ICCV 2021 论文列表

The point cloud representation of an object can have a large geometric variation in view of inconsistent data acquisition procedure, which thus leads to domain discrepancy due to diverse and uncontrollable shape representation cross datasets. To improve discrimination on unseen distribution of point-based geometries in a practical and feasible perspective, this paper proposes a new method of geometry-aware self-training (GAST) for unsupervised domain adaptation of object point cloud classification. Specifically, this paper aims to learn a domain-shared representation of semantic categories, via two novel self-supervised geometric learning tasks as feature regularization. On one hand, the representation learning is empowered by a linear mixup of point cloud samples with their self-generated rotation labels, to capture a global topological configuration of local geometries. On the other hand, a diverse point distribution across datasets can be normalized with a novel curvature-aware distortion localization. Experiments on the PointDA-10 dataset show that our GAST method can significantly outperform the state-of-the-art methods.

Monocular 3D object detection has received increasing attention due to the wide application in autonomous driving. Existing works mainly focus on introducing geometry projection to predict depth priors for each object. Despite their impressive progress, these methods neglect the geometry leverage effect of the projection process, which leads to uncontrollable inferences and damage the training efficiency. In this paper, we propose a Geometry Uncertainty Projection Network (GUP Net) to handle these problems, which can guide the model to learn more reliable depth outputs. The overall framework combines the uncertainty inference and the hierarchical task learning to reduce the negative effects of the geometry leverage. Specifically, an Uncertainty Geometry Projection module is proposed to obtain the geometry guided uncertainty of the inferred depth, which can not only benefit the geometry learning but also provide more reliable depth inferences to reduce the uncontrollableness caused by the geometry leverage. Besides, to reduce the instability in the training process caused by the geometry leverage effect, we propose a Hierarchical Task Learning strategy to control the overall optimization process. This learning algorithm can monitor the situation of each task through a well designed learning situation indicator and adaptively assign the proper loss weights for different tasks according to their learning situation and the hierarchical structure, which can significantly improve the stability and the efficiency of the training process. Extensive experiments demonstrate the effectiveness of the proposed method.The overall model can infer more reliable depth and location information than existing methods, which achieves the state-of-the-art performance on the KITTI benchmark.

Simulators can efficiently generate large amounts of labeled synthetic data with perfect supervision for hard-to-label tasks like semantic segmentation. However, they introduce a domain gap that severely hurts real-world performance. We propose to use self-supervised monocular depth estimation as a proxy task to bridge this gap and improve sim-to-real unsupervised domain adaptation (UDA). Our Geometric Unsupervised Domain Adaptation method (GUDA) learns a domain-invariant representation via a multi-task objective combining synthetic semantic supervision with real-world geometric constraints on videos. GUDA establishes a new state of the art in UDA for semantic segmentation on three benchmarks, outperforming methods that use domain adversarial learning, self-training, or other self-supervised proxy tasks. Furthermore, we show that our method scales well with the quality and quantity of synthetic data while also improving depth prediction.

Pixel-to-mesh has wide applications, especially in virtual or augmented reality, animation and game industry. However, existing mesh reconstruction models perform unsatisfactorily in local geometry details due to ignoring mesh topology information during learning. Besides, most methods are constrained by the initial template, which cannot reconstruct meshes of various genus. In this work, we propose a geometric granularity-aware pixel-to-mesh framework with a fidelity-selection-and-guarantee strategy, which explicitly addresses both challenges. First, a geometry structure extractor is proposed for detecting local high structured parts and capturing local spatial feature. Second, we apply it to facilitate pixel-to-mesh mapping and resolve coarse details problem caused by the neglect of structural information in previous practices. Finally, a mesh edit module is proposed to encourage non-zero genus topology to emergence by fine-grained topology modification and a patching algorithm is introduced to repair the non-closed boundaries. Extensive experimental results, both quantitatively and visually have demonstrated the high reconstruction fidelity achieved by the proposed framework.

Deep Learning architectures, albeit successful in mostcomputer vision tasks, were designed for data with an un-derlying Euclidean structure, which is not usually fulfilledsince pre-processed data may lie on a non-linear space.In this paper, we propose a geometry aware deep learn-ing approach using rigid and non rigid transformation opti-mization for skeleton-based action recognition. Skeleton se-quences are first modeled as trajectories on Kendall's shapespace and then mapped to the linear tangent space. The re-sulting structured data are then fed to a deep learning archi-tecture, which includes a layer that optimizes over rigid andnon rigid transformations of the 3D skeletons, followed bya CNN-LSTM network. The assessment on two large scaleskeleton datasets, namely NTU-RGB+D and NTU-RGB+D120, has proven that the proposed approach outperformsexisting geometric deep learning methods and exceeds re-cently published approaches with respect to the majority of configurations.

In this paper, we propose a novel method for representation and classification of two-person interactions from 3D skeleton sequences. The key idea of our approach is to use Gaussian distributions to capture statistics on Rn and those on the space of symmetric positive definite (SPD) matrices. The main challenge is how to parametrize those distributions. Towards this end, we develop methods for embedding Gaussian distributions in matrix groups based on the theory of Lie groups and Riemannian symmetric spaces. Our method relies on the Riemannian geometry of the underlying manifolds and has the advantage of encoding high-order statistics from 3D joint positions. We show that the proposed method achieves competitive results in two-person interaction recognition on two large-scale benchmarks for 3D human activity understanding.

Contrastive learning methods have significantly narrowed the gap between supervised and unsupervised learning on computer vision tasks. In this paper, we explore their application to geo-located datasets, e.g. remote sensing, where unlabeled data is often abundant but labeled data is scarce. We first show that due to their different characteristics, a non-trivial gap persists between contrastive and supervised learning on standard benchmarks. To close the gap, we propose novel training methods that exploit the spatio-temporal structure of remote sensing data. We leverage spatially aligned images over time to construct temporal positive pairs in contrastive learning and geo-location to design pre-text tasks. Our experiments show that our proposed method closes the gap between contrastive and supervised learning on image classification, object detection and semantic segmentation for remote sensing. Moreover, we demonstrate that the proposed method can also be applied to geo-tagged ImageNet images, improving downstream performance on various tasks.

This paper presents a novel task together with a new benchmark for detecting generic, taxonomy-free event boundaries that segment a whole video into chunks. Conventional work in temporal video segmentation and action detection focuses on localizing pre-defined action categories and thus does not scale to generic videos. Cognitive Science has known since last century that humans consistently segment videos into meaningful temporal chunks. This segmentation happens naturally, without pre-defined event categories and without being explicitly asked to do so. Here, we repeat these cognitive experiments on mainstream CV datasets; with our novel annotation guideline which addresses the complexities of taxonomy-free event boundary annotation, we introduce the task of Generic Event Boundary Detection (GEBD) and the new benchmark Kinetics-GEBD. We view GEBD as an important stepping stone towards understanding the video as a whole, and believe it has been previously neglected due to a lack of proper task definition and annotations. Through experiment and human study we demonstrate the value of the annotations. Further, we benchmark supervised and un-supervised GEBD approaches on the TAPOS dataset and our Kinetics-GEBD. We release our annotations and baseline codes at CVPR'21 LOVEU Challenge: https://sites.google.com/view/loveucvpr21.

Transformers are increasingly dominating multi-modal reasoning tasks, such as visual question answering, achieving state-of-the-art results thanks to their ability to contextualize information using the self-attention and co-attention mechanisms. These attention modules also play a role in other computer vision tasks including object detection and image segmentation. Unlike Transformers that only use self-attention, Transformers with co-attention require to consider multiple attention maps in parallel in order to highlight the information that is relevant to the prediction in the model's input. In this work, we propose the first method to explain prediction by any Transformer-based architecture, including bi-modal Transformers and Transformers with co-attentions. We provide generic solutions and apply these to the three most commonly used of these architectures: (i) pure self-attention, (ii) self-attention combined with co-attention, and (iii) encoder-decoder attention. We show that our method is superior to all existing methods which are adapted from single modality explainability.

