CVPR 2021 Papers

Recently unsupervised domain adaptation for the semantic segmentation task has become more and more popular due to the high-cost of pixel-level annotation on real-world images. However, most domain adaptation methods are only restricted to single-source-single-target pair, and can not be directly extended to multiple target domains. In this work, we propose a collaborative learning framework to achieve unsupervised multi-target domain adaptation. An unsupervised domain adaptation expert model is first trained for each source-target pair and is further encouraged to collaborate with each other through a bridge built between different target domains. These expert models are further improved by adding the regularization of making the consistent pixel-wise prediction for each sample with the same structured context. To obtain a single model that works across multiple target domains, we propose to simultaneously learn a student model which is trained to not only imitate the output of each expert on the corresponding target domain but also to pull different expert close to each other with regularization on their weights. Extensive experiments demonstrate that the proposed method can effectively exploit rich structured information contained in both labeled source domain and multiple unlabeled target domains. Not only does it perform well across multiple target domains but also performs favorably against state-of-the-art unsupervised domain adaptation methods specially trained on a single source-target pair.

Image restoration tasks demand a complex balance between spatial details and high-level contextualized information while recovering images. In this paper, we propose a novel synergistic design that can optimally balance these competing goals. Our main proposal is a multi-stage architecture, that progressively learns restoration functions for the degraded inputs, thereby breaking down the overall recovery process into more manageable steps. Specifically, our model first learns the contextualized features using encoder-decoder architectures and later combines them with a high-resolution branch that retains local information. At each stage, we introduce a novel per-pixel adaptive design that leverages in-situ supervised attention to reweight the local features. A key ingredient in such a multi-stage architecture is the information exchