ICCV 2021 Papers

Self-supervised learning has been successfully applied to pre-train video representations, which aims at efficient adaptation from pre-training domain to downstream tasks. Existing approaches merely leverage contrastive loss to learn instance-level discrimination. However, lack of category information will lead to hard-positive problem that constrains the generalization ability of this kind of methods. We find that the multi-task process of meta learning can provide a solution to this problem. In this paper, we propose a Meta-Contrastive Network (MCN), which combines the contrastive learning and meta learning, to enhance the learning ability of existing self-supervised approaches. Our method contains two training stages based on model-agnostic meta learning (MAML), each of which consists of a contrastive branch and a meta branch. Extensive evaluations demonstrate the effectiveness of our method. For two downstream tasks, i.e., video action recognition and video retrieval, MCN outperforms state-of-the-art approaches on UCF101 and HMDB51 datasets. To be more specific, with R(2+1)D backbone, MCN achieves Top-1 accuracies of 84.8% and 54.5% for video action recognition, as well as 52.5% and 23.7% for video retrieval.

Animals have evolved highly functional visual systems to understand motion, assisting perception even under complex environments. In this paper, we work towards developing a computer vision system able to segment objects by exploiting motion cues, i.e. motion segmentation. To achieve this, we introduce a simple variant of the Transformer to segment optical flow frames into primary objects and the background, which can be trained in a self-supervised manner, i.e. without using any manual annotations. Despite using only optical flow, and no appearance information, as input, our approach achieves superior results compared to previous state-of-the-art self-supervised methods on public benchmarks (DAVIS2016, SegTrackv2, FBMS59), while being an order of magnitude faster. On a challenging camouflage dataset (MoCA), we significantly outperform other self-supervised approaches, and are competitive with the top supervised approach, highlighting the importance of motion cues and the potential bias towards appearance in existing video segmentation models.

Vessel segmentation is critically essential for diagnosinga series of diseases, e.g., coronary artery disease and retinal disease. However, annotating vessel segmentation maps of medical images is notoriously challenging due to the tiny and complex vessel structures, leading to insufficient available annotated datasets for existing supervised methods and domain adaptation methods. The subtle structures and confusing background of medical images further suppress the efficacy of unsupervised methods. In this paper, we propose a self-supervised vessel segmentation method via adversarial learning. Our method learns vessel representations by training an attention-guided generator and a segmentation generator to simultaneously synthesize fake vessels and segment vessels out of coronary angiograms. To support the research, we also build the first X-ray angiography coronary vessel segmentation dataset, named XCAD. We evaluate our method extensively on multiple vessel segmentation datasets, including the XCAD dataset, the DRIVE dataset,and the STARE dataset. The experimental results show our method suppresses unsupervised methods significantly and achieves competitive performance compared with supervised methods and traditional methods.

Traditional methods for RGB hand mesh recovery usually need to train a separate model for each dataset with the corresponding ground truth and are hardly adapted to new scenarios without the ground truth for supervision. To address the problem, we propose a self-supervised framework for hand mesh estimation, where we pre-learn hand priors from existing hand datasets and transfer the priors to new scenarios without any landmark annotations. The proposed approach takes binocular images as input and mainly relies on left-right consistency constraints including appearance consensus and shape consistency to train the model to estimate the hand mesh in new scenarios. We conduct experiments on the widely used stereo hand dataset, and the experimental results verify that our model can get comparable performance compared with state-of-the-art methods even without the corresponding landmark annotations. To further evaluate our model, we collect a large real binocular dataset. The experimental results on the collected real dataset also verify the effectiveness of our model qualitatively.

Self-supervised representation learning is able to learn semantically meaningful features; however, much of its recent success relies on multiple crops of an image with very few objects. Instead of learning view-invariant representation from simple images, humans learn representations in a complex world with changing scenes by observing object movement, deformation, pose variation, and ego motion. Motivated by this ability, we present a new self-supervised learning representation framework that can be directly deployed on a video stream of complex scenes with many moving objects. Our framework features a simple flow equivariance objective that encourages the network to predict the features of another frame by applying a flow transformation to the features of the current frame. Our representations, learned from high-resolution raw video, can be readily used for downstream tasks on static images. Readout experiments on challenging semantic segmentation, instance segmentation, and object detection benchmarks show that we are able to outperform representations obtained from previous state-of-the-art methods including SimCLR and BYOL.