We propose a new generative model for layout generation. We generate layouts in three steps. First, we generate the layout elements as nodes in a layout graph. Second, we compute constraints between layout elements as edges in the layout graph. Third, we solve for the final layout using constrained optimization. For the first two steps, we build on recent transformer architectures. The layout optimization implements the constraints efficiently. We show three practical contributions compared to the state of the art: our work requires no user input, produces higher quality layouts, and enables many novel capabilities for conditional layout generation.

Inferring objects and their relationships from an image in the form of a scene graph is useful in many applications at the intersection of vision and language. We consider a challenging problem of compositional generalization that emerges in this task due to a long tail data distribution. Current scene graph generation models are trained on a tiny fraction of the distribution corresponding to the most frequent compositions, e.g. <cup, on, table>. However, test images might contain zero- and few-shot compositions of objects and relationships, e.g. <cup, on, surfboard>. Despite each of the object categories and the predicate (e.g. 'on') being frequent in the training data, the models often fail to properly understand such unseen or rare compositions. To improve generalization, it is natural to attempt increasing the diversity of the training distribution. However, in the graph domain this is non-trivial. To that end, we propose a method to synthesize rare yet plausible scene graphs by perturbing real ones. We then propose and empirically study a model based on conditional generative adversarial networks (GANs) that allows us to generate visual features of perturbed scene graphs and learn from them in a joint fashion. When evaluated on the Visual Genome dataset, our approach yields marginal, but consistent improvements in zero- and few-shot metrics. We analyze the limitations of our approach indicating promising directions for future research.

Deformable templates are essential to large-scale medical image registration, segmentation, and population analysis. Current conventional and deep network-based methods for template construction use only regularized registration objectives and often yield templates with blurry and/or anatomically implausible appearance, confounding downstream biomedical interpretation. We reformulate deformable registration and conditional template estimation as an adversarial game wherein we encourage realism in the moved templates with a generative adversarial registration framework conditioned on flexible image covariates. The resulting templates exhibit significant gain in specificity to attributes such as age and disease, better fit underlying group-wise spatiotemporal trends, and achieve improved sharpness and centrality. These improvements enable more accurate population modeling with diverse covariates for standardized downstream analyses and easier anatomical delineation for structures of interest.

Recent progress in stochastic motion prediction, i.e., predicting multiple possible future human motions given a single past pose sequence, has led to producing truly diverse future motions and even providing control over the motion of some body parts. However, to achieve this, the state-of-the-art method requires learning several mappings for diversity and a dedicated model for controllable motion prediction. In this paper, we introduce a unified deep generative network for both diverse and controllable motion prediction. To this end, we leverage the intuition that realistic human motions consist of smooth sequences of valid poses, and that, given limited data, learning a pose prior is much more tractable than a motion one. We therefore design a generator that predicts the motion of different body parts sequentially, and introduce a normalizing flow based pose prior, together with a joint angle loss, to achieve motion realism.Our experiments on two standard benchmark datasets, Human3.6M and HumanEva-I, demonstrate that our approach outperforms the state-of-the-art baselines in terms of both sample diversity and accuracy. The code is available at https://github.com/wei-mao-2019/gsps

Segmenting objects in videos is a fundamental computer vision task. The current deep learning based paradigm offers a powerful, but data-hungry solution. However, current datasets are limited by the cost and human effort of annotating object masks in videos. This effectively limits the performance and generalization capabilities of existing video segmentation methods. To address this issue, we explore weaker form of bounding box annotations. We introduce a method for generating segmentation masks from per-frame bounding box annotations in videos. To this end, we propose a spatio-temporal aggregation module that effectively mines consistencies in the object and background appearance across multiple frames. We use our predicted accurate masks to train video object segmentation (VOS) networks for the tracking domain, where only manual bounding box annotations are available. The additional data provides substantially better generalization performance, leading to state-of-the-art results on standard tracking benchmarks. The code and models are available at https://github.com/visionml/pytracking.

