CVPR 2022 Papers

Our lives can be seen as a complex weaving of activities; we switch from one activity to another, to maximise our achievements or in reaction to demands placed upon us. Observing a video of unscripted daily activities, we parse the video into its constituent activity threads through a process we call unweaving. To accomplish this, we introduce a video representation explicitly capturing activity threads called a thread bank, along with a neural controller capable of detecting goal changes and continuations of past activities, together forming UnweaveNet. We train and evaluate UnweaveNet on sequences from the unscripted egocentric dataset EPIC-KITCHENS. We propose and showcase the efficacy of pretraining UnweaveNet in a self-supervised manner.

Cluster discrimination is an effective pretext task for unsupervised representation learning, which often consists of two phases: clustering and discrimination. Clustering is to assign each instance a pseudo label that will be used to learn representations in discrimination. The main challenge resides in clustering since prevalent clustering methods (e.g., k-means) have to run in a batch mode. Besides, there can be a trivial solution consisting of a dominating cluster. To address these challenges, we first investigate the objective of clustering-based representation learning. Based on this, we propose a novel clustering-based pretext task with online Constrained K-means (CoKe). Compared with the balanced clustering that each cluster has exactly the same size, we only constrain the minimal size of each cluster to flexibly capture the inherent data structure. More importantly, our online assignment method has a theoretical guarantee to approach the global optimum. By decoupling clustering and discrimination, CoKe can achieve competitive performance when optimizing with only a single view from each instance. Extensive experiments on ImageNet and other benchmark data sets verify both the efficacy and efficiency of our proposal.

Understanding realistic visual scene images together with language descriptions is a fundamental task towards generic visual understanding. Previous works have shown compelling comprehensive results by building hierarchical structures for visual scenes (e.g., scene graphs) and natural languages (e.g., dependency trees), individually. However, how to construct a joint vision-language (VL) structure has barely been investigated. More challenging but worthwhile, we introduce a new task that targets on inducing such a joint VL structure in an unsupervised manner. Our goal is to bridge the visual scene graphs and linguistic dependency trees seamlessly. Due to the lack of VL structural data, we start by building a new dataset VLParse. Rather than using labor-intensive labeling from scratch, we propose an automatic alignment procedure to produce coarse structures followed by human refinement to produce high-quality ones. Moreover, we benchmark our dataset by proposing a contrastive learning (CL)-based framework VLGAE, short for Vision-Language Graph Autoencoder. Our model obtains superior performance on two derived tasks, i.e., language grammar induction and VL phrase grounding. Ablations show the effectiveness of both visual cues and dependency relationships on fine-grained VL structure construction.

Self-supervised learning is a promising unsupervised learning framework that has achieved success with large floating point networks. But such networks are not readily deployable to edge devices. To accelerate deployment of models with the benefit of unsupervised representation learning to such resource limited devices for various downstream tasks, we propose a self-supervised learning method for binary networks that uses a moving target network. In particular, we propose to jointly train a randomly initialized classifier, attached to a pretrained floating point feature extractor, with a binary network. Additionally, we propose a feature similarity loss, a dynamic loss balancing and modified multi-stage training to further improve the accuracy, and call our method BURN. Our empirical validations over five downstream tasks using seven datasets show that BURN outperforms self-supervised baselines for binary networks and sometimes outperforms supervised pretraining. Code is availabe at https://github.com/naver-ai/burn.

Unsupervised video representation learning has made remarkable achievements in recent years. However, most existing methods are designed and optimized for video classification. These pre-trained models can be sub-optimal for temporal localization tasks due to the inherent discrepancy between video-level classification and clip-level localization. To bridge this gap, we make the first attempt to propose a self-supervised pretext task, coined as Pseudo Action Localization (PAL) to Unsupervisedly Pre-train feature encoders for Temporal Action Localization tasks (UP-TAL). Specifically, we first randomly select temporal regions, each of which contains multiple clips, from one video as pseudo actions and then paste them onto different temporal positions of the other two videos. The pretext task is to align the features of pasted pseudo action regions from two synthetic videos and maximize the agreement between them. Compared to the existing unsupervised video representation learning approaches, our PAL adapts better to downstream TAL tasks by introducing a temporal equivariant contrastive learning paradigm in a temporally dense and scale-aware manner. Extensive experiments show that PAL can utilize large-scale unlabeled video data to significantly boost the performance of existing TAL methods. Our codes and models will be made publicly available at https://github.com/zhang-can/UP-TAL.

