ICCV 2021 论文列表

Deep learning based visual trackers entail offline pre-training on large volumes of video datasets with accurate bounding box annotations that are labor-expensive to achieve. We present a new framework to facilitate bounding box annotations for video sequences, which investigates a selection-and-refinement strategy to automatically improve the preliminary annotations generated by tracking algorithms. A temporal assessment network (T-Assess Net) is proposed which is able to capture the temporal coherence of target locations and select reliable tracking results by measuring their quality. Meanwhile, a visual-geometry refinement network (VG-Refine Net) is also designed to further enhance the selected tracking results by considering both target appearance and temporal geometry constraints, allowing inaccurate tracking results to be corrected. The combination of the above two networks provides a principled approach to ensure the quality of automatic video annotation. Experiments on large scale tracking benchmarks demonstrate that our method can deliver highly accurate bounding box annotations and significantly reduce human labor by 94.0%, yielding an effective means to further boost tracking performance with augmented training data.

We introduce Video Transformer (VidTr) with separable-attention for video classification. Comparing with commonly used 3D networks, VidTr is able to aggregate spatio-temporal information via stacked attentions and provide better performance with higher efficiency. We first introduce the vanilla video transformer and show that the transformer module is able to perform spatio-temporal modeling from raw pixels, but with heavy memory usage. We then present VidTr which reduces the memory cost by 3.3xwhile keeping the same performance. To further optimize the model, we propose the standard deviation based topK pooling for attention, which reduces the computation by dropping non-informative features along temporal dimension. VidTr achieves state-of-the-art performance on five commonly used datasets with lower computational requirements, showing both the efficiency and effectiveness of our design. Finally, error analysis and visualization show that VidTr is especially good at predicting actions that require long-term temporal reasoning.

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we will release code and models.

In this paper, we introduce a novel self-supervised visual representation learning method which understands both images and videos in a joint learning fashion. The proposed neural network architecture and objectives are designed to obtain two different Convolutional Neural Networks for solving visual recognition tasks in the domain of videos and images. Our method called Video/Image for Visual Contrastive Learning of Representation(Vi2CLR) uses unlabeled videos to exploit dynamic and static visual cues for self-supervised and instances similarity/dissimilarity learning. Vi2CLR optimization pipeline consists of visual clustering part and representation learning based on groups of similar positive instances within a cluster and negative ones from other clusters and learning visual clusters and their distances. We show how a joint self-supervised visual clustering and instance similarity learning with 2D (image) and 3D (video) CovNet encoders yields such robust and near to supervised learning performance. We extensively evaluate the method on downstream tasks like large scale action recognition and image and object classification on datasets like Kinetics, ImageNet, Pascal VOC'07 and UCF101 and achieve outstanding results compared to state-of-the-art self-supervised methods. To the best of our knowledge, the Vi2CLR is the first of its kind self-supervised neural network to tackle both video and image recognition task simultaneously by only using one source of data.

To improve the generalization of detectors, for domain adaptive object detection (DAOD), recent advances mainly explore aligning feature-level distributions between the source and single-target domain, which may neglect the impact of domain-specific information existing in the aligned features. Towards DAOD, it is important to extract domain-invariant object representations. To this end, in this paper, we try to disentangle domain-invariant representations from domain-specific representations. And we propose a novel disentangled method based on vector decomposition. Firstly, an extractor is devised to separate domain-invariant representations from the input, which are used for extracting object proposals. Secondly, domain-specific representations are introduced as the differences between the input and domain-invariant representations. Through the difference operation, the gap between the domain-specific and domain-invariant representations is enlarged, which promotes domain-invariant representations to contain more domain-irrelevant information. In the experiment, we separately evaluate our method on the single- and compound-target case. For the single-target case, experimental results of four domain-shift scenes show our method obtains a significant performance gain over baseline methods. Moreover, for the compound-target case (i.e., the target is a compound of two different domains without domain labels), our method outperforms baseline methods by around 4%, which demonstrates the effectiveness of our method.

