CVPR 2022 论文列表

The mainstream paradigm behind continual learning has been to adapt the model parameters to non-stationary data distributions, where catastrophic forgetting is the central challenge. Typical methods rely on a rehearsal buffer or known task identity at test time to retrieve learned knowledge and address forgetting, while this work presents a new paradigm for continual learning that aims to train a more succinct memory system without accessing task identity at test time. Our method learns to dynamically prompt (L2P) a pre-trained model to learn tasks sequentially under different task transitions. In our proposed framework, prompts are small learnable parameters, which are maintained in a memory space. The objective is to optimize prompts to instruct the model prediction and explicitly manage task-invariant and task-specific knowledge while maintaining model plasticity. We conduct comprehensive experiments under popular image classification benchmarks with different challenging continual learning settings, where L2P consistently outperforms prior state-of-the-art methods. Surprisingly, L2P achieves competitive results against rehearsal-based methods even without a rehearsal buffer and is directly applicable to challenging task-agnostic continual learning. Source code is available at https://github.com/google-research/l2p.

This paper studies the task of One-Shot image Generation (OSG), where generation network learned on base dataset should be generalizable to synthesize images of novel categories with only one available sample per novel category. Most existing methods for feature transfer in one-shot image generation only learn reusable features implicitly on pre-training tasks. Such methods would be likely to overfit pre-training tasks. In this paper, we propose a novel model to explicitly learn and memorize reusable features that can help hallucinate novel category images. To be specific, our algorithm learns to decompose image features into the Category-Related (CR) and Category-Independent (CI) features. Our model learning to memorize class-independent CI features which are further utilized by our feature hallucination component to generate target novel category images. We validate our model on several benchmarks. Extensive experiments demonstrate that our model effectively boosts the OSG performance and can generate compelling and diverse samples.

We present a framework for modeling interactional communication in dyadic conversations: given multimodal inputs of a speaker, we autoregressively output multiple possibilities of corresponding listener motion. We combine the motion and speech audio of the speaker using a motion-audio cross attention transformer. Furthermore, we enable non-deterministic prediction by learning a discrete latent representation of realistic listener motion with a novel motion-encoding VQ-VAE. Our method organically captures the multimodal and non-deterministic nature of nonverbal dyadic interactions. Moreover, it produces realistic 3D listener facial motion synchronous with the speaker (see video). We demonstrate that our method outperforms baselines qualitatively and quantitatively via a rich suite of experiments. To facilitate this line of research, we introduce a novel and large in-the-wild dataset of dyadic conversations. Code, data, and videos available at http://people.eecs.berkeley.edu/ evonne_ng/projects/learning2listen/.

Meta-learning enables models to adapt to new environments rapidly with a few training examples. Current gradient-based meta-learning methods concentrate on finding good initialization (meta-weights) for learners but ignore the impact of neural architectures. In this paper, we aim to obtain better meta-learners by co-optimizing the architecture and meta-weights simultaneously. Existing NAS-based meta-learning methods apply a two-stage strategy, i.e., first searching architectures and then re-training meta-weights on the searched architecture. However, this two-stage strategy would break the mutual impact of the architecture and meta-weights since they are optimized separately. Differently, we propose progressive connection consolidation, fixing the architecture layer by layer, in which the layer with the largest weight value would be fixed first. In this way, we can jointly search architectures and train the meta-weights on fixed layers. Besides, to improve the generalization performance of the searched meta-learner on all tasks, we propose a more effective rule for co-optimization, namely Connection-Adaptive Meta-learning (CAML). By searching only once, we can obtain both adaptive architecture and meta-weights for meta-learning. Extensive experiments show that our method achieves state-of-the-art performance with 3x less computational cost, revealing our method's effectiveness and efficiency.

Catastrophic forgetting (CF) frequently occurs when learning with non-stationary data distribution. The CF issue remains nearly unexplored and is more challenging when meta-learning on a sequence of domains (datasets), called sequential domain meta-learning (SDML). In this work, we propose a simple yet effective learning to learn approach, i.e., meta optimizer, to mitigate the CF problem in SDML. We first apply the proposed meta optimizer to the simplified setting of SDML, domain-aware meta-learning, where the domain labels and boundaries are known during the learning process. We propose dynamically freezing the network and incorporating it with the proposed meta optimizer by considering the domain nature during meta training. In addition, we extend the meta optimizer to the more general setting of SDML, domain-agnostic meta-learning, where domain labels and boundaries are unknown during the learning process. We propose a domain shift detection technique to capture latent domain change and equip the meta optimizer with it to work in this setting. The proposed meta optimizer is versatile and can be easily integrated with several existing meta-learning algorithms. Finally, we construct a challenging and large-scale benchmark consisting of 10 heterogeneous domains with a super long task sequence consisting of 100K tasks. We perform extensive experiments on the proposed benchmark for both settings and demonstrate the effectiveness of our proposed method, outperforming current strong baselines by a large margin.

