framework. Compared to existing noise-robust methods on image tasks, our PNAL is noise-rate blind, to cope with the spatially variant noise rate problem specific to point clouds. Specifically, we propose a novel point-wise confidence selection to obtain reliable labels based on the historical predictions of each point. A novel cluster-wise label correction is proposed with a voting strategy to generate the best possible label taking the neighbor point correlations into consideration. We conduct extensive experiments to demonstrate the effectiveness of PNAL on both synthetic and real-world noisy datasets. In particular, even with 60% symmetric noisy labels, our proposed method produces much better results than its baseline counterpart without PNAL and is comparable to the ideal upper bound trained on a completely clean dataset. Moreover, we fully re-labeled the validation set of a popular but noisy real-world scene dataset ScanNetV2 to make it clean, for rigorous experiment and future research. Our code and data will be released.

The core of deep metric learning (DML) involves learning visual similarities in high-dimensional embedding space. One of the main challenges is to generalize from seen classes of training data to unseen classes of test data. Recent works have focused on exploiting past embeddings to increase the number of instances for the seen classes. Such methods achieve performance improvement via augmentation, while the strong focus on seen classes still remains. This can be undesirable for DML, where training and test data exhibit entirely different classes. In this work, we present a novel training strategy for DML called MemVir. Unlike previous works, MemVir memorizes both embedding features and class weights to utilize them as additional virtual classes. The exploitation of virtual classes not only utilizes augmented information for training but also alleviates a strong focus on seen classes for better generalization. Moreover, we embed the idea of curriculum learning by slowly adding virtual classes for a gradual increase in learning difficulty, which improves the learning stability as well as the final performance. MemVir can be easily applied to many existing loss functions without any modification. Extensive experimental results on famous benchmarks demonstrate the superiority of MemVir over state-of-the-art competitors. Code of MemVir is publicly available.

Anomaly detection (AD) aims to address the task of classification or localization of image anomalies. This paper addresses two pivotal issues of reconstruction-based approaches to AD in images, namely, model adaptation and reconstruction gap. The former generalizes an AD model to tackling a broad range of object categories, while the latter provides useful clues for localizing abnormal regions. At the core of our method is an unsupervised universal model, termed as Metaformer, which leverages both meta-learned model parameters to achieve high model adaptation capability and instance-aware attention to emphasize the focal regions for localizing abnormal regions, i.e., to explore the reconstruction gap at those regions of interest. We justify the effectiveness of our method with SOTA results on the MVTec AD dataset of industrial images and highlight the adaptation flexibility of the universal Metaformer with multi-class and few-shot scenarios.

Tracking by detection, the dominant approach for online multi-object tracking, alternates between localization and association steps. As a result, it strongly depends on the quality of instantaneous observations, often failing when objects are not fully visible. In contrast, tracking in humans is underlined by the notion of object permanence: once an object is recognized, we are aware of its physical existence and can approximately localize it even under full occlusions. In this work, we introduce an end-to-end trainable approach for joint object detection and tracking that is capable of such reasoning. We build on top of the recent CenterTrack architecture, which takes pairs of frames as input, and extend it to videos of arbitrary length. To this end, we augment the model with a spatio-temporal, recurrent memory module, allowing it to reason about object locations and identities in the current frame using all the previous history. It is, however, not obvious how to train such an approach. We study this question on a new, large-scale, synthetic dataset for multi-object tracking, which provides ground truth annotations for invisible objects, and propose several approaches for supervising tracking behind occlusions. Our model, trained jointly on synthetic and real data, outperforms the state of the art on KITTI and MOT17 datasets thanks to its robustness to occlusions.