CVPR 2021 Papers

Existing person search methods integrate person detection and re-identification (re-ID) module into a unified system. Though promising results have been achieved, the misalignment problem, which commonly occurs in person search, limits the discriminative feature representation for re-ID. To overcome this limitation, we introduce a novel framework to learn the discriminative representation by utilizing prototype in OIM loss. Unlike conventional methods using prototype as a representation of person identity, we utilize it as guidance to allow the attention network to consistently highlight multiple instances across different poses. Moreover, we propose a new prototype update scheme with adaptive momentum to increase the discriminative ability across different instances. Extensive ablation experiments demonstrate that our method can significantly enhance the feature discriminative power, outperforming the state-of-the-art results on two person search benchmarks including CUHK-SYSU and PRW.

Few-shot learning is a challenging task, which aims to learn a classifier for novel classes with few examples. Pre-training based meta-learning methods effectively tackle the problem by pre-training a feature extractor and then fine-tuning it through the nearest centroid based meta-learning. However, results show that the fine-tuning step makes very marginal improvements. In this paper, 1) we figure out the key reason, i.e., in the pre-trained feature space, the base classes already form compact clusters while novel classes spread as groups with large variances, which implies that fine-tuning the feature extractor is less meaningful; 2) instead of fine-tuning the feature extractor, we focus on estimating more representative prototypes during meta-learning. Consequently, we propose a novel prototype completion based meta-learning framework. This framework first introduces primitive knowledge (i.e., class-level part or attribute annotations) and extracts representative attribute features as priors. Then, we design a prototype completion network to learn to complete prototypes with these priors. To avoid the prototype completion error caused by primitive knowledge noises or class differences, we further develop a Gaussian based prototype fusion strategy that combines the mean-based and completed prototypes by exploiting the unlabeled samples. Extensive experiments show that our method: (i) can obtain more accurate prototypes; (ii) outperforms state-of-the-art techniques by 2% 9% in terms of classification accuracy. Our code is available online.

Despite the impressive performance in many individual tasks, deep neural networks suffer from catastrophic forgetting when learning new tasks incrementally. Recently, various incremental learning methods have been proposed, and some approaches achieved acceptable performance relying on stored data or complex generative models. However, storing data from previous tasks is limited by memory or privacy issues, and generative models are usually unstable and inefficient in training. In this paper, we propose a simple non-exemplar based method named PASS, to address the catastrophic forgetting problem in incremental learning. On the one hand, we propose to memorize one class-representative prototype for each old class and adopt prototype augmentation (protoAug) in the deep feature space to maintain the decision boundary of previous tasks. On the other hand, we employ self-supervised learning (SSL) to learn more generalizable and transferable features for other tasks, which demonstrates the effectiveness of SSL in incremental learning. Experimental results on benchmark datasets show that our approach significantly outperforms non-exemplar based methods, and achieves comparable performance compared to exemplar based approaches.

Ever since Machine Learning as a Service emerges as a viable business that utilizes deep learning models to generate lucrative revenue, Intellectual Property Right (IPR) has become a major concern because these deep learning models can easily be replicated, shared, and re-distributed by any unauthorized third parties. To the best of our knowledge, one of the prominent deep learning models - Generative Adversarial Networks (GANs) which has been widely used to create photorealistic image are totally unprotected despite the existence of pioneering IPR protection methodology for Convolutional Neural Networks (CNNs). This paper therefore presents a complete protection framework in both black-box and white-box settings to enforce IPR protection on GANs. Empirically, we show that the proposed method does not compromise the original GANs performance (i.e. image generation, image super-resolution, style transfer), and at the same time, it is able to withstand both removal and ambiguity attacks against embedded watermarks.

Contrastive learning methods for unsupervised visual representation learning have reached remarkable levels of transfer performance. We argue that the power of contrastive learning has yet to be fully unleashed, as current methods are trained only on instance-level pretext tasks, leading to representations that may be sub-optimal for downstream tasks requiring dense pixel predictions. In this paper, we introduce pixel-level pretext tasks for learning dense feature representations. The first task directly applies contrastive learning at the pixel level. We additionally propose a pixel-to-propagation consistency task that produces better results, even surpassing the state-of-the-art approaches by a large margin. Specifically, it achieves 60.2 AP, 41.4 / 40.5 mAP and 77.2 mIoU when transferred to Pascal VOC object detection (C4), COCO object detection (FPN / C4) and Cityscapes semantic segmentation using a ResNet-50 backbone network, which are 2.6 AP, 0.8 / 1.0 mAP and 1.0 mIoU better than the previous best methods built on instance-level contrastive learning. Moreover, the pixel-level pretext tasks are found to be effective for pre-training not only regular backbone networks but also head networks used for dense downstream tasks, and are complementary to instance-level contrastive methods. These results demonstrate the strong potential of defining pretext tasks at the pixel level, and suggest a new path forward in unsupervised visual representation learning. Code is available at https://github.com/zdaxie/PixPro.

