CVPR 2021 Papers

Self-supervised learning presents a remarkable performance to utilize unlabeled data for various video tasks. In this paper, we focus on applying the power of self-supervised methods to improve semi-supervised action proposal generation. Particularly, we design a Self-supervised Semi-supervised Temporal Action Proposal (SSTAP) framework. The SSTAP contains two crucial branches, i.e., temporal-aware semi-supervised branch and relation-aware self-supervised branch. The semi-supervised branch improves the proposal model by introducing two temporal perturbations, i.e., temporal feature shift and temporal feature flip, in the mean teacher framework. The self-supervised branch defines two pretext tasks, including masked feature reconstruction and clip-order prediction, to learn the relation of temporal clues. By this means, SSTAP can better explore unlabeled videos, and improve the discriminative abilities of learned action features. We extensively evaluate the proposed SSTAP on THUMOS14 and ActivityNet v1.3 datasets. The experimental results demonstrate that SSTAP significantly outperforms state-of-the-art semi-supervised methods and even matches fully-supervised methods. The code will be released once this paper is accepted.

We present self-supervised geometric perception (SGP), the first general framework to learn a feature descriptor for correspondence matching without any ground-truth geometric model labels (e.g., camera poses, rigid transformations). Our first contribution is to formulate geometric perception as an optimization problem that jointly optimizes the feature descriptor and the geometric models given a large corpus of visual measurements (e.g., images, point clouds). Under this optimization formulation, we show that two important streams of research in vision, namely robust model fitting and deep feature learning, correspond to optimizing one block of the unknown variables while fixing the other block. This analysis naturally leads to our second contribution - the SGP algorithm that performs alternating minimization to solve the joint optimization. SGP iteratively executes two meta-algorithms: a teacher that performs robust model fitting given learned features to generate geometric pseudo-labels, and a student that performs deep feature learning under noisy supervision of the pseudo-labels. As a third contribution, we apply SGP to two perception problems on large-scale real datasets, namely relative camera pose estimation on MegaDepth and point cloud registration on 3DMatch. We demonstrate that SGP achieves state-of-the-art performance that is on-par or superior to the supervised oracles trained using ground-truth labels.

We propose a new generative model for 3D garment deformations that enables us to learn, for first time, a data-driven method for virtual try-on that effectively addresses garment-body collisions. In contrast to existing methods that require an undesirable postprocessing step to fix garment-body interpenetrations at test time, our approach directly outputs 3D garment configurations that do not collide with the underlying body. Key to our success is a new canonical space for garments that removes pose-and-shape deformations already captured by a new diffused human body model, which extrapolates body surface properties such as skinning weights and blendshapes to any 3D point. We leverage this representation to train a generative model with a novel self-supervised collision term that learns to reliably solve garment-body interpenetrations. We extensively evaluate and compare our results with recently proposed data-driven methods, and show that our method is the first to successfully address garment-body contact in unseen body shapes and motions, without compromising the realism and detail.

We propose an approach to domain adaptation for semantic segmentation that is both practical and highly accurate. In contrast to previous work, we abandon the use of computationally involved adversarial objectives, network ensembles and style transfer. Instead, we employ standard data augmentation techniques - photometric noise, flipping and scaling - and ensure consistency of the semantic predictions across these image transformations. We develop this principle in a lightweight self-supervised framework trained on co-evolving pseudo labels without the need for cumbersome extra training rounds. Simple in training from a practitioner's standpoint, our approach is remarkably effective. We achieve significant improvements of the state-of-the-art segmentation accuracy after adaptation, consistent both across different choices of the backbone architecture and adaptation scenarios.

Recent single-view 3D reconstruction methods reconstruct object's shape and texture from a single image with only 2D image-level annotation. However, without explicit 3D attribute-level supervision, it is still difficult to achieve satisfying reconstruction accuracy. In this paper, we propose a Self-supervised Mesh Reconstruction (SMR) approach to enhance 3D mesh attribute learning process. Our approach is motivated by observations that (1) 3D attributes from interpolation and prediction should be consistent, and (2) feature representation of landmarks from all images should be consistent. By only requiring silhouette mask annotation, our SMR can be trained in an end-to-end manner and generalizes to reconstruct natural objects of birds, cows, motorbikes, etc. Experiments demonstrate that our approach improves both 2D supervised and unsupervised 3D mesh reconstruction on multiple datasets. We also show that our model can be adapted to other image synthesis tasks, e.g., novel view generation, shape transfer, and texture transfer, with promising results. Our code is publicly available at https://github.com/Jia-Research-Lab.