The distributional shift issue between training and test sets is a critical challenge in machine learning, and is aggravated when models capture unintended decision rules with spurious correlations. Although existing works often handle this issue using human supervision, the availability of the proper annotations is impractical and even unrealistic. To better tackle this challenge, we propose a simple but effective debiasing technique in an unsupervised manner. Specifically, we perform clustering on the feature embedding space and identify pseudo-bias-attributes by taking advantage of the clustering results even without an explicit attribute supervision. Then, we employ a novel cluster-based reweighting scheme for learning debiased representation; this prevents minority groups from being ignored for minimizing the overall loss, which is desirable for worst-case generalization. The extensive experiments demonstrate the outstanding performance of our approach on multiple standard benchmarks, which is even as competitive as the supervised method. We plan to release the source code of our work for better reproducibility.

Unsupervised image-to-image translation aims to learn the translation between two visual domains without paired data. Despite the recent progress in image translation models, it remains challenging to build mappings between complex domains with drastic visual discrepancies. In this work, we present a novel framework, Generative Prior-guided UNsupervised Image-to-image Translation (GP-UNIT), to improve the overall quality and applicability of the translation algorithm. Our key insight is to leverage the generative prior from pre-trained class-conditional GANs (e.g., BigGAN) to learn rich content correspondences across various domains. We propose a novel coarse-to-fine scheme: we first distill the generative prior to capture a robust coarse-level content representation that can link objects at an abstract semantic level, based on which fine-level content features are adaptively learned for more accurate multi-level content correspondences. Extensive experiments demonstrate the superiority of our versatile framework over state-of-the-art methods in robust, high-quality and diversified translations, even for challenging and distant domains.

Estimating homography from an image pair is a fundamental problem in image alignment. Unsupervised learning methods have received increasing attention in this field due to their promising performance and label-free training. However, existing methods do not explicitly consider the problem of plane induced parallax, which will make the predicted homography compromised on multiple planes. In this work, we propose a novel method HomoGAN to guide unsupervised homography estimation to focus on the dominant plane. First, a multi-scale transformer network is designed to predict homography from the feature pyramids of input images in a coarse-to-fine fashion. Moreover, we propose an unsupervised GAN to impose coplanarity constraint on the predicted homography, which is realized by using a generator to predict a mask of aligned regions, and then a discriminator to check if two masked feature maps can be induced by a single homography. To validate the effectiveness of HomoGAN and its components, we conduct extensive experiments on a large-scale dataset, and results show that our matching error is 22% lower than the previous SOTA method. Our code will be publicly available.

Previous advances in object tracking mostly reported on favorable illumination circumstances while neglecting performance at nighttime, which significantly impeded the development of related aerial robot applications. This work instead develops a novel unsupervised domain adaptation framework for nighttime aerial tracking (named UDAT). Specifically, a unique object discovery approach is provided to generate training patches from raw nighttime tracking videos. To tackle the domain discrepancy, we employ a Transformer-based bridging layer post to the feature extractor to align image features from both domains. With a Transformer day/night feature discriminator, the daytime tracking model is adversarially trained to track at night. Moreover, we construct a pioneering benchmark namely NAT2021 for unsupervised domain adaptive nighttime tracking, which comprises a test set of 180 manually annotated tracking sequences and a train set of over 276k unlabelled nighttime tracking frames. Exhaustive experiments demonstrate the robustness and domain adaptability of the proposed framework in nighttime aerial tracking. The code and benchmark are available at https://github.com/vision4robotics/UDAT.

Image deraining is a typical low-level image restoration task, which aims at decomposing the rainy image into two distinguishable layers: the clean image layer and the rain layer. Most of the existing learning-based deraining methods are supervisedly trained on synthetic rainy-clean pairs. The domain gap between the synthetic and real rains makes them less generalized to different real rainy scenes. Moreover, the existing methods mainly utilize the property of the two layers independently, while few of them have considered the mutually exclusive relationship between the two layers. In this work, we propose a novel non-local contrastive learning (NLCL) method for unsupervised image deraining. Consequently, we not only utilize the intrinsic self-similarity property within samples but also the mutually exclusive property between the two layers, so as to better differ the rain layer from the clean image. Specifically, the non-local self-similarity image layer patches as the positives are pulled together and similar rain layer patches as the negatives are pushed away. Thus the similar positive/negative samples that are close in the original space benefit us to enrich more discriminative representation. Apart from the self-similarity sampling strategy, we analyze how to choose an appropriate feature encoder in NLCL. Extensive experiments on different real rainy datasets demonstrate that the proposed method obtains state-of-the-art performance in real deraining.