HD map reconstruction is crucial for autonomous driving. LiDAR-based methods are limited due to the deployed expensive sensors and time-consuming computation. Camera-based methods usually need to separately perform road segmentation and view transformation, which often causes distortion and the absence of content. To push the limits of the technology, we present a novel framework that enables reconstructing a local map formed by road layout and vehicle occupancy in the bird's-eye view given a front-view monocular image only. In particular, we propose a cross-view transformation module, which takes the constraint of cycle consistency between views into account and makes full use of their correlation to strengthen the view transformation and scene understanding. Considering the relationship between vehicles and roads, we also design a context-aware discriminator to further refine the results. Experiments on public benchmarks show that our method achieves the state-of-the-art performance in the tasks of road layout estimation and vehicle occupancy estimation. Especially for the latter task, our model outperforms all competitors by a large margin. Furthermore, our model runs at 35 FPS on a single GPU, which is efficient and applicable for real-time panorama HD map reconstruction.

Distortion rectification is often required for fisheye images. The generation-based method is one mainstream solution due to its label-free property, but its naive skip-connection and overburdened decoder will cause blur and incomplete correction. First, the skip-connection directly transfers the image features, which may introduce distortion and cause incomplete correction. Second, the decoder is overburdened during simultaneously reconstructing the content and structure of the image, resulting in vague performance. To solve these two problems, in this paper, we focus on the interpretable correction mechanism of the distortion rectification network and propose a feature-level correction scheme. We embed a correction layer in skip-connection and leverage the appearance flows in different layers to pre-correct the image features. Consequently, the decoder can easily reconstruct a plausible result with the remaining distortion-less information. In addition, we propose a parallel complementary structure. It effectively reduces the burden of the decoder by separating content reconstruction and structure correction. Subjective and objective experiment results on different datasets demonstrate the superiority of our method.

In this paper, we propose a progressive unsupervised learning (PUL) framework, which entirely removes the need for annotated training videos in visual tracking. Specifically, we first learn a background discrimination (BD) model that effectively distinguishes an object from background in a contrastive learning way. We then employ the BD model to progressively mine temporal corresponding patches (i.e., patches connected by a track) in sequential frames. As the BD model is imperfect and thus the mined patch pairs are noisy, we propose a noise-robust loss function to more effectively learn temporal correspondences from this noisy data. We use the proposed noise robust loss to train backbone networks of Siamese trackers. Without online fine-tuning or adaptation, our unsupervised real-time Siamese trackers can outperform state-of-the-art unsupervised deep trackers and achieve competitive results to the supervised baselines.

Video inpainting aims to fill spatio-temporal "corrupted" regions with plausible content. To achieve this goal, it is necessary to find correspondences from neighbouring frames to faithfully hallucinate the unknown content. Current methods achieve this goal through attention, flow-based warping, or 3D temporal convolution. However, flow-based warping can create artifacts when optical flow is not accurate, while temporal convolution may suffer from spatial misalignment. We propose `Progressive Temporal Feature Alignment Network', which progressively enriches features extracted from the current frame with the feature warped from neighbouring frames using optical flow. Our approach corrects the spatial misalignment in the temporal feature propagation stage, greatly improving visual quality and temporal consistency of the inpainted videos. Using the proposed architecture, we achieve state-of-the-art performance on the DAVIS and FVI datasets compared to existing deep learning approaches. Code is available at https://github.com/MaureenZOU/TSAM.

Unsupervised learning methods have recently shown their competitiveness against supervised training. Typically, these methods use a single objective to train the entire network. But one distinct advantage of unsupervised over supervised learning is that the former possesses more variety and freedom in designing the objective. In this work, we explore new dimensions of unsupervised learning by proposing the Progressive Stage-wise Learning (PSL) framework. For a given unsupervised task, we design multi-level tasks and define different learning stages for the deep network. Early learning stages are forced to focus on low-level tasks while late stages are guided to extract deeper information through harder tasks. We discover that by progressive stage-wise learning, unsupervised feature representation can be effectively enhanced. Our extensive experiments show that PSL consistently improves results for the leading unsupervised learning methods.