Convolutional Neural Networks have achieved remarkable success in face recognition, in part due to the abundant availability of data. However, the data used for training CNNs is often imbalanced. Prior works largely focus on the long-tailed nature of face datasets in data volume per identity, or focus on single bias variation. In this paper, we show that many bias variations such as ethnicity, head pose, occlusion and blur can jointly affect the accuracy significantly. We propose a sample level weighting approach termed Multi-variation Cosine Margin (MvCoM), to simultaneously consider the multiple variation factors, which orthogonally enhances the face recognition losses to incorporate the importance of training samples. Further, we leverage a learning to learn approach, guided by a held-out meta learning set and use an additive modeling to predict the MvCoM. Extensive experiments on challenging face recognition benchmarks demonstrate the advantages of our method in jointly handling imbalances due to multiple variations.

Deep neural network (DNN) suffers from catastrophic forgetting when learning incrementally, which greatly limits its applications. Although maintaining a handful of samples (called "exemplars") of each task could alleviate forgetting to some extent, existing methods are still limited by the small number of exemplars since these exemplars are too few to carry enough task-specific knowledge, and therefore the forgetting remains. To overcome this problem, we propose to "imagine" diverse counterparts of given exemplars referring to the abundant semantic-irrelevant information from unlabeled data. Specifically, we develop a learnable feature generator to diversify exemplars by adaptively generating diverse counterparts of exemplars based on semantic information from exemplars and semantically-irrelevant information from unlabeled data. We introduce semantic contrastive learning to enforce the generated samples to be semantic consistent with exemplars and perform semanticdecoupling contrastive learning to encourage diversity of generated samples. The diverse generated samples could effectively prevent DNN from forgetting when learning new tasks. Our method does not bring any extra inference cost and outperforms state-of-the-art methods on two benchmarks CIFAR-100 and ImageNet-Subset by a clear margin.

This paper presents an unpaired method for creating line drawings from photographs. Current methods often rely on high quality paired datasets to generate line drawings. However, these datasets often have limitations due to the subjects of the drawings belonging to a specific domain, or in the amount of data collected. Although recent work in unsupervised image-to-image translation has shown much progress, the latest methods still struggle to generate compelling line drawings. We observe that line drawings are encodings of scene information and seek to convey 3D shape and semantic meaning. We build these observations into a set of objectives and train an image translation to map photographs into line drawings. We introduce a geometry loss which predicts depth information from the image features of a line drawing, and a semantic loss which matches the CLIP features of a line drawing with its corresponding photograph. Our approach outperforms state-of-the-art unpaired image translation and line drawing generation methods on creating line drawings from arbitrary photographs.

We propose the Model Quality Network, MQ-Net in short, for predicting the quality, e.g. the pose error of essential matrices, of models generated inside RANSAC. It replaces the traditionally used scoring techniques, e.g., inlier counting of RANSAC, truncated loss of MSAC, and the marginalization-based loss of MAGSAC++. Moreover, Minimal samples Filtering Network (MF-Net) is proposed for the early rejection of minimal samples that likely lead to degenerate models or to ones that are inconsistent with the scene geometry, e.g., due to the chirality constraint. We show on 54450 image pairs from public real-world datasets that the proposed MQ-Net leads to results superior to the state-of-the-art in terms of accuracy by a large margin. The proposed MF-Net accelerates the fundamental matrix estimation by five times and significantly reduces the essential matrix estimation time while slightly improving accuracy as well. Also, we show experimentally that consensus maximization, i.e. inlier counting, is not an inherently good measure of the model quality for relative pose estimation. The code and models will be made publicly available.