Few-shot class-incremental learning is to recognize the new classes given few samples and not forget the old classes. It is a challenging task since representation optimization and prototype reorganization can only be achieved under little supervision. To address this problem, we propose a novel incremental prototype learning scheme. Our scheme consists of a random episode selection strategy that adapts the feature representation to various generated incremental episodes to enhance the corresponding extensibility, and a self-promoted prototype refinement mechanism which strengthens the expression ability of the new class by explicitly considering the dependencies among different classes. Particularly, a dynamic relation projection module is proposed to calculate the relation matrix in a shared embedding space and leverage it as the factor for bootstrapping the update of prototypes. Extensive experiments on three benchmark datasets demonstrate the above-par incremental performance, outperforming state-of-the-art methods by a margin of 13%, 17% and 11%, respectively.

Due to the scarcity of annotated scene flow data, self-supervised scene flow learning in point clouds has attracted increasing attention. In the self-supervised manner, establishing correspondences between two point clouds to approximate scene flow is an effective approach. Previous methods often obtain correspondences by applying point-wise matching that only takes the distance on 3D point coordinates into account, introducing two critical issues: (1) it overlooks other discriminative measures, such as color and surface normal, which often bring fruitful clues for accurate matching; and (2) it often generates sub-par performance, as the matching is operated in an unconstrained situation, where multiple points can be ended up with the same corresponding point. To address the issues, we formulate this matching task as an optimal transport problem. The output optimal assignment matrix can be utilized to guide the generation of pseudo ground truth. In this optimal transport, we design the transport cost by considering multiple descriptors and encourage one-to-one matching by mass equality constraints. Also, constructing a graph on the points, a random walk module is introduced to encourage the local consistency of the pseudo labels. Comprehensive experiments on FlyingThings3D and KITTI show that our method achieves state-of-the-art performance among self-supervised learning methods. Our self-supervised method even performs on par with some supervised learning approaches, although we do not need any ground truth flow for training.

Few-shot segmentation has been attracting a lot of attention due to its effectiveness to segment unseen object classes with a few annotated samples. Most existing approaches use masked Global Average Pooling (GAP) to encode an annotated support image to a feature vector to facilitate query image segmentation. However, this pipeline unavoidably loses some discriminative information due to the average operation. In this paper, we propose a simple but effective self-guided learning approach, where the lost critical information is mined. Specifically, through making an initial prediction for the annotated support image, the covered and uncovered foreground regions are encoded to the primary and auxiliary support vectors using masked GAP, respectively. By aggregating both the primary and auxiliary support vectors, better segmentation performance is obtained on query images. Enlightened by our self-guided module for 1-shot segmentation, we propose a cross-guided module for multiple shot segmentation, where the final mask is fused using predictions from multiple annotated samples with high-quality support vectors contributing more and vice versa. This module improves the final prediction in the inference stage without re-training. Extensive experiments show that our approach achieves new state-of-the-art performances on both PASCAL-5i and COCO-20i datasets. Source code will be released once the paper is accepted.

Although existing fully-supervised defocus blur detection (DBD) models significantly improve performance, training such deep models requires abundant pixel-level manual annotation, which is highly time-consuming and error-prone. Addressing this issue, this paper makes an effort to train a deep DBD model without using any pixel-level annotation. The core insight is that a defocus blur region/focused clear area can be arbitrarily pasted to a given realistic full blurred image/full clear image without affecting the judgment of the full blurred image/full clear image. Specifically, we train a generator G in an adversarial manner against dual discriminators Dc and Db. G learns to produce a DBD mask that generates a composite clear image and a composite blurred image through copying the focused area and unfocused region from corresponding source image to another full clear image and full blurred image. Then, Dc and Db can not distinguish them from realistic full clear image and full blurred image simultaneously, achieving a self-generated DBD by an implicit manner to define what a defocus blur area is. Besides, we propose a bilateral triplet-excavating constraint to avoid the degenerate problem caused by the case one discriminator defeats the other one. Comprehensive experiments on two widely-used DBD datasets demonstrate the superiority of the proposed approach. Source codes are available at: https://github.com/shangcai1/SG.

Pre-trained models play an important role in deep learning based text detectors. However, most methods ignore the gap between natural images and scene text images and directly apply ImageNet for pre-training. To address such a problem, some of them firstly pre-train the model using a large amount of synthetic data and then fine-tune it on target datasets, which is task-specific and has limited generalization capability. In this paper, we focus on providing general pre-trained models for text detectors. Considering the importance of exploring text contents for text detection, we propose STKM (Self-attention based Text Knowledge Mining), which consists of a CNN Encoder and a Self-attention Decoder, to learn general prior knowledge for text detection from SynthText. Given only image level text labels, Self-attention Decoder directly decodes features extracted from CNN Encoder to texts without requirement of detection, which guides the CNN backbone to explicitly learn discriminative semantic representations ignored by previous approaches. After that, the text knowledge learned by the backbone can be transferred to various text detectors to significantly improve their detection performance (e.g., 5.89% higher F-measure for EAST on ICDAR15 dataset) without bells and whistles. Pre-trained model is available at: https://github.com/CVI-SZU/STKM