Invariance and equivariance to the rotation group have been widely discussed in the 3D deep learning community for pointclouds. Yet most proposed methods either use complex mathematical tools that may limit their accessibility, or are tied to specific input data types and network architectures. In this paper, we introduce a general framework built on top of what we call Vector Neuron representations for creating SO(3)-equivariant neural networks for pointcloud processing. Extending neurons from 1D scalars to 3D vectors, our vector neurons enable a simple mapping of SO(3) actions to latent spaces thereby providing a framework for building equivariance in common neural operations -- including linear layers, non-linearities, pooling, and normalizations. Due to their simplicity, vector neurons are versatile and, as we demonstrate, can be incorporated into diverse network architecture backbones, allowing them to process geometry inputs in arbitrary poses. Despite its simplicity, our method performs comparably well in accuracy and generalization with other more complex and specialized state-of-the-art methods on classification and segmentation tasks. We also show for the first time a rotation equivariant reconstruction network. Source code is available at https://github.com/FlyingGiraffe/vnn.

Data augmentation is an intuitive step towards solving the problem of few-shot classification. However, ensuring both discriminability and diversity in the augmented samples is challenging. To address this, we propose a feature disentanglement framework that allows us to augment features with randomly sampled intra-class variations while preserving their class-discriminative features. Specifically, we disentangle a feature representation into two components: one represents the intra-class variance and the other encodes the class-discriminative information. We assume that the intra-class variance induced by variations in poses, backgrounds, or illumination conditions is shared across all classes and can be modelled via a common distribution. Then we sample features repeatedly from the learned intra-class variability distribution and add them to the class-discriminative features to get the augmented features. Such a data augmentation scheme ensures that the augmented features inherit crucial class-discriminative features while exhibiting large intra-class variance. Our method significantly outperforms the state-of-the-art methods on multiple challenging fine-grained few-shot image classification benchmarks. Code is available at: https://github.com/cvlab-stonybrook/vfd-iccv21

In crowd counting, due to the problem of laborious labelling, it is perceived intractability of collecting a new large-scale dataset which has plentiful images with large diversity in density, scene, etc. Thus, for learning a general model, training with data from multiple different datasets might be a remedy and be of great value. In this paper, we resort to the multi-domain joint learning and propose a simple but effective Domain-specific Knowledge Propagating Network (DKPNet) for unbiasedly learning the knowledge from multiple diverse data domains at the same time. It is mainly achieved by proposing the novel Variational Attention(VA) technique for explicitly modeling the attention distributions for different domains. And as an extension to VA, Intrinsic Variational Attention(InVA) is proposed to handle the problems of over-lapped domains and sub-domains. Extensive experiments have been conducted to validate the superiority of our DKPNet over several popular datasets, including ShanghaiTech A/B, UCF-QNRF and NWPU.

We propose a versatile deep image compression network based on Spatial Feature Transform (SFT), which takes a source image and a corresponding quality map as inputs and produce a compressed image with variable rates. Our model covers a wide range of compression rates using a single model, which is controlled by arbitrary pixel-wise quality maps. In addition, the proposed framework allows us to perform task-aware image compressions for various tasks, e.g., classification, by efficiently estimating optimized quality maps specific to target tasks for our encoding network. This is even possible with a pretrained network without learning separate models for individual tasks. Our algorithm achieves outstanding rate-distortion trade-off compared to the approaches based on multiple models that are optimized separately for several different target rates. At the same level of compression, the proposed approach successfully improves performance on image classification and text region quality preservation via task-aware quality map estimation without additional model training. The code is available at the project website https://github.com/micmic123/QmapCompression.

Deep generative models can synthesize photorealistic images of human faces with novel identities.However, a key challenge to the wide applicability of such techniques is to provide independent control over semantically meaningful parameters: appearance, head pose, face shape, and facial expressions. In this paper, we propose VariTex - to the best of our knowledge the first method that learns a variational latent feature space of neural face textures, which allows sampling of novel identities. We combine this generative model with a parametric face model and gain explicit control over head pose and facial expressions. To generate complete images of human heads, we propose an additive decoder that adds plausible details such as hair. A novel training scheme enforces a pose-independent latent space and in consequence, allows learning a one-to-many mapping between latent codes and pose-conditioned exterior regions. The resulting method can generate geometrically consistent images of novel identities under fine-grained control over head pose, face shape, and facial expressions. This facilitates a broad range of downstream tasks, like sampling novel identities, changing the head pose, expression transfer, and more.

