CVPR 2021 Papers

Visual grounding, which aims to build a correspondence between visual objects and their language entities, plays a key role in cross-modal scene understanding. One promising and scalable strategy for learning visual grounding is to utilize weak supervision from only image-caption pairs. Previous methods typically rely on matching query phrases directly to a precomputed, fixed object candidate pool, which leads to inaccurate localization and ambiguous matching due to lack of semantic relation constraints. In our paper, we propose a novel context-aware weakly-supervised learning method that incorporates coarse-to-fine object refinement and entity relation modeling into a two-stage deep network, capable of producing more accurate object representation and matching. To effectively train our network, we introduce a self-taught regression loss for the proposal locations and a classification loss based on parsed entity relations. Extensive experiments on two public benchmarks Flickr30K Entities and ReferItGame demonstrate the efficacy of our weakly grounding framework. The results show that we outperform the previous methods by a considerable margin, achieving 59.27% top-1 accuracy in Flickr30K Entities and 37.68% in the ReferItGame dataset respectively.

Few-shot learning aims to correctly recognize query samples from unseen classes given a limited number of support samples, often by relying on global embeddings of images. In this paper, we propose to equip the backbone network with an attention agent, which is trained by reinforcement learning. The policy gradient algorithm is employed to train the agent towards adaptively localizing the representative regions on feature maps over time. We further design a reward function based on the prediction of the held-out data, thus helping the attention mechanism to generalize better across the unseen classes. The extensive experiments show, with the help of the reinforced attention, that our embedding network has the capability to progressively generate a more discriminative representation in few-shot learning. Moreover, experiments on the task of image classification also show the effectiveness of the proposed design.

Effective regularization techniques are highly desired in deep learning for alleviating overfitting and improving generalization. This work proposes a new regularization scheme, based on the understanding that the flat local minima of the empirical risk cause the model to generalize better. This scheme is referred to as adversarial model perturbation (AMP), where instead of directly minimizing the empirical risk, an alternative "AMP loss" is minimized via SGD. Specifically, the AMP loss is obtained from the empirical risk by applying the "worst" norm-bounded perturbation on each point in the parameter space. Comparing with most existing regularization schemes, AMP has strong theoretical justifications, in that minimizing the AMP loss can be shown theoretically to favour flat local minima of the empirical risk. Extensive experiments on various modern deep architectures establish AMP as a new state of the art among regularization schemes. Our code is available at https://github.com/hiyouga/AMP-Regularizer.

Recent years have witnessed the rapid progress of generative adversarial networks (GANs). However, the success of the GAN models hinges on a large amount of training data. This work proposes a regularization approach for training robust GAN models on limited data. We theoretically show a connection between the regularized loss and an f-divergence called LeCam-divergence, which we find is more robust under limited training data. Extensive experiments on several benchmark datasets demonstrate that the proposed regularization scheme 1) improves the generalization performance and stabilizes the learning dynamics of GAN models under limited training data, and 2) complements the recent data augmentation methods. These properties facilitate training GAN models to achieve state-of-the-art performance when only limited training data of the ImageNet benchmark is available. The source code is available at https://github.com/google/lecam-gan.

Data augmentation is an effective regularization strategy to alleviate the overfitting, which is an inherent drawback of the deep neural networks. However, data augmentation is rarely considered for point cloud processing despite many studies proposing various augmentation methods for image data. Actually, regularization is essential for point clouds since lack of generality is more likely to occur in point cloud due to small datasets. This paper proposes a Rigid Subset Mix (RSMix), a novel data augmentation method for point clouds that generates a virtual mixed sample by replacing part of the sample with shape-preserved subsets from another sample. RSMix preserves structural information of the point cloud sample by extracting subsets from each sample without deformation using a neighboring function. The neighboring function was carefully designed considering unique properties of point cloud, unordered structure and non-grid. Experiments verified that RSMix successfully regularized the deep neural networks with remarkable improvement for shape classification. We also analyzed various combinations of data augmentations including RSMix with single and multi-view evaluations, based on abundant ablation studies.