Synthetic data is emerging as a promising solution to the scalability issue of supervised deep learning, especially when real data are difficult to acquire or hard to annotate. Synthetic data generation, however, can itself be prohibitively expensive when domain experts have to manually and painstakingly oversee the process. Moreover, neural networks trained on synthetic data often do not perform well on real data because of the domain gap. To solve these challenges, we propose Sim2SG, a self-supervised automatic scene generation technique for matching the distribution of real data. Importantly, Sim2SG does not require supervision from the real-world dataset, thus making it applicable in situations for which such annotations are difficult to obtain. Sim2SG is designed to bridge both the content and appearance gaps, by matching the content of real data, and by matching the features in the source and target domains. We select scene graph (SG) generation as the downstream task, due to the limited availability of labeled datasets. Experiments demonstrate significant improvements over leading baselines in reducing the domain gap both qualitatively and quantitatively, on several synthetic datasets as well as the real-world KITTI dataset.

Supervised deep learning-based hash and vector quantization are enabling fast and large-scale image retrieval systems. By fully exploiting label annotations, they are achieving outstanding retrieval performances compared to the conventional methods. However, it is painstaking to assign labels precisely for a vast amount of training data, and also, the annotation process is error-prone. To tackle these issues, we propose the first deep unsupervised image retrieval method dubbed Self-supervised Product Quantization (SPQ) network, which is label-free and trained in a self-supervised manner. We design a Cross Quantized Contrastive learning strategy that jointly learns codewords and deep visual descriptors by comparing individually transformed images (views). Our method analyzes the image contents to extract descriptive features, allowing us to understand image representations for accurate retrieval. By conducting extensive experiments on benchmarks, we demonstrate that the proposed method yields state-of-the-art results even without supervised pretraining.

Pretraining on large labeled datasets is a prerequisite to achieve good performance in many computer vision tasks like image recognition, video understanding etc. However, pretraining is not widely used for 3D recognition tasks where state-of-the-art methods train models from scratch. A primary reason is the lack of large annotated datasets because 3D data labelling is time-consuming. Recent work shows that self-supervised learning is useful to pretrain models in 3D but requires multi-view data and point correspondences. We present a simple self-supervised pretraining method that can work with single-view depth scans acquired by varied sensors, without 3D registration and point correspondences. We pretrain standard point cloud and voxel based model architectures, and show that joint pretraining further improves performance. We evaluate our models on 9 benchmarks for object detection, semantic segmentation, and object classification, where they achieve state-of-the-art results. Most notably, we set a new state-of-the-art for object detection on ScanNet (69.0% mAP) and SUNRGBD (63.5% mAP). Our pretrained models are label efficient and improve performance for classes with few examples.

We present SSOD -- the first end-to-end analysis-by-synthesis framework with controllable GANs for the task of self-supervised object detection. We use collections of real-world images without bounding box annotations to learn to synthesize and detect objects. We leverage controllable GANs to synthesize images with pre-defined object properties and use them to train object detectors. We propose a tight end-to-end coupling of the synthesis and detection networks to optimally train our system. Finally, we also propose a method to optimally adapt SSOD to an intended target data without requiring labels for it. For the task of car detection, on the challenging KITTI and Cityscapes datasets, we show that SSOD outperforms the prior state-of-the-art purely image-based self-supervised object detection method Wetectron. Even without requiring any 3DCAD assets, it also surpasses the state-of-the-art rendering-based method Meta-Sim2. Our work advances the field of self-supervised object detection by introducing a successful new paradigm of using controllable GAN-based image synthesis for it and by significantly improving the base-line accuracy of the task. We open-source our code athttps://github.com/NVlabs/SSOD.

We consider using untrained neural networks to solve the reconstruction problem of snapshot compressive imaging (SCI), which uses a two-dimensional (2D) detector to capture a high-dimensional (usually 3D) data-cube in a compressed manner. Various SCI systems have been built in recent years to capture data such as high-speed videos, hyperspectral images, and the state-of-the-art reconstruction is obtained by the deep neural networks. However, most of these networks are trained in an end-to-end manner by a large amount of corpus with sometimes simulated ground truth, measurement pairs. In this paper, inspired by the untrained neural networks such as deep image priors (DIP) and deep decoders, we develop a framework by integrating DIP into the plug-and-play regime, leading to a self-supervised network for spectral SCI reconstruction. Extensive synthetic and real data results show that the proposed algorithm without training is capable of achieving competitive results to the training based networks. Furthermore, by integrating the proposed method with a pre-trained deep denoising prior, we have achieved higher performance than existing state-of-the-art.