We present a novel approach for unsupervised activity segmentation which uses video frame clustering as a pretext task and simultaneously performs representation learning and online clustering. This is in contrast with prior works where representation learning and clustering are often performed sequentially. We leverage temporal information in videos by employing temporal optimal transport. In particular, we incorporate a temporal regularization term which preserves the temporal order of the activity into the standard optimal transport module for computing pseudo-label cluster assignments. The temporal optimal transport module enables our approach to learn effective representations for unsupervised activity segmentation. Furthermore, previous methods require storing learned features for the entire dataset before clustering them in an offline manner, whereas our approach processes one mini-batch at a time in an online manner. Extensive evaluations on three public datasets, i.e. 50-Salads, YouTube Instructions, and Breakfast, and our dataset, i.e., Desktop Assembly, show that our approach performs on par with or better than previous methods, despite having significantly less memory constraints.

Recent progress towards learning from limited supervision has encouraged efforts towards designing models that can recognize novel classes at test time (generalized zero-shot learning or GZSL). GZSL approaches assume knowledge of all classes, with or without labeled data, before-hand. However, practical scenarios demand models that are adaptable and can handle dynamic addition of new seen and unseen classes on the fly (i.e continual generalized zero-shot learning or CGZSL). One solution is to sequentially retrain and reuse conventional GZSL methods, however, such an approach suffers from catastrophic forgetting leading to suboptimal generalization performance. A few recent efforts towards tackling CGZSL have been limited by difference in settings, practicality, data splits and protocols followed - inhibiting fair comparison and a clear direction forward. Motivated from these observations, in this work, we firstly consolidate the different CGZSL setting variants and propose a new Online CGZSL setting which is more practical and flexible. Secondly, we introduce a unified feature-generative framework for CGZSL that leverages bi-directional incremental alignment to dynamically adapt to addition of new classes with or without labeled data that arrive over time in any of these CGZSL settings. Our comprehensive experiments and analysis on five benchmark datasets and comparison with baselines show that our approach consistently outperforms existing methods, especially on the more practical Online setting.

Learning single image deraining (SID) networks from an unpaired set of clean and rainy images is practical and valuable as acquiring paired real-world data is almost infeasible. However, without the paired data as the supervision, learning a SID network is challenging. Moreover, simply using existing unpaired learning methods (e.g., unpaired adversarial learning and cycle-consistency constraints) in the SID task is insufficient to learn the underlying relationship from rainy inputs to clean outputs as there exists significant domain gap between the rainy and clean images. In this paper, we develop an effective unpaired SID adversarial framework which explores mutual properties of the unpaired exemplars by a dual contrastive learning manner in a deep feature space, named as DCD-GAN. The proposed method mainly consists of two cooperative branches: Bidirectional Translation Branch (BTB) and Contrastive Guidance Branch (CGB). Specifically, BTB exploits full advantage of the circulatory architecture of adversarial consistency to generate abundant exemplar pairs and excavates latent feature distributions between two domains by equipping it with bidirectional mapping. Simultaneously, CGB implicitly constrains the embeddings of different exemplars in the deep feature space by encouraging the similar feature distributions closer while pushing the dissimilar further away, in order to better facilitate rain removal and help image restoration. Extensive experiments demonstrate that our method performs favorably against existing unpaired deraining approaches on both synthetic and real-world datasets, and generates comparable results against several fully-supervised or semi-supervised models.

In this paper, we present a general-purpose solution to cartoon image synthesis with unpaired training data. In contrast to previous works learning pre-defined cartoon styles for specified usage scenarios (portrait or scene), we aim to train a common cartoon translator which can not only simultaneously render exaggerated anime faces and realistic cartoon scenes, but also provide flexible user controls for desired cartoon styles. It is challenging due to the complexity of the task and the absence of paired data. The core idea of the proposed method is to introduce gated cycle mapping, that utilizes a novel gated mapping unit to produce the category-specific style code and embeds this code into cycle networks to control the translation process. For the concept of category, we classify images into different categories (e.g., 4 types: photo/cartoon portrait/scene) and learn finer-grained category translations rather than overall mappings between two domains (e.g., photo and cartoon). Furthermore, the proposed method can be easily extended to cartoon video generation with an auxiliary dataset and a new adaptive style loss. Experimental results demonstrate the superiority of the proposed method over the state of the art and validate its effectiveness in the brand-new task of general cartoon image synthesis.

Existing person re-identification (ReID) methods typically directly load the pre-trained ImageNet weights for initialization. However, as a fine-grained classification task, ReID is more challenging and exists a large domain gap between ImageNet classification. Inspired by the great success of self-supervised representation learning with contrastive objectives, in this paper, we design an Unsupervised Pre-training framework for ReID based on the contrastive learning (CL) pipeline, dubbed UP-ReID. During the pre-training, we attempt to address two critical issues for learning fine-grained ReID features: (1) the augmentations in CL pipeline may distort the discriminative clues in person images. (2) the fine-grained local features of person images are not fully-explored. Therefore, we introduce an (I^2-)regularization in the UP-ReID, which is instantiated as two constraints coming from global image aspect and local patch aspect: a global consistency is enforced between augmented and original person images to increase robustness to augmentation, while an intrinsic contrastive constraint among local patches of each image is employed to fully explore the local discriminative clues. Extensive experiments on multiple popular Re-ID datasets, including PersonX, Market1501, CUHK03, and MSMT17, demonstrate that our UP-ReID pre-trained model can significantly benefit the downstream ReID fine-tuning and achieve state-of-the-art performance.