In this paper, we address the problems of rain streaks and rain accumulation removal in video, by developing a self-aligned network with transmission-depth consistency. Existing video based deraining method focus only on rain streak removal, and commonly use optical flow to align the rain video frames. However, besides rain streaks, rain accummulation can considerably degrade visibility; and, optical flow estimation in a rain video is still erroneous, making the deraining performance tend to be inaccurate. Our method employs deformable convolution layers in our encoder to achieve feature-level frame alignment, and hence avoids using optical flow. For rain streaks, our method predicts the current frame from its adjacent frames, such that rain streaks that appear randomly in the temporal domain can be removed. For rain accumulation, our method employs transmission-depth consistency to resolve the ambiguity between the depth and water-droplet density. Our network estimates the depth from consecutive rain-accumulation-removal outputs, and we calculate the transmission map using a commonly used physics model. To ensure photometric-temporal and depth-temporal consistencies, our network also estimate the camera poses, so that we can warp one frame to its adjacent frames. Experimental results show that our method is effective in removing both rain streaks and rain accumulation. Our results outperform those of state-of-the-art methods quantitatively and qualitatively.

Instance segmentation has witnessed a remarkable progress on class-balanced benchmarks. However, they fail to perform as accurately in real-world scenarios, where the category distribution of objects naturally comes with a long tail. Instances of head classes dominate a long-tailed dataset and they serve as negative samples of tail categories. The overwhelming gradients of negative samples on tail classes lead to a biased learning process for classifiers. Consequently, objects of tail categories are more likely to be misclassified as backgrounds or head categories. To tackle this problem, we propose Seesaw Loss to dynamically re-balance gradients of positive and negative samples for each category, with two complementary factors, i.e., mitigation factor and compensation factor. The mitigation factor reduces punishments to tail categories w.r.t. the ratio of cumulative training instances between different categories. Meanwhile, the compensation factor increases the penalty of misclassified instances to avoid false positives of tail categories. We conduct extensive experiments on Seesaw Loss with mainstream frameworks and different data sampling strategies. With a simple end-to-end training pipeline, Seesaw Loss obtains significant gains over Cross-Entropy Loss, and achieves state-of-the-art performance on LVIS dataset without bells and whistles. Code is available at https://github.com/open-mmlab/mmdetection.

Nowadays, we have witnessed the early progress on learning the association between voice and face automatically, which brings a new wave of studies to the computer vision community. However, most of the prior arts along this line (a) merely adopt local information to perform modality alignment and (b) ignore the diversity of learning difficulty across different subjects. In this paper, we propose a novel framework to jointly address the above-mentioned issues. Targeting at (a), we propose a two-level modality alignment loss where both global and local information are considered. Compared with the existing methods, we introduce a global loss into the modality alignment process. The global component of the loss is driven by the accuracy of the identity classification. Theoretically, we show that minimizing the loss could maximize the distance between embeddings across different identities while minimizing the distance between embeddings belonging to the same identity, in a global sense (instead of a mini-batch). Targeting at (b), we propose a dynamic reweighting scheme to better explore the hard but valuable identities while filtering out the unlearnable and noisy identities. Experiments show that the proposed method outperforms the previous methods in multiple settings, including voice-face matching, verification and retrieval.

We propose a new flash technique for low-light imaging, using deep-red light as an illuminating source. Our main observation is that in a dim environment, the human eye mainly uses rods for the perception of light, which are not sensitive to wavelengths longer than 620nm, yet the camera sensor still has a spectral response. We propose a novel modulation strategy when training a modern CNN model for guided image filtering, fusing a noisy RGB frame and a flash frame. This fusion network is further extended for video reconstruction. We have built a prototype with minor hardware adjustments and tested the new flash technique on a variety of static and dynamic scenes. The experimental results demonstrate that our method produces compelling reconstructions, even in extra dim conditions.

We study on joint learning of Convolutional Neural Network (CNN) and Transformer for vision-language pre-training (VLPT) which aims to learn cross-modal alignments from millions of image-text pairs. State-of-the-art approaches extract salient image regions and align regions with words step-by-step. As region-based representations usually represent parts of an image, it is challenging for existing models to fully understand the semantics from paired natural languages. In this paper, we propose SOHO to ""See Out of tHe bOx"" that takes a full image as input, and learns vision-language representation in an end-to-end manner. SOHO does not require bounding box annotations, while enables 10 times faster inference than region-based approaches. In particular, SOHO learns to extract comprehensive yet compact image features through a visual dictionary (VD) that facilitates cross-modal understanding. VD is designed to represent consistent visual abstractions of similar semantics, and VD can be further updated on-the-fly during pre-training. We conduct experiments on four well-established vision-language tasks by following standard VLPT settings. SOHO achieves absolute gains of 2.0% R@1 score on MSCOCO text retrieval 5k test split, 1.5% accuracy on NLVR2 test-P split, 6.7% accuracy on SNLI-VE test split, respectively.