We consider the problem of recovering a single person's 3D human mesh from in-the-wild crowded scenes. While much progress has been in 3D human mesh estimation, existing methods struggle when test input has crowded scenes. The first reason for the failure is a domain gap between training and testing data. A motion capture dataset, which provides accurate 3D labels for training, lacks crowd data and impedes a network from learning crowded scene-robust image features of a target person. The second reason is a feature processing that spatially averages the feature map of a localized bounding box containing multiple people. Averaging the whole feature map makes a target person's feature indistinguishable from others. We present 3DCrowdNet that firstly explicitly targets in-the-wild crowded scenes and estimates a robust 3D human mesh by addressing the above issues. First, we leverage 2D human pose estimation that does not require a motion capture dataset with 3D labels for training and does not suffer from the domain gap. Second, we propose a joint-based regressor that distinguishes a target person's feature from others. Our joint-based regressor preserves the spatial activation of a target by sampling features from the target's joint locations and regresses human model parameters. As a result, 3DCrowdNet learns target-focused features and effectively excludes the irrelevant features of nearby persons. We conduct experiments on various benchmarks and prove the robustness of 3DCrowdNet to the in-the-wild crowded scenes both quantitatively and qualitatively. Codes are available here: https://github.com/hongsukchoi/3DCrowdNet_RELEASE

Modern camera localization methods that use image retrieval, feature matching, and 3D structure-based pose estimation require long-term storage of numerous scene images or a vast amount of image features. This can make them unsuitable for resource constrained VR/AR devices and also raises serious privacy concerns. We present a new learned camera localization technique that eliminates the need to store features or a detailed 3D point cloud. Our key idea is to implicitly encode the appearance of a sparse yet salient set of 3D scene points into a convolutional neural network (CNN) that can detect these scene points in query images whenever they are visible. We refer to these points as scene landmarks. We also show that a CNN can be trained to regress bearing vectors for such landmarks even when they are not within the camera's field-of-view. We demonstrate that the predicted landmarks yield accurate pose estimates and that our method outperforms DSAC*, the state-of-the-art in learned localization. Furthermore, extending HLoc (an accurate method) by combining its correspondences with our predictions, boosts its accuracy even further.

Current 3D object detectors for autonomous driving are almost entirely trained on human-annotated data. Although of high quality, the generation of such data is laborious and costly, restricting them to a few specific locations and object types. This paper proposes an alternative approach entirely based on unlabeled data, which can be collected cheaply and in abundance almost everywhere on earth. Our approach leverages several simple common sense heuristics to create an initial set of approximate seed labels. For example, relevant traffic participants are generally not persistent across multiple traversals of the same route, do not fly, and are never under ground. We demonstrate that these seed labels are highly effective to bootstrap a surprisingly accurate detector through repeated self-training without a single human annotated label. Code is available at https://github.com/YurongYou/MODEST.

Learning personalized models for user-customized computer-vision tasks is challenging due to the limited private-data and computation available on each edge device. Decentralized learning (DL) can exploit the images distributed over devices on a network topology to train a global model but is not designed to train personalized models for different tasks or optimize the topology. Moreover, the mixing weights used to aggregate neighbors' gradient messages in DL can be sub-optimal for personalization since they are not adaptive to different nodes/tasks and learning stages. In this paper, we dynamically update the mixing-weights to improve the personalized model for each node's task and meanwhile learn a sparse topology to reduce communication costs. Our first approach, "learning to collaborate (L2C)", directly optimizes the mixing weights to minimize the local validation loss per node for a pre-defined set of nodes/tasks. In order to produce mixing weights for new nodes or tasks, we further develop "meta-L2C", which learns an attention mechanism to automatically assign mixing weights by comparing two nodes' model updates. We evaluate both methods on diverse benchmarks and experimental settings for image classification. Thorough comparisons to both classical and recent methods for IID/non-IID decentralized and federated learning demonstrate our method's advantages in identifying collaborators among nodes, learning sparse topology, and producing better personalized models with low communication and computational cost.

Anticipating future events is an essential feature for intelligent systems and embodied AI. However, compared to the traditional recognition task, the uncertainty of future and reasoning ability requirement make the anticipation task very challenging and far beyond solved. In this filed, previous methods usually care more about the model architecture design or but few attention has been put on how to train an anticipation model with a proper learning policy. To this end, in this work, we propose a novel training scheme called Dynamic Context Removal (DCR), which dynamically schedule the visibility of observed future in the learning procedure. It follows the human-like curriculum learning process, i.e., gradually removing the event context to increase the anticipation difficulty till satisfying the final anticipation target. Our learning scheme is plug-and-play and easy to integrate any reasoning model including transformer and LSTM, with advantages in both effectiveness and efficiency. In extensive experiments, the proposed method achieves state-of-the-art on four widely-used benchmarks. Our code and models are publicly released at https://github.com/AllenXuuu/DCR.

