CVPR 2021 论文列表

Unsupervised Domain Adaptation (UDA) aims to learn a predictor model for an unlabeled dataset by transferring knowledge from a labeled source data, which has been trained on similar tasks. However, most of these conventional UDA approaches have a strong assumption of having access to the source data during training, which may not be very practical due to privacy, security and storage concerns. A recent line of work addressed this problem and proposed an algorithm that transfers knowledge to the unlabeled target domain only from a single learned source model without requiring access to the source data. However, for adaptation purpose, if there are multiple trained source models available to choose from, this method has to go through adapting each and every model individually, to check for the best source. Thus, we ask the question: can we find the optimal combination of source models, with no source data and without target labels, whose performance is no worse than the single best source? To answer this, we propose a novel and efficient algorithm which automatically combines the source models with suitable weights in such a way that it performs at least as good as the best source model. We provide intuitive theoretical insights to justify our claim. Moreover, extensive experiments are conducted on several benchmark datasets to show the effectiveness of our algorithm, where in most cases, our method not only reaches best source accuracy but also outperform it.

Understanding the 3D world from 2D projected natural images is a fundamental challenge in computer vision and graphics. Recently, an unsupervised learning approach has garnered considerable attention owing to its advantages in data collection. However, to mitigate training limitations, typical methods need to impose assumptions for viewpoint distribution (e.g., a dataset containing various viewpoint images) or object shape (e.g., symmetric objects). These assumptions often restrict applications; for instance, the application to non-rigid objects or images captured from similar viewpoints (e.g., flower or bird images) remains a challenge. To complement these approaches, we propose aperture rendering generative adversarial networks (AR-GANs), which equip aperture rendering on top of GANs, and adopt focus cues to learn the depth and depth-of-field (DoF) effect of unlabeled natural images. To address the ambiguities triggered by unsupervised setting (i.e., ambiguities between smooth texture and out-of-focus blurs, and between foreground and background blurs), we develop DoF mixture learning, which enables the generator to learn real image distribution while generating diverse DoF images. In addition, we devise a center focus prior to guiding the learning direction. In the experiments, we demonstrate the effectiveness of AR-GANs in various datasets, such as flower, bird, and face images, demonstrate their portability by incorporating them into other 3D representation learning GANs, and validate their applicability in shallow DoF rendering.

Recently, numerous works have attempted to learn 3D reconstructors of textured 3D models of visual categories given a training set of annotated static images of objects. In this paper, we seek to decrease the amount of needed supervision by leveraging a collection of object-centric videos captured in-the-wild without requiring any manual 3D annotations. Since existing category-centric datasets are insufficient for this problem, we contribute with a large-scale crowd-sourced dataset of object-centric videos suitable for this task. We further propose a novel method that learns via differentiable rendering of a predicted implicit surface of the scene. Here, inspired by classic multi-view stereo methods, our key technical contribution is a novel warp-conditioned implicit shape function, which is robust to the noise in the SfM video reconstructions that supervise our learning. Our evaluation demonstrates performance improvements over several deep monocular reconstruction baselines on 2 existing benchmarks and on our novel dataset.

Robust model fitting is a core algorithm in a large number of computer vision applications. Solving this problem efficiently for highly contaminated datasets is, however, still challenging due to its underlying computational complexity. Recent attention has been focused on learning-based algorithms. However, most approaches are supervised (which require a large amount of labelled training data). In this paper, we introduce a novel unsupervised learning framework that learns to directly solve robust model fitting. Unlike other methods, our work is agnostic to the underlying input features, and can be easily generalized to a wide variety of LP-type problems with quasi-convex residuals. We empirically show that our method outperforms existing unsupervised learning approaches, and achieves competitive results compared to traditional methods on several important computer vision problems.

Most of the existing literature regarding hyperbolic embedding concentrate upon supervised learning, whereas the use of unsupervised hyperbolic embedding is less well explored. In this paper, we analyze how unsupervised tasks can benefit from learned representations in hyperbolic space. To explore how well the hierarchical structure of unlabeled data can be represented in hyperbolic spaces, we design a novel hyperbolic message passing auto-encoder whose overall auto-encoding is performed in hyperbolic space. The proposed model conducts auto-encoding the networks via fully utilizing hyperbolic geometry in message passing. Through extensive quantitative and qualitative analyses, we validate the properties and benefits of the unsupervised hyperbolic representations. Codes are available at https://github.com/junhocho/HGCAE.