Remarkable results have been achieved by DCNN based self-supervised depth estimation approaches. However, most of these approaches can only handle either day-time or night-time images, while their performance degrades for all-day images due to large domain shift and the variation of illumination between day and night images. To relieve these limitations, we propose a domain-separated network for self-supervised depth estimation of all-day images. Specifically, to relieve the negative influence of disturbing terms (illumination, etc.), we partition the information of day and night image pairs into two complementary sub-spaces: private and invariant domains, where the former contains the unique information (illumination, etc.) of day and night images and the latter contains essential shared information (texture, etc.). Meanwhile, to guarantee that the day and night images contain the same information, the domain-separated network takes the day-time images and corresponding night-time images (generated by GAN) as input, and the private and invariant feature extractors are learned by orthogonality and similarity loss, where the domain gap can be alleviated, thus better depth maps can be expected. Meanwhile, the reconstruction and photometric losses are utilized to estimate complementary information and depth maps effectively. Experimental results demonstrate that our approach achieves state-of-the-art depth estimation results for all-day images on the challenging Oxford RobotCar dataset, proving the superiority of our proposed approach.

The lack of clean images undermines the practicability of supervised image prior learning methods, of which the training schemes require a large number of clean images. To free image prior learning from the image collection burden, a novel Self-Supervised learning method for Gaussian Mixture Model (SS-GMM) is proposed in this paper. It can simultaneously achieve the noise level estimation and the image prior learning directly from only a single noisy image. This work is derived from our study on eigenvalues of the GMM's covariance matrix. Through statistical experiments and theoretical analysis, we conclude that (1) covariance eigenvalues for clean images hold the sparsity; and that (2) those for noisy images contain sufficient information for noise estimation. The first conclusion inspires us to impose a sparsity constraint on covariance eigenvalues during the learning process to suppress the influence of noise. The second conclusion leads to a self-contained noise estimation module of high accuracy in our proposed method. This module serves to estimate the noise level and automatically determine the specific level of the sparsity constraint. Our final derived method requires only minor modifications to the standard expectation-maximization algorithm. This makes it easy to implement. Very interestingly, the GMM learned via our proposed self-supervised learning method can even achieve better image denoising performance than its supervised counterpart, i.e., the EPLL. Also, it is on par with the state-of-the-art self-supervised deep learning method, i.e., the Self2Self. Code is available at https://github.com/HUST-Tan/SS-GMM.

To learn distinguishable patterns, most of recent works in vehicle re-identification (ReID) struggled to redevelop official benchmarks to provide various supervisions, which requires prohibitive human labors. In this paper, we seek to achieve the similar goal but do not involve more human efforts. To this end, we introduce a novel framework, which successfully encodes both geometric local features and global representations to distinguish vehicle instances, optimized only by the supervision from official ID labels. Specifically, given our insight that objects in ReID share similar geometric characteristics, we propose to borrow self-supervised representation learning to facilitate geometric features discovery. To condense these features, we introduce an interpretable attention module, with the core of local maxima aggregation instead of fully automatic learning, whose mechanism is completely understandable and whose response map is physically reasonable. To the best of our knowledge, we are the first that perform self-supervised learning to discover geometric features. We conduct comprehensive experiments on three most popular datasets for vehicle ReID, i.e., VeRi-776, CityFlow-ReID, and VehicleID. We report our state-of-the-art (SOTA) performances and promising visualization results. We also show the excellent scalability of our approach on other ReID related tasks, i.e., person ReID and multi-target multi-camera (MTMC) vehicle tracking.

With wide applications of image editing tools, forged images (splicing, copy-move, removal and etc.) have been becoming great public concerns. Although existing image forgery localization methods could achieve fairly good results on several public datasets, most of them perform poorly when the forged images are JPEG compressed as they are usually done in social networks. To tackle this issue, in this paper, a self-supervised domain adaptation network, which is composed of a backbone network with Siamese architecture and a compression approximation network (ComNet), is proposed for JPEG-resistant image forgery localization. To improve the performance against JPEG compression, ComNet is customized to approximate the JPEG compression operation through self-supervised learning, generating JPEG-agent images with general JPEG compression characteristics. The backbone network is then trained with domain adaptation strategy to localize the tampering boundary and region, and alleviate the domain shift between uncompressed and JPEG-agent images. Extensive experimental results on several public datasets show that the proposed method outperforms or rivals to other state-of-the-art methods in image forgery localization, especially for JPEG compression with unknown QFs.