Face restoration is important in face image processing, and has been widely studied in recent years. However, previous works often fail to generate plausible high quality (HQ) results for real-world low quality (LQ) face images. In this paper, we propose a new progressive semantic-aware style transformation framework, named PSFR-GAN, for face restoration. Specifically, instead of using an encoder-decoder framework as previous methods, we formulate the restoration of LQ face images as a multi-scale progressive restoration procedure through semantic-aware style transformation. Given a pair of LQ face image and its corresponding parsing map, we first generate a multi-scale pyramid of the inputs, and then progressively modulate different scale features from coarse-to-fine in a semantic-aware style transfer way. Compared with previous networks, the proposed PSFR-GAN makes full use of the semantic (parsing maps) and pixel (LQ images) space information from different scales of input pairs. In addition, we further introduce a semantic aware style loss which calculates the feature style loss for each semantic region individually to improve the details of face textures. Finally, we pretrain a face parsing network which can generate decent parsing maps from real-world LQ face images. Experiment results show that our model trained with synthetic data can produce more realistic high-resolution results for synthetic LQ inputs than state-of-the-art methods and generalize better to natural LQ face images.

The objective of this work is to segment high-resolution images without overloading GPU memory usage or losing the fine details in the output segmentation map. The memory constraint means that we must either downsample the big image or divide the image into local patches for separate processing. However, the former approach would lose the fine details, while the latter can be ambiguous due to the lack of a global picture. In this work, we present MagNet, a multi-scale framework that resolves local ambiguity by looking at the image at multiple magnification levels. MagNet has multiple processing stages, where each stage corresponds to a magnification level, and the output of one stage is fed into the next stage for coarse-to-fine information propagation. Each stage analyzes the image at a higher resolution than the previous stage, recovering the previously lost details due to the lossy downsampling step, and the segmentation output is progressively refined through the processing stages. Experiments on three high-resolution datasets of urban views, aerial scenes, and medical images shows that MagNet consistently outperforms the state-of-the-art methods by a significant margin.

Human multimodal emotion recognition involves time-series data of different modalities, such as natural language, visual motions, and acoustic behaviors. Due to the variable sampling rates for sequences from different modalities, the collected multimodal streams are usually unaligned. The asynchrony across modalities increases the difficulty on conducting efficient multimodal fusion. Hence, this work mainly focuses on multimodal fusion from unaligned multimodal sequences. To this end, we propose the Progressive Modality Reinforcement (PMR) approach based on the recent advances of crossmodal transformer. Our approach introduces a message hub to exchange information with each modality. The message hub sends common messages to each modality and reinforces their features via crossmodal attention. In turn, it also collects the reinforced features from each modality and uses them to generate a reinforced common message. By repeating the cycle process, the common message and the modalities' features can progressively complement each other. Finally, the reinforced features are used to make predictions for human emotion. Comprehensive experiments on different human multimodal emotion recognition benchmarks clearly demonstrate the superiority of our approach.

Single domain generalization is a challenging case of model generalization, where the models are trained on a single domain and tested on other unseen domains. A promising solution is to learn cross-domain invariant representations by expanding the coverage of the training domain. These methods have limited generalization performance gains in practical applications due to the lack of appropriate safety and effectiveness constraints. In this paper, we propose a novel learning framework called progressive domain expansion network (PDEN) for single domain generalization. The domain expansion subnetwork and representation learning subnetwork in PDEN mutually benefit from each other by joint learning. For the domain expansion subnetwork, multiple domains are progressively generated in order to simulate various photometric and geometric transforms in unseen domains. A series of strategies are introduced to guarantee the safety and effectiveness of the expanded domains. For the domain invariant representation learning subnetwork, contrastive learning is introduced to learn the domain invariant representation in which each class is well clustered so that a better decision boundary can be learned to improve it's generalization. Extensive experiments on classification and segmentation have shown that PDEN can achieve up to 15.28% improvement compared with the state-of-the-art single-domain generalization methods.

State-of-the-art scene text detection methods usually model the text instance with local pixels or components from the bottom-up perspective and, therefore, are sensitive to noises and dependent on the complicated heuristic post-processing especially for arbitrary-shape texts. To relieve these two issues, instead, we propose to progressively evolve the initial text proposal to arbitrarily shaped text contours in a top-down manner. The initial horizontal text proposals are generated by estimating the center and size of texts. To reduce the range of regression, the first stage of the evolution predicts the corner points of oriented text proposals from the initial horizontal ones. In the second stage, the contours of the oriented text proposals are iteratively regressed to arbitrarily shaped ones. In the last iteration of this stage, we rescore the confidence of the final localized text by utilizing the cues from multiple contour points, rather than the single cue from the initial horizontal proposal center that may be out of arbitrary-shape text regions. Moreover, to facilitate the progressive contour evolution, we design a contour information aggregation mechanism to enrich the feature representation on text contours by considering both the circular topology and semantic context. Experiments conducted on CTW1500, Total-Text, ArT, and TD500 have demonstrated that the proposed method especially excels in line-level arbitrary-shape texts. Code is available at http://github.com/dpengwen/PCR.