This paper tackles a new photometric stereo task, named universal photometric stereo. Unlike existing tasks that assumed specific physical lighting models; hence, drastically limited their usability, a solution algorithm of this task is supposed to work for objects with diverse shapes and materials under arbitrary lighting variations without assuming any specific models. To solve this extremely challenging task, we present a purely data-driven method, which eliminates the prior assumption of lighting by replacing the recovery of physical lighting parameters with the extraction of the generic lighting representation, named global lighting contexts. We use them like lighting parameters in a calibrated photometric stereo network to recover surface normal vectors pixelwisely. To adapt our network to a wide variety of shapes, materials and lightings, it is trained on a new synthetic dataset which simulates the appearance of objects in the wild. Our method is compared with other state-of-the-art uncalibrated photometric stereo methods on our test data to demonstrate the significance of our method.

Learning from a label distribution has achieved promising results on ordinal regression tasks such as facial age and head pose estimation wherein, the concept of adaptive label distribution learning (ALDL) has drawn lots of attention recently for its superiority in theory. However, compared with the methods assuming fixed form label distribution, ALDL methods have not achieved better performance. We argue that existing ALDL algorithms do not fully exploit the intrinsic properties of ordinal regression. In this paper, we emphatically summarize that learning an adaptive label distribution on ordinal regression tasks should follow three principles. First, the probability corresponding to the ground-truth should be the highest in label distribution. Second, the probabilities of neighboring labels should decrease with the increase of distance away from the ground-truth, i.e., the distribution is unimodal. Third, the label distribution should vary with samples changing, and even be distinct for different instances with the same label, due to the different levels of difficulty and ambiguity. Under the premise of these principles, we propose a novel loss function for fully adaptive label distribution learning, namely unimodal-concentrated loss. Specifically, the unimodal loss derived from the learning to rank strategy constrains the distribution to be unimodal. Furthermore, the estimation error and the variance of the predicted distribution for a specific sample are integrated into the proposed concentrated loss to make the predicted distribution maximize at the ground-truth and vary according to the predicting uncertainty. Extensive experimental results on typical ordinal regression tasks including age and head pose estimation, show the superiority of our proposed unimodal-concentrated loss compared with existing loss functions.

This paper aims to improve panoptic segmentation for real-world applications in three ways. First, we present a label policy that unifies four of the most popular panoptic segmentation datasets for autonomous driving. We also clean up label confusion by adding the new vehicle labels pickup and van. Full relabeling information for the popular Mapillary Vistas, IDD, and Cityscapes dataset are provided to add these new labels to existing setups. Second, we introduce Wilddash2 (WD2), a new dataset and public benchmark service for panoptic segmentation. The dataset consists of more than 5000 unique driving scenes from all over the world with a focus on visually challenging scenes, such as diverse weather conditions, lighting situations, and camera characteristics. We showcase experimental visual hazard classifiers which help to pre-filter challenging frames during dataset creation. Finally, to characterize the robustness of algorithms in out-of-distribution situations, we introduce hazard-aware and negative testing for panoptic segmentation as well as statistical significance calculations that increase confidence for both concepts. Additionally, we present a novel technique for visualizing panoptic segmentation errors. Our experiments show the negative impact of visual hazards on panoptic segmentation quality. Additional data from the WD2 dataset improves performance for visually challenging scenes and thus robustness in real-world scenarios.

Slow shutter speed and long exposure time of frame-based cameras often cause visual blur and loss of inter-frame information, degenerating the overall quality of captured videos. To this end, we present a unified framework of event-based motion deblurring and frame interpolation for blurry video enhancement, where the extremely low latency of events is leveraged to alleviate motion blur and facilitate intermediate frame prediction. Specifically, the mapping relation between blurry frames and sharp latent images is first predicted by a learnable double integral network, and a fusion network is then proposed to refine the coarse results via utilizing the information from consecutive blurry inputs and the concurrent events. By exploring the mutual constraints among blurry frames, latent images, and event streams, we further propose a self-supervised learning framework to enable network training with real-world blurry videos and events. Extensive experiments demonstrate that our method compares favorably against the state-of-the-art approaches and achieves remarkable performance on both synthetic and real-world datasets.