Cryo-Electron Tomography (cryo-ET) is a powerful tool for 3D cellular visualization. Due to instrumental limitations, cryo-ET images and their volumetric reconstruction suffer from extremely low signal-to-noise ratio. In this paper, we propose a novel end-to-end self-supervised learning model, the Sparsity Constrained Network (SC-Net), to restore volumetric image from single noisy data in cryo-ET. The proposed method only requires a single noisy data as training input and no ground-truth is needed in the whole training procedure. A new target function is proposed to preserve both local smoothness and detailed structure. Additionally, a novel procedure for the simulation of electron tomographic photographing is designed to help the evaluation of methods. Experiments are done on three simulated data and four real-world data. The results show that our method could produce a strong enhancement for a single very noisy cryo-ET volumetric data, which is much better than the state-of-the-art Noise2Void, and with a competitive performance comparing with Noise2Noise. Code is available at https://github.com/icthrm/SC-Net.

Recently, self-supervised learning (SSL) has been proved very effective and it can help boost the performance in learning representations from unlabeled data in the image domain. Yet, very little is explored about its usefulness in 3D skeleton-based action recognition understanding. Directly applying existing SSL techniques for 3D skeleton learning, however, suffers from trivial solutions and imprecise representations. To tackle these drawbacks, we consider perceiving the consistency and continuity of motion at different playback speeds are two critical issues. To this end, we propose a novel SSL method to learn the 3D skeleton representation in an efficacious way. Specifically, by constructing a positive clip (speed-changed) and a negative clip (motion-broken) of the sampled action sequence, we encourage the positive pairs closer while pushing the negative pairs to force the network to learn the intrinsic dynamic motion consistency information. Moreover, to enhance the learning features, skeleton interpolation is further exploited to model the continuity of human skeleton data. To validate the effectiveness of the proposed method, extensive experiments are conducted on Kinetics, NTU60, NTU120, and PKUMMD datasets with several alternative network architectures. Experimental evaluations demonstrate the superiority of our approach and through which, we can gain significant performance improvement without using extra labeled data.

Encouraged by the success of contrastive learning on image classification tasks, we propose a new self-supervised method for the structured regression task of 3D hand pose estimation. Contrastive learning makes use of unlabeled data for the purpose of representation learning via a loss formulation that encourages the learned feature representations to be invariant under any image transformation. For 3D hand pose estimation, it too is desirable to have invariance to appearance transformation such as color jitter. However, the task requires equivariance under affine transformations, such as rotation and translation. To address this issue, we propose an equivariant contrastive objective and demonstrate its effectiveness in the context of 3D hand pose estimation. We experimentally investigate the impact of invariant and equivariant contrastive objectives and show that learning equivariant features leads to better representations for the task of 3D hand pose estimation. Furthermore, we show that standard ResNets with sufficient depth, trained on additional unlabeled data, attain improvements of up to 14.5% in PA-EPE on FreiHAND and thus achieves state-of-the-art performance without any task specific, specialized architectures. Code and models are available at https://ait.ethz.ch/projects/2021/PeCLR

We present a conditional estimation (CEST) framework to learn 3D facial parameters from 2D single-view images by self-supervised training from videos. CEST is based on the process of analysis by synthesis, where the 3D facial parameters (shape, reflectance, viewpoint, and illumination) are estimated from the face image, and then recombined to reconstruct the 2D face image. In order to learn semantically meaningful 3D facial parameters without explicit access to their labels, CEST couples the estimation of different 3D facial parameters by taking their statistical dependency into account. Specifically, the estimation of any 3D facial parameter is not only conditioned on the given image, but also on the facial parameters that have already been derived. Moreover, the reflectance symmetry and consistency among the video frames are adopted to improve the disentanglement of facial parameters. Together with a novel strategy for incorporating the reflectance symmetry and consistency, CEST can be efficiently trained with in-the-wild video clips. Both qualitative and quantitative experiments demonstrate the effectiveness of CEST.

In this paper, we seek reasons for the two major failure cases in Semantic Segmentation (SS): 1) missing small objects or minor object parts, and 2) mislabeling minor parts of large objects as wrong classes. We have an interesting finding that Failure-1 is due to the underuse of detailed features and Failure-2 is due to the underuse of visual contexts. To help the model learn a better trade-off, we introduce several Self-Regulation (SR) losses for training SS neural networks. By "self", we mean that the losses are from the model per se without using any additional data or supervision. By applying the SR losses, the deep layer features are regulated by the shallow ones to preserve more details; meanwhile, shallow layer classification logits are regulated by the deep ones to capture more semantics. We conduct extensive experiments on both weakly and fully supervised SS tasks, and the results show that our approach consistently surpasses the baselines. We also validate that SR losses are easy to implement in various state-of-the-art SS models, e.g., SPGNet and OCRNet, incurring little computational overhead during training and none for testing