Despite the recent advances in multiple object tracking (MOT), achieved by joint detection and tracking, dealing with long occlusions remains a challenge. This is due to the fact that such techniques tend to ignore the long-term motion information. In this paper, we introduce a probabilistic autoregressive motion model to score tracklet proposals by directly measuring their likelihood. This is achieved by training our model to learn the underlying distribution of natural tracklets. As such, our model allows us not only to assign new detections to existing tracklets, but also to inpaint a tracklet when an object has been lost for a long time, e.g., due to occlusion, by sampling tracklets so as to fill the gap caused by misdetections. Our experiments demonstrate the superiority of our approach at tracking objects in challenging sequences; it outperforms the state of the art in most standard MOT metrics on multiple MOT benchmark datasets, including MOT16, MOT17, and MOT20.

Model protection is vital when deploying Convolutional Neural Networks (CNNs) for commercial services, due to the massive costs of training them. In this work, we propose a selective encryption (SE) algorithm to protect CNN models from unauthorized access, with a unique feature of providing hierarchical services to users. Our algorithm firstly selects important model parameters via the proposed Probabilistic Selection Strategy (PSS). It then encrypts the most important parameters with the designed encryption method called Distribution Preserving Random Mask (DPRM), so as to maximize the performance degradation by encrypting only a very small portion of model parameters. We also design a set of access permissions, using which different amount of most important model parameters can be decrypted. Hence, different levels of model performance can be naturally provided for users. Experimental results demonstrate that the proposed scheme could effectively protect the classification model VGG19 by merely encrypting 8% parameters of convolutional layers. We also implement the proposed model protection scheme in the denoising model DnCNN, showcasing the hierarchical denoising services.

To generate "accurate" scene graphs, almost all exist-ing methods predict pairwise relationships in a determin-istic manner. However, we argue that visual relationshipsare often semantically ambiguous. Specifically, inspired bylinguistic knowledge, we classify the ambiguity into threetypes: Synonymy Ambiguity, Hyponymy Ambiguity, andMulti-view Ambiguity. The ambiguity naturally leads to theissue ofimplicit multi-label, motivating the need for diversepredictions. In this work, we propose a novel plug-and-play Probabilistic Uncertainty Modeling (PUM) module. Itmodels each union region as a Gaussian distribution, whosevariance measures the uncertainty of the corresponding vi-sual content. Compared to the conventional determinis-tic methods, such uncertainty modeling brings stochasticityof feature representation, which naturally enables diversepredictions. As a byproduct, PUM also manages to covermore fine-grained relationships and thus alleviates the is-sue of bias towards frequent relationships. Extensive exper-iments on the large-scale Visual Genome benchmark showthat combining PUM with newly proposed ResCAGCN canachieve state-of-the-art performances, especially under themean recall metric. Furthermore, we show the universal ef-fectiveness of PUM by plugging it into some existing modelsand provide insightful analysis of its ability to generate di-verse yet plausible visual relationships.

Semantic correspondence is a fundamental problem in computer vision, which aims at establishing dense correspondences across images depicting different instances under the same category. This task is challenging due to large intra-class variations and a severe lack of ground truth. A popular solution is to learn correspondences from synthetic data. However, because of the limited intra-class appearance and background variations within synthetically generated training data, the model's capability for handling "real" image pairs using such strategy is intrinsically constrained. We address this problem with the use of a novel Probabilistic Model Distillation (PMD) approach which transfers knowledge learned by a probabilistic teacher model on synthetic data to a static student model with the use of unlabeled real image pairs. A probabilistic supervision reweighting (PSR) module together with a confidence-aware loss (CAL) is used to mine the useful knowledge and alleviate the impact of errors. Experimental results on a variety of benchmarks show that our PMD achieves state-of-the-art performance. To demonstrate the generalizability of our approach, we extend PMD to incorporate stronger supervision for better accuracy -- the probabilistic teacher is trained with stronger key-point supervision. Again, we observe the superiority of our PMD. The extensive experiments verify that PMD is able to infer more reliable supervision signals from the probabilistic teacher for representation learning and largely alleviate the influence of errors in pseudo labels.