Being able to infer 3D structures from 2D images with geometric principles, vanishing points have been a well-recognized concept in 3D vision research. It has been widely used in autonomous driving, SLAM, and AR/VR for applications including road direction estimation, camera calibration, and camera pose estimation. Existing vanishing point detection methods often need to trade off between robustness, precision, and inference speed. In this paper, we introduce VaPiD, a novel neural network-based rapid Vanishing Point Detector that achieves unprecedented efficiency with learned vanishing point optimizers. The core of our method contains two components: a vanishing point proposal network that gives a set of vanishing point proposals as coarse estimations; and a neural vanishing point optimizer that iteratively optimizes the positions of the vanishing point proposals to achieve high-precision levels. Extensive experiments on both synthetic and real-world datasets show that our method provides competitive, if not better, performance as compared to the previous state-of-the-art vanishing point detection approaches, while being significantly faster.

We present VSAC, a RANSAC-type robust estimator with a number of novelties. It benefits from the introduction of the concept of independent inliers that improves significantly the efficacy of the dominant plane handling and also allows near error-free rejection of incorrect models, without false positives. The local optimization process and its application is improved so that it is run on average only once. Further technical improvements include adaptive sequential hypothesis verification and efficient model estimation via Gaussian elimination. Experiments on four standard datasets show that VSAC is significantly faster than all its predecessors and runs on average in 1-2 ms, on a CPU. It is two orders of magnitude faster and yet as precise as MAGSAC++, the currently most accurate estimator of two-view geometry. In the repeated runs on EVD, HPatches, PhotoTourism, and Kusvod2 datasets, it never failed.

In recent years, sparse voxel-based methods have become the state-of-the-arts for 3D semantic segmentation of indoor scenes, thanks to the powerful 3D CNNs. Nevertheless, being oblivious to the underlying geometry, voxel-based methods suffer from ambiguous features on spatially close objects and struggle with handling complex and irregular geometries due to the lack of geodesic information. In view of this, we present Voxel-Mesh Network (VMNet), a novel 3D deep architecture that operates on the voxel and mesh representations leveraging both the Euclidean and geodesic information. Intuitively, the Euclidean information extracted from voxels can offer contextual cues representing interactions between nearby objects, while the geodesic information extracted from meshes can help separate objects that are spatially close but have disconnected surfaces. To incorporate such information from the two domains, we design an intra-domain attentive module for effective feature aggregation and an inter-domain attentive module for adaptive feature fusion. Experimental results validate the effectiveness of VMNet: specifically, on the challenging ScanNet dataset for large-scale segmentation of indoor scenes, it outperforms the state-of-the-art SparseConvNet and MinkowskiNet (74.6% vs 72.5% and 73.6% in mIoU) with a simpler network structure (17M vs 30M and 38M parameters). Code release: https://github.com/hzykent/VMNet

Cognitive grammar suggests that the acquisition of language grammar is grounded within visual structures. While grammar is an essential representation of natural language, it also exists ubiquitously in vision to represent the hierarchical part-whole structure. In this work, we study grounded grammar induction of vision and language in a joint learning framework. Specifically, we present VLGrammar, a method that uses compound probabilistic context-free grammars (compound PCFGs) to induce the language grammar and the image grammar simultaneously. We propose a novel contrastive learning framework to guide the joint learning of both modules. To provide a benchmark for the grounded grammar induction task, we collect a large-scale dataset, PartIt, which contains human-written sentences that describe part-level semantics for 3D objects. Experiments on the PartIt dataset show that VLGrammar outperforms all baselines in image grammar induction and language grammar induction. The learned VLGrammar naturally benefits related downstream tasks. Specifically, it improves the image unsupervised clustering accuracy by 30%, and performs well in image retrieval and text retrieval. Notably, the induced grammar shows superior generalizability by easily generalizing to unseen categories.

Lane detection plays a key role in autonomous driving. While car cameras always take streaming videos on the way, current lane detection works mainly focus on individual images (frames) by ignoring dynamics along the video. In this work, we collect a new video instance lane detection (VIL-100) dataset, which contains 100 videos with in total 10,000 frames, acquired from different real traffic scenarios. All the frames in each video are manually annotated to a high-quality instance-level lane annotation, and a set of frame-level and video-level metrics are included for quantitative performance evaluation. Moreover, we propose a new baseline model, named multi-level memory aggregation network (MMA-Net), for video instance lane detection. In our approach, the representation of current frame is enhanced by attentively aggregating both local and global memory features from other frames. Experiments on the new collected dataset show that the proposed MMA-Net outperforms state-of-the-art lane detection methods and video object segmentation methods. We release our dataset and code at https://github.com/yujun0-0/MMA-Net.