Multi-object tracking from RGB-D video sequences is a challenging problem due to the combination of changing viewpoints, motion, and occlusions over time. We observe that having the complete geometry of objects aids in their tracking, and thus propose to jointly infer the complete geometry of objects as well as track them, for rigidly moving objects over time. Our key insight is that inferring the complete geometry of the objects significantly helps in tracking. By hallucinating unseen regions of objects, we can obtain additional correspondences between the same instance, thus providing robust tracking even under strong change of appearance. From a sequence of RGB-D frames, we detect objects in each frame and learn to predict their complete object geometry as well as a dense correspondence mapping into a canonical space. This allows us to derive 6DoF poses for the objects in each frame, along with their correspondence between frames, providing robust object tracking across the RGB-D sequence. Experiments on both synthetic and real-world RGB-D data demonstrate that we achieve state-of-the-art performance on 3D multi-object tracking. Furthermore, we show that our object completion significantly helps tracking, providing an improvement of 8% in mean MOTA.

Training deep neural networks requires gradient estimation from data batches to update parameters. Gradients per parameter are averaged over a set of data and this has been presumed to be safe for privacy-preserving training in joint, collaborative, and federated learning applications. Prior work only showed the possibility of recovering input data given gradients under very restrictive conditions - a single input point, or a network with no non-linearities, or a small 32x32 px input batch. Therefore, averaging gradients over larger batches was thought to be safe. In this work, we introduce GradInversion, using which input images from a larger batch (8 - 48 images) can also be recovered for large networks such as ResNets (50 layers), on complex datasets such as ImageNet (1000 classes, 224x224 px). We formulate an optimization task that converts random noise into natural images, matching gradients while regularizing image fidelity. We also propose an algorithm for target class label recovery given gradients. We further propose a group consistency regularization framework, where multiple agents starting from different random seeds work together to find an enhanced reconstruction of original data batch. We show that gradients encode a surprisingly large amount of information, such that all the individual images can be recovered with high fidelity via GradInversion, even for complex datasets, deep networks, and large batch sizes.

Neural Architecture Search (NAS), together with model scaling, has shown remarkable progress in designing high accuracy and fast convolutional architecture families. However, as neither NAS nor model scaling considers sufficient hardware architecture details, they do not take full advantage of the emerging datacenter (DC) accelerators. In this paper, we search for fast and accurate CNN model families for efficient inference on DC accelerators. We first analyze DC accelerators and find that existing CNNs suffer from insufficient operational intensity, parallelism, and execution efficiency and exhibit FLOPs-latency nonproportionality. These insights let us create a DC-accelerator-optimized search space, with space-to-depth, space-to-batch, hybrid fused convolution structures with vanilla and depthwise convolutions, and block-wise activation functions. We further propose a latency-aware compound scaling (LACS), the first multi-objective compound scaling method optimizing both accuracy and latency. Our LACS discovers that network depth should grow much faster than image size and network width, which is quite different from the observations from previous compound scaling. With the new search space and LACS, our search and scaling on datacenter accelerators results in a new model series named EfficientNet-X. EfficientNet-X is up to more than 2X faster than EfficientNet (a model series with state-of-the-art trade-off on FLOPs and accuracy) on TPUv3 and GPUv100, with comparable accuracy. EfficientNet-X is also up to 7X faster than recent RegNet and ResNeSt on TPUv3 and GPUv100. Source code is at https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet/tpu

In one-shot NAS, sub-networks need to be searched from the supernet to meet different hardware constraints. However, the search cost is high and N times of searches are needed for N different constraints. In this work, we propose a novel search strategy called architecture generator to search sub-networks by generating them, so that the search process can be much more efficient and flexible. With the trained architecture generator, given target hardware constraints as the input, N good architectures can be generated for N constraints by just one forward pass without re-searching and supernet retraining. Moreover, we propose a novel single-path supernet, called unified supernet, to further improve search efficiency and reduce GPU memory consumption of the architecture generator. With the architecture generator and the unified supernet, we propose a flexible and efficient one-shot NAS framework, called Searching by Generating NAS (SGNAS). With the pre-trained supernt, the search time of SGNAS for N different hardware constraints is only 5 GPU hours, which is 4N times faster than previous SOTA single-path methods. After training from scratch, the top1-accuracy of SGNAS on ImageNet is 77.1%, which is comparable with the SOTAs. The code is available at: https://github.com/eric8607242/SGNAS.