Domain adaptation (DA) aims at transferring knowledge from a labeled source domain to an unlabeled target domain. Though many DA theories and algorithms have been proposed, most of them are tailored into classification settings and may fail in regression tasks, especially in the practical keypoint detection task. To tackle this difficult but significant task, we present a method of regressive domain adaptation (RegDA) for unsupervised keypoint detection. Inspired by the latest theoretical work, we first utilize an adversarial regressor to maximize the disparity on the target domain and train a feature generator to minimize this disparity. However, due to the high dimension of the output space, this regressor fails to detect samples that deviate from the support of the source. To overcome this problem, we propose two important ideas. First, based on our observation that the probability density of the output space is sparse, we introduce a spatial probability distribution to describe this sparsity and then use it to guide the learning of the adversarial regressor. Second, to alleviate the optimization difficulty in the high-dimensional space, we innovatively convert the minimax game in the adversarial training to the minimization of two opposite goals. Extensive experiments show that our method brings large improvement by 8% to 11% in terms of PCK on different datasets.

Image composition plays a common but important role in photo editing. To acquire photo-realistic composite images, one must adjust the appearance and visual style of the foreground to be compatible with the background. Existing deep learning methods for harmonizing composite images directly learn an image mapping network from the composite to real one, without explicit exploration on visual style consistency between the background and the foreground images. To ensure the visual style consistency between the foreground and the background, in this paper, we treat image harmonization as a style transfer problem. In particular, we propose a simple yet effective Region-aware Adaptive Instance Normalization (RAIN) module, which explicitly formulates the visual style from the background and adaptively applies them to the foreground. With our settings, our RAIN module can be used as a drop-in module for existing image harmonization networks and is able to bring significant improvements. Extensive experiments on the existing image harmonization benchmark datasets show the superior capability of the proposed method. Code is available at https://github.com/junleen/RainNet .

Determining which image regions to concentrate is critical for Human-Object Interaction (HOI) detection. Conventional HOI detectors focus on either detected human and object pairs or pre-defined interaction locations, which limits learning of the effective features. In this paper, we reformulate HOI detection as an adaptive set prediction problem, with this novel formulation, we propose an Adaptive Set-based one-stage framework (AS-Net) with parallel instance and interaction branches. To attain this, we map a trainable interaction query set to an interaction prediction set with transformer. Each query adaptively aggregates the interaction-relevant features from global contexts through multi-head co-attention. Besides, the training process is supervised adaptively by matching each ground-truth with the interaction prediction. Furthermore, we design an effective instance-aware attention module to introduce instructive features from the instance branch into the interaction branch. Our method outperforms previous state-of-the-art methods without any extra human pose and language features on three challenging HOI detection datasets. Especially, we achieve over 31% relative improvement on a large scale HICO-DET dataset. Code is available at https://github.com/yoyomimi/AS-Net.

Unsupervised object re-identification targets at learning discriminative representations for object retrieval without any annotations. Clustering-based methods conduct training with the generated pseudo labels and currently dominate this research direction. However, they still suffer from the issue of pseudo label noise. To tackle the challenge, we propose to properly estimate pseudo label similarities between consecutive training generations with clustering consensus and refine pseudo labels with temporally propagated and ensembled pseudo labels. To the best of our knowledge, this is the first attempt to leverage the spirit of temporal ensembling to improve classification with dynamically changing classes over generations. The proposed pseudo label refinery strategy is simple yet effective and can be seamlessly integrated into existing clustering-based unsupervised re-identification methods. With our proposed approach, state-of-the-art method can be further boosted with up to 8.8% mAP improvements on the challenging MSMT17 dataset.