Learning feature embedding directly from images without any human supervision is a very challenging and essential task in the field of computer vision and machine learning. Following the paradigm in supervised manner, most existing unsupervised metric learning approaches mainly focus on binary similarity in Euclidean space. However, these methods cannot achieve promising performance in many practical applications, where the manual information is lacking and data exhibits non-Euclidean latent anatomy. To address this limitation, we propose an Unsupervised Hyperbolic Metric Learning method with Hierarchical Similarity. It considers the natural hierarchies of data by taking advantage of Hyperbolic metric learning and hierarchical clustering, which can effectively excavate richer similarity information beyond binary in modeling. More importantly, we design a new loss function to capture the hierarchical similarity among samples to enhance the stability of the proposed method. Extensive experimental results on benchmark datasets demonstrate that our method achieves state-of-the-art performance compared with current unsupervised deep metric learning approaches.

Human pose estimation is a major computer vision problem with applications ranging from augmented reality and video capture to surveillance and movement tracking. In the medical context, the latter may be an important biomarker for neurological impairments in infants. Whilst many methods exist, their application has been limited by the need for well annotated large datasets and the inability to generalize to humans of different shapes and body compositions, e.g. children and infants. In this paper we present a novel method for learning pose estimators for human adults and infants in an unsupervised fashion. We approach this as a learnable template matching problem facilitated by deep feature extractors. Human-interpretable landmarks are estimated by transforming a template consisting of predefined body parts that are characterized by 2D Gaussian distributions. Enforcing a connectivity prior guides our model to meaningful human shape representations. We demonstrate the effectiveness of our approach on two different datasets including adults and infants.

Unsupervised feature learning has made great strides with contrastive learning based on instance discrimination and invariant mapping, as benchmarked on curated class-balanced datasets. However, natural data could be highly correlated and long-tail distributed. Natural between-instance similarity conflicts with the presumed instance distinction, causing unstable training and poor performance. Our idea is to discover and integrate between-instance similarity into contrastive learning, not directly by instance grouping, but by cross-level discrimination (CLD) between instances and local instance groups. While invariant mapping of each instance is imposed by attraction within its augmented views, between-instance similarity emerges from common repulsion against instance groups. Our batch-wise and cross-view comparisons also greatly improve the positive/negative sample ratio of contrastive learning and achieve better invariant mapping. To effect both grouping and discrimination objectives, we impose them on features separately derived from a shared representation. In addition, we propose normalized projection heads and unsupervised hyper-parameter tuning for the first time. Our extensive experimentation demonstrates that CLD is a lean and powerful add-on to existing methods (e.g., NPID, MoCo, InfoMin, BYOL) on highly correlated, long-tail, or balanced datasets. It not only achieves new state-of-the-art on self-supervision, semi-supervision, and transfer learning benchmarks, but also beats MoCo v2 and SimCLR on every reported performance attained with a much larger compute. CLD effectively extends unsupervised learning to natural data and brings it closer to real-world applications.

We propose a method to disentangle linear-encoded facial semantics from StyleGAN without external supervision. The method derives from linear regression and sparse representation learning concepts to make the disentangled latent representations easily interpreted as well. We start by coupling StyleGAN with a stabilized 3D deformable facial reconstruction method to decompose single-view GAN generations into multiple semantics. Latent representations are then extracted to capture interpretable facial semantics. In this work, we make it possible to get rid of labels for disentangling meaningful facial semantics. Also, we demonstrate that the guided extrapolation along the disentangled representations can help with data augmentation, which sheds light on handling unbalanced data. Finally, we provide an analysis of our learned localized facial representations and illustrate that the semantic information is encoded, which surprisingly complies with human intuition. The overall unsupervised design brings more flexibility to representation learning in the wild.

Instance segmentation is an active topic in computer vision that is usually solved by using supervised learning approaches over very large datasets composed of object level masks. Obtaining such a dataset for any new domain can be very expensive and time-consuming. In addition, models trained on certain annotated categories do not generalize well to unseen objects. The goal of this paper is to propose a method that can perform unsupervised discovery of long-tail categories in instance segmentation, through learning instance embeddings of masked regions. Leveraging rich relationship and hierarchical structure between objects in the images, we propose self-supervised losses for learning mask embeddings. Trained on COCO dataset without additional annotations of the long-tail objects, our model is able to discover novel and more fine-grained objects than the common categories in COCO. We show that the model achieves competitive quantitative results on LVIS as compared to the supervised and partially supervised methods.