Attribution map visualization has arisen as one of the most effective techniques to understand the underlying inference process of Convolutional Neural Networks. In this task, the goal is to compute an score for each image pixel related to its contribution to the network output. In this paper, we introduce Disentangled Masked Backpropagation (DMBP), a novel gradient-based method that leverages on the piecewise linear nature of ReLU networks to decompose the model function into different linear mappings. This decomposition aims to disentangle the attribution maps into positive, negative and nuisance factors by learning a set of variables masking the contribution of each filter during back-propagation. A thorough evaluation over standard architectures (ResNet50 and VGG16) and benchmark datasets (PASCAL VOC and ImageNet) demonstrates that DMBP generates more visually interpretable attribution maps than previous approaches. Additionally, we quantitatively show that the maps produced by our method are more consistent with the true contribution of each pixel to the final network output.

Deep neural networks have significantly improved appearance-based gaze estimation accuracy. However, it still suffers from unsatisfactory performance when generalizing the trained model to new domains, e.g., unseen environments or persons. In this paper, we propose a plug-and-play gaze adaptation framework (PnP-GA), which is an ensemble of networks that learn collaboratively with the guidance of outliers. Since our proposed framework does not require ground-truth labels in the target domain, the existing gaze estimation networks can be directly plugged into PnP-GA and generalize the algorithms to new domains. We test PnP-GA on four gaze domain adaptation tasks, ETH-to-MPII, ETH-to-EyeDiap, Gaze360-to-MPII, and Gaze360-to-EyeDiap. The experimental results demonstrate that the PnP-GA framework achieves considerable performance improvements of 36.9%, 31.6%, 19.4%, and 11.8% over the baseline system. The proposed framework also outperforms the state-of-the-art domain adaptation approaches on gaze domain adaptation tasks.

Both generalized and incremental few-shot learning have to deal with three major challenges: learning novel classes from only few samples per class, preventing catastrophic forgetting of base classes, and classifier calibration across novel and base classes. In this work we propose a three-stage framework that allows to explicitly and effectively address these challenges. While the first phase learns base classes with many samples, the second phase learns a calibrated classifier for novel classes from few samples while also preventing catastrophic forgetting. In the final phase, calibration is achieved across all classes. We evaluate the proposed framework on four challenging benchmark datasets for image and video few-shot classification and obtain state-of-the-art results for both generalized and incremental few shot learning.

Domain adaptation (DA) aims to transfer the knowledge learned from source domain to an unlabeled target domain. Some recent works tackle source-free domain adaptation (SFDA) where only source pre-trained model is available for adaptation to target domain. However those methods does not consider keeping source performance which is of high practical value in real world application. In this paper, we propose a new domain adaptation paradigm denoted as Generalized Source-free Domain Adaptation (G-SFDA), where the learned model needs to perform well on both target and source domains, with only access to current unlabeled target data during adaptation. First, we propose local structure clustering (LSC), aiming to cluster the target features with its semantically similar neighbors, which successfully adapts the model to target domain in absence of source data. Second, we propose randomly generated domain attention (RGDA), it produces binary domain specific attention to activate different feature channels for different domains, meanwhile the domain attention will be utilized to regularize the gradient during adaptation to keep source information. In the experiments, for target performance our method is on par with or better than existing DA and SFDA methods, specifically achieves state-of-the-art performance (85.4%) on VisDA, and our method works well for all domains after adapting to single or multiple target domains.

We consider the shuffled linear regression problem where the correspondences between covariates and responses are unknown. While the existing formulation assumes an ideal underlying bijection in which all pieces of data should match, such an assumption barely holds in real-world applications due to either missing data or outliers. Therefore, in this work, we generalize the formulation of shuffled linear regression to a broader range of conditions where only part of the data should correspond. Moreover, we present a remarkably simple yet effective optimization algorithm with guaranteed global convergence. Distinct tasks validate the effectiveness of the proposed method.