The two-stage methods for instance segmentation, e.g. Mask R-CNN, have achieved excellent performance recently. However, the segmented masks are still very coarse due to the downsampling operations in both the feature pyramid and the instance-wise pooling process, especially for large objects. In this work, we propose a new method called RefineMask for high-quality instance segmentation of objects and scenes, which incorporates fine-grained features during the instance-wise segmenting process in a multi-stage manner. Through fusing more detailed information stage by stage, RefineMask is able to refine high-quality masks consistently. RefineMask succeeds in segmenting hard cases such as bent parts of objects that are over-smoothed by most previous methods and outputs accurate boundaries. Without bells and whistles, RefineMask yields significant gains of 2.6, 3.4, 3.8 AP over Mask R-CNN on COCO, LVIS, and Cityscapes benchmarks respectively at a small amount of additional computational cost. Furthermore, our single-model result outperforms the winner of the LVIS Challenge 2020 by 1.3 points on the LVIS test-dev set and establishes a new state-of-the-art. Code will be available at https://github.com/zhanggang001/RefineMask.

Knowledge distillation is a method of transferring the knowledge from a pretrained complex teacher model to a student model, so a smaller network can replace a large teacher network at the deployment stage. To reduce the necessity of training a large teacher model, the recent literatures introduced a self-knowledge distillation, which trains a student network progressively to distill its own knowledge without a pretrained teacher network. While Self-knowledge distillation is largely divided into a data augmentation based approach and an auxiliary network based approach, the data augmentation approach looses its local information in the augmentation process, which hinders its applicability to diverse vision tasks, such as semantic segmentation. Moreover, these knowledge distillation approaches do not receive the refined feature maps, which are prevalent in the object detection and semantic segmentation community. This paper proposes a novel self-knowledge distillation method, Feature Refinement via Self-Knowledge Distillation (FRSKD), which utilizes an auxiliary self-teacher network to transfer a refined knowledge for the classifier network. Our proposed method, FRSKD, can utilize both soft label and feature-map distillations for the self-knowledge distillation. Therefore, FRSKD can be applied to classification, and semantic segmentation, which emphasize preserving the local information. We demonstrate the effectiveness of FRSKD by enumerating its performance improvements in diverse tasks and benchmark datasets. The implemented code will be open-sourced.

Grounding referring expressions in RGBD image has been an emerging field. We present a novel task of 3D visual grounding in single-view RGBD image where the referred objects are often only partially scanned due to occlusion. In contrast to previous works that directly generate object proposals for grounding in the 3D scenes, we propose a bottom-up approach to gradually aggregate content-aware information, effectively addressing the challenge posed by the partial geometry. Our approach first fuses the language and the visual features at the bottom level to generate a heatmap that coarsely localizes the relevant regions in the RGBD image. Then our approach conducts an adaptive feature learning based on the heatmap and performs the object-level matching with another visio-linguistic fusion to finally ground the referred object. We evaluate the proposed method by comparing to the state-of-the-art methods on both the RGBD images extracted from the ScanRefer dataset and our newly collected SUNRefer dataset. Experiments show that our method outperforms the previous methods by a large margin (by 11.2% and 15.6% Acc@0.5) on both datasets.

Convolutional Neural Networks (CNNs) often fail to maintain their performance when they confront new test domains, which is known as the problem of domain shift. Recent studies suggest that one of the main causes of this problem is CNNs' strong inductive bias towards image styles (i.e. textures) which are sensitive to domain changes, rather than contents (i.e. shapes). Inspired by this, we propose to reduce the intrinsic style bias of CNNs to close the gap between domains. Our Style-Agnostic Networks (SagNets) disentangle style encodings from class categories to prevent style biased predictions and focus more on the contents. Extensive experiments show that our method effectively reduces the style bias and makes the model more robust under domain shift. It achieves remarkable performance improvements in a wide range of cross-domain tasks including domain generalization, unsupervised domain adaptation, and semi-supervised domain adaptation on multiple datasets.

Learning non-rigid registration in an end-to-end manner is challenging due to the inherent high degrees of freedom and the lack of labeled training data. In this paper, we resolve these two challenges simultaneously. First, we propose to represent the non-rigid transformation with a point-wise combination of several rigid transformations. This representation not only makes the solution space well-constrained but also enables our method to be solved iteratively with a recurrent framework, which greatly reduces the difficulty of learning. Second, we introduce a differentiable loss function that measures the 3D shape similarity on the projected multi-view 2D depth images so that our full framework can be trained end-to-end without ground truth supervision. Extensive experiments on several different datasets demonstrate that our proposed method outperforms the previous state-of-the-art by a large margin.