Hough voting, as has been demonstrated in VoteNet, is effective for 3D object detection, where voting is a key step. In this paper, we propose a novel VoteNet-based 3D detector with vote enhancement to improve the detection accuracy in cluttered indoor scenes. It addresses the limitations of current voting schemes, i.e., votes from neighboring objects and background have significant negative impacts Specifically, before voting, we replace the classic MLP with the proposed Attentive MLP (AMLP) in the backbone network to get better feature description of seed points. During voting, we design a new vote attraction loss (VALoss) to enforce vote centers to locate closely and compactly to the corresponding object centers. After voting, we then devise a vote weighting module to integrate the foreground/background prediction into the vote aggregation process to enhance the capability of the original VoteNet to handle noise from background voting. The three proposed strategies all contribute to more effective voting and improved performance, resulting in a novel 3D object detector, termed VENet. Experiments show that our method outperforms state-of-the-art methods on benchmark datasets. Ablation studies demonstrate the effectiveness of the proposed components.

Real world images often gets corrupted due to unwanted reflections and their removal is highly desirable. A major share of such images originate from smart phone cameras capable of very high resolution captures. Most of the existing methods either focus on restoration quality by compromising on processing speed and memory requirements or, focus on removing reflections at very low resolutions, there by limiting their practical deploy-ability. We propose a light weight deep learning model for reflection removal using a novel scale space architecture. Our method processes the corrupted image in two stages, a Low Scale Sub-network (LSSNet) to process the lowest scale and a Progressive Inference (PI) stage to process all the higher scales. In order to reduce the computational complexity, the sub-networks in PI stage are designed to be much shallower than LSSNet. Moreover, we employ weight sharing between various scales within the PI stage to limit the model size. This also allows our method to generalize to very high resolutions without explicit retraining. Our method is superior both qualitatively and quantitatively compared to the state of the art methods and at the same time 20x faster with 50x less number of parameters compared to the most recent state-of-the-art algorithm RAGNet. We implemented our method on an android smart phone, where a high resolution 12 MP image is restored in under 5 seconds.

Being able to learn dense semantic representations of images without supervision is an important problem in computer vision. However, despite its significance, this problem remains rather unexplored, with a few exceptions that considered unsupervised semantic segmentation on small-scale datasets with a narrow visual domain. In this paper, we make a first attempt to tackle the problem on datasets that have been traditionally utilized for the supervised case. To achieve this, we introduce a two-step framework that adopts a predetermined mid-level prior in a contrastive optimization objective to learn pixel embeddings. This marks a large deviation from existing works that relied on proxy tasks or end-to-end clustering. Additionally, we argue about the importance of having a prior that contains information about objects, or their parts, and discuss several possibilities to obtain such a prior in an unsupervised manner. Experimental evaluation shows that our method comes with key advantages over existing works. First, the learned pixel embeddings can be directly clustered in semantic groups using K-Means on PASCAL. Under the fully unsupervised setting, there is no precedent in solving the semantic segmentation task on such a challenging benchmark. Second, our representations can improve over strong baselines when transferred to new datasets, e.g. COCO and DAVIS. The code is available.

We present an image segmentation algorithm that is developed in an unsupervised deep learning framework. The delineation of object boundaries often fails due to the nuisance factors such as illumination changes and occlusions. Thus, we initially propose an unsupervised image decomposition algorithm to obtain an intrinsic representation that is robust with respect to undesirable bias fields based on a multiplicative image model. The obtained intrinsic image is subsequently provided to an unsupervised segmentation procedure that is developed based on a piecewise smooth model. The segmentation model is further designed to incorporate a geometric constraint imposed in the generative adversarial network framework where the discrepancy between the distribution of partitioning functions and the distribution of prior shapes is minimized. We demonstrate the effectiveness and robustness of the proposed algorithm in particular with bias fields and occlusions using simple yet illustrative synthetic examples and a benchmark dataset for image segmentation.