Most existing CNN-based super-resolution (SR) methods are developed based on an assumption that the degradation is fixed and known (e.g., bicubic downsampling). However, these methods suffer a severe performance drop when the real degradation is different from their assumption. To handle various unknown degradations in real-world applications, previous methods rely on degradation estimation to reconstruct the SR image. Nevertheless, degradation estimation methods are usually time-consuming and may lead to SR failure due to large estimation errors. In this paper, we propose an unsupervised degradation representation learning scheme for blind SR without explicit degradation estimation. Specifically, we learn abstract representations to distinguish various degradations in the representation space rather than explicit estimation in the pixel space. Moreover, we introduce a Degradation-Aware SR (DASR) network with flexible adaption to various degradations based on the learned representations. It is demonstrated that our degradation representation learning scheme can extract discriminative representations to obtain accurate degradation information. Experiments on both synthetic and real images show that our network achieves state-of-the-art performance for the blind SR task. Code is available at: https://github.com/LongguangWang/DASR.

Most 3D shape completion approaches rely heavily on partial-complete shape pairs and learn in a fully supervised manner. Despite their impressive performances on in-domain data, when generalizing to partial shapes in other forms or real-world partial scans, they often obtain unsatisfactory results due to domain gaps. In contrast to previous fully supervised approaches, in this paper we present ShapeInversion, which introduces Generative Adversarial Network (GAN) inversion to shape completion for the first time. ShapeInversion uses a GAN pre-trained on complete shapes by searching for a latent code that gives a complete shape that best reconstructs the given partial input. In this way, ShapeInversion no longer needs paired training data, and is capable of incorporating the rich prior captured in a well-trained generative model. On the ShapeNet benchmark, the proposed ShapeInversion outperforms the SOTA unsupervised method, and is comparable with supervised methods that are learned using paired data. It also demonstrates remarkable generalization ability, giving robust results for real-world scans and partial inputs of various forms and incompleteness levels. Importantly, ShapeInversion naturally enables a series of additional abilities thanks to the involvement of a pre-trained GAN, such as producing multiple valid complete shapes for an ambiguous partial input, as well as shape manipulation and interpolation.

The main difficulty of person re-identification (ReID) lies in collecting annotated data and transferring the model across different domains. This paper presents UnrealPerson, a novel pipeline that makes full use of unreal image data to decrease the costs in both the training and deployment stages. Its fundamental part is a system that can generate synthesized images of high-quality and from controllable distributions. Instance-level annotation goes with the synthesized data and is almost free. We point out some details in image synthesis that largely impact the data quality. With 3,000 IDs and 120,000 instances, our method achieves a 38.5% rank-1 accuracy when being directly transferred to MSMT17. It almost doubles the former record using synthesized data and even surpasses previous direct transfer records using real data. This offers a good basis for unsupervised domain adaption, where our pre-trained model is easily plugged into the state-of-the-art algorithms towards higher accuracy. In addition, the data distribution can be flexibly adjusted to fit some corner ReID scenarios, which widens the application of our pipeline. We publish our data synthesis toolkit and synthesized data in https://github.com/FlyHighest/UnrealPerson.

Image-to-image translation aims to preserve source contents while translating to discriminative target styles between two visual domains. Most works apply adversarial learning in the ambient image space, which could be computationally expensive and challenging to train. In this paper, we propose to deploy an energy-based model (EBM) in the latent space of a pretrained autoencoder for this task. The pretrained autoencoder serves as both a latent code extractor and an image reconstruction worker. Our model, LETIT, is based on the assumption that two domains share the same latent space, where latent representation is implicitly decomposed as a content code and a domain-specific style code. Instead of explicitly extracting the two codes and applying adaptive instance normalization to combine them, our latent EBM can implicitly learn to transport the source style code to the target style code while preserving the content code, an advantage over existing image translation methods. This simplified solution is also more efficient in the one-sided unpaired image translation setting. Qualitative and quantitative comparisons demonstrate superior translation quality and faithfulness for content preservation. Our model is the first to be applicable to 1024x1024-resolution unpaired image translation to the best of our knowledge. Code is available at https://github.com/YangNaruto/latent-energy-transport.