Deep learning models suffer from catastrophic forgetting when trained in an incremental learning setting. In this work, we propose a novel approach to address the task incremental learning problem, which involves training a model on new tasks that arrive in an incremental manner. The task incremental learning problem becomes even more challenging when the test set contains classes that are not part of the train set, i.e., a task incremental generalized zero-shot learning problem. Our approach can be used in both the zero-shot and non zero-shot task incremental learning settings. Our proposed method uses weight rectifications and affine transformations in order to adapt the model to different tasks that arrive sequentially. Specifically, we adapt the network weights to work for new tasks by "rectifying" the weights learned from the previous task. We learn these weight rectifications using very few parameters. We additionally learn affine transformations on the outputs generated by the network in order to better adapt them for the new task. We perform experiments on several datasets in both zero-shot and non zero-shot task incremental learning settings and empirically show that our approach achieves state-of-the-art results. Specifically, our approach outperforms the state-of-the-art non zero-shot task incremental learning method by over 5% on the CIFAR-100 dataset. Our approach also significantly outperforms the state-of-the-art task incremental generalized zero-shot learning method by absolute margins of 6.91% and 6.33% for the AWA1 and CUB datasets, respectively. We validate our approach using various ablation studies.

Deep denoiser, the deep network for denoising, has been the focus of the recent development on image denoising. In the last few years, there is an increasing interest in developing unsupervised deep denoisers which only call unorganized noisy images without ground truth for training. Nevertheless, the performance of these unsupervised deep denoisers is not competitive to their supervised counterparts. Aiming at developing a more powerful unsupervised deep denoiser, this paper proposed a data augmentation technique, called recorrupted-to-recorrupted (R2R), to address the overfitting caused by the absence of truth images. For each noisy image, we showed that the cost function defined on the noisy/noisy image pairs constructed by the R2R method is statistically equivalent to its supervised counterpart defined on the noisy/truth image pairs. Extensive experiments showed that the proposed R2R method noticeably outperformed existing unsupervised deep denoisers, and is competitive to representative supervised deep denoisers.

In this paper, we introduce the new task of reconstructing 3D human pose from a single image in which we can see the person and the person's image through a mirror. Compared to general scenarios of 3D pose estimation from a single view, the mirror reflection provides an additional view for resolving the depth ambiguity. We develop an optimization-based approach that exploits mirror symmetry constraints for accurate 3D pose reconstruction. We also provide a method to estimate the surface normal of the mirror from vanishing points in the single image. To validate the proposed approach, we collect a large-scale dataset named Mirrored-Human, which covers a large variety of human subjects, poses and backgrounds. The experiments demonstrate that, when trained on Mirrored-Human with our reconstructed 3D poses as pseudo ground-truth, the accuracy and generalizability of existing single-view 3D pose estimators can be largely improved.

Aligning distributions of view representations is a core component of today's state of the art models for deep multi-view clustering. However, we identify several drawbacks with naively aligning representation distributions. We demonstrate that these drawbacks both lead to less separable clusters in the representation space, and inhibit the model's ability to prioritize views. Based on these observations, we develop a simple baseline model for deep multi-view clustering. Our baseline model avoids representation alignment altogether, while performing similar to, or better than, the current state of the art. We also expand our baseline model by adding a contrastive learning component. This introduces a selective alignment procedure that preserves the model's ability to prioritize views. Our experiments show that the contrastive learning component enhances the baseline model, improving on the current state of the art by a large margin on several datasets.

Standard methods for video recognition use large CNNs designed to capture spatio-temporal data. However, training these models requires a large amount of labeled training data, containing a wide variety of actions, scenes, settings and camera viewpoints. In this paper, we show that current convolutional neural network models are unable to recognize actions from camera viewpoints not present in their training data (i.e., unseen view action recognition). To address this, we develop approaches based on 3D pose and introduce a new geometric convolutional layer that can learn viewpoint invariant representations. Further, we introduce a new, challenging dataset for unseen view recognition and show the approaches ability to learn viewpoint invariant representations.