In this paper, we focus on the Audio-Visual Question Answering (AVQA) task, which aims to answer questions regarding different visual objects, sounds, and their associations in videos. The problem requires comprehensive multimodal understanding and spatio-temporal reasoning over audio-visual scenes. To benchmark this task and facilitate our study, we introduce a large-scale MUSIC-AVQA dataset, which contains more than 45K question-answer pairs covering 33 different question templates spanning over different modalities and question types. We develop several baselines and introduce a spatio-temporal grounded audio-visual network for the AVQA problem. Our results demonstrate that AVQA benefits from multisensory perception and our model outperforms recent A-, V-, and AVQA approaches. We believe that our built dataset has the potential to serve as testbed for evaluating and promoting progress in audio-visual scene understanding and spatio-temporal reasoning. Code and dataset: http://gewu-lab.github.io/MUSIC-AVQA/

State-of-the-art methods for self-supervised sequential action alignment rely on deep networks that find correspondences across videos in time. They either learn frame-to-frame mapping across sequences, which does not leverage temporal information, or assume monotonic alignment between each video pair, which ignores variations in the order of actions. As such, these methods are not able to deal with common real-world scenarios that involve background frames or videos that contain non-monotonic sequence of actions. In this paper, we propose an approach to align sequential actions in the wild that involve diverse temporal variations. To this end, we propose an approach to enforce temporal priors on the optimal transport matrix, which leverages temporal consistency, while allowing for variations in the order of actions. Our model accounts for both monotonic and non-monotonic sequences and handles background frames that should not be aligned. We demonstrate that our approach consistently outperforms the state-of-the-art in self-supervised sequential action representation learning on four different benchmark datasets.

Few-shot learning (FSL) aims to learn a classifier that can be easily adapted to accommodate new tasks, given only a few examples. To handle the limited-data in few-shot regimes, recent methods tend to collectively use a set of local features to densely represent an image instead of using a mixed global feature. They generally explore a unidirectional paradigm, e.g., find the nearest support feature for every query feature and aggregate these local matches for a joint classification. In this paper, we propose a novel Mutual Centralized Learning (MCL) to fully affiliate these two disjoint dense features sets in a bidirectional paradigm. We first associate each local feature with a particle that can bidirectionally random walk in a discrete feature space. To estimate the class probability, we propose the dense features' accessibility that measures the expected number of visits to the dense features of that class in a Markov process. We relate our method to learning a centrality on an affiliation network and demonstrate its capability to be plugged in existing methods by highlighting centralized local features. Experiments show that our method achieves the new state-of-the-art.

This paper proposes using a sparse-structured multivariate Gaussian to provide a closed-form approximator for the output of probabilistic ensemble models used for dense image prediction tasks. This is achieved through a convolutional neural network that predicts the mean and covariance of the distribution, where the inverse covariance is parameterised by a sparsely structured Cholesky matrix. Similarly to distillation approaches, our single network is trained to maximise the probability of samples from pre-trained probabilistic models, in this work we use a fixed ensemble of networks. Once trained, our compact representation can be used to efficiently draw spatially correlated samples from the approximated output distribution. Importantly, this approach captures the uncertainty and structured correlations in the predictions explicitly in a formal distribution, rather than implicitly through sampling alone. This allows direct introspection of the model, enabling visualisation of the learned structure. Moreover, this formulation provides two further benefits: estimation of a sample probability, and the introduction of arbitrary spatial conditioning at test time. We demonstrate the merits of our approach on monocular depth estimation and show that the advantages of our approach are obtained with comparable quantitative performance.

Estimating keypoint scale and orientation is crucial to extracting invariant features under significant geometric changes. Recently, the estimators based on self-supervised learning have been designed to adapt to complex imaging conditions. Such learning-based estimators generally predict a single scalar for the keypoint scale or orientation, called hard estimators. However, hard estimators are difficult to handle the local patches containing structures of different objects or multiple edges. In this paper, a Soft Self-Supervised Estimator (S3Esti) is proposed to overcome this problem by learning to predict multiple scales and orientations. S3Esti involves three core factors. First, the estimator is constructed to predict the discrete distributions of scales and orientations. The elements with high confidence will be kept as the final scales and orientations. Second, a probabilistic covariant loss is proposed to improve the consistency of the scale and orientation distributions under different transformations. Third, an optimization algorithm is designed to minimize the loss function, whose convergence is proved in theory. When combined with different keypoint extraction models, S3Esti generally improves over 50% accuracy in image matching tasks under significant viewpoint changes. In the 3D reconstruction task, S3Esti decreases more than 10% reprojection error and improves the number of registered images.