Machine learning models are known to be vulnerable to adversarial attacks, namely perturbations of the data that lead to wrong predictions despite being imperceptible. However, the existence of "universal" attacks (i.e., unique perturbations that transfer across different data points) has only been demonstrated for images to date. Part of the reason lies in the lack of a common domain, for geometric data such as graphs, meshes, and point clouds, where a universal perturbation can be defined. In this paper, we offer a change in perspective and demonstrate the existence of universal attacks for geometric data (shapes). We introduce a computational procedure that operates entirely in the spectral domain, where the attacks take the form of small perturbations to short eigenvalue sequences; the resulting geometry is then synthesized via shape-from-spectrum recovery. Our attacks are universal, in that they transfer across different shapes, different representations (meshes and point clouds), and generalize to previously unseen data.

Methods for object detection and segmentation rely on large scale instance-level annotations for training, which are difficult and time-consuming to collect. Efforts to alleviate this look at varying degrees and quality of supervision. Weakly-supervised approaches draw on image-level labels to build detectors/segmentors, while zero/few-shot methods assume abundant instance-level data for a set of base classes, and none to a few examples for novel classes. This taxonomy has largely siloed algorithmic designs. In this work, we aim to bridge this divide by proposing an intuitive and unified semi-supervised model that is applicable to a range of supervision: from zero to a few instance-level samples per novel class. For base classes, our model learns a mapping from weakly-supervised to fully-supervised detectors/segmentors. By learning and leveraging visual and lingual similarities between the novel and base classes, we transfer those mappings to obtain detectors/segmentors for novel classes; refining them with a few novel class instance-level annotated samples, if available. The overall model is end-to-end trainable and highly flexible. Through extensive experiments on MS-COCO and Pascal VOC benchmark datasets we show improved performance in a variety of settings.

Action recognition has been heavily employed in many applications such as autonomous vehicles, surveillance, etc, where its robustness is a primary concern. In this paper, we examine the robustness of state-of-the-art action recognizers against adversarial attack, which has been rarely investigated so far. To this end, we propose a new method to attack action recognizers which rely on the 3D skeletal motion. Our method involves an innovative perceptual loss which ensures the imperceptibility of the attack. Empirical studies demonstrate that our method is effective in both white-box and black-box scenarios. Its generalizability is evidenced on a variety of action recognizers and datasets. Its versatility is shown in different attacking strategies. Its deceitfulness is proven in extensive perceptual studies. Our method shows that adversarial attack on 3D skeletal motions, one type of time-series data, is significantly different from traditional adversarial attack problems. Its success raises serious concern on the robustness of action recognizers and provides insights on potential improvements.

Unsupervised contrastive learning has achieved outstanding success, while the mechanism of contrastive loss has been less studied. In this paper, we concentrate on the understanding of the behaviours of unsupervised contrastive loss. We will show that the contrastive loss is a hardness-aware loss function, and the temperature t controls the strength of penalties on hard negative samples. The previous study has shown that uniformity is a key property of contrastive learning. We build relations between the uniformity and the temperature t. We will show that uniformity helps the contrastive learning to learn separable features, however excessive pursuit to the uniformity makes the contrastive loss not tolerant to semantically similar samples, which may break the underlying semantic structure and be harmful to the formation of features useful for downstream tasks. This is caused by the inherent defect of the instance discrimination objective. Specifically, instance discrimination objective tries to push all different instances apart, ignoring the underlying relations between samples. Pushing semantically consistent samples apart has no positive effect for acquiring a prior informative to general downstream tasks. A well-designed contrastive loss should have some extents of tolerance to the closeness of semantically similar samples. Therefore, we find that the contrastive loss meets a uniformity-tolerance dilemma, and a good choice of temperature can compromise these two properties properly to both learn separable features and tolerant to semantically similar samples, improving the feature qualities and the downstream performances.

Perceptual metrics based on features of deep Convolutional Neural Networks (CNNs) have shown remarkable success when used as loss functions in a range of computer vision problems and significantly outperform classical losses such as L1 or L2 in pixel space. The source of this success remains somewhat mysterious, especially since a good loss does not require a particular CNN architecture nor a particular training method. In this paper we show that similar success can be achieved even with losses based on features of a deep CNN with random filters. We use the tool of infinite CNNs to derive an analytical form for perceptual similarity in such CNNs, and prove that the perceptual distance between two images is equivalent to the maximum mean discrepancy (MMD) distance between local distributions of small patches in the two images. We use this equivalence to propose a simple metric for comparing two images which directly computes the MMD between local distributions of patches in the two images. Our proposed metric is simple to understand, requires no deep networks, and gives comparable performance to perceptual metrics in a range of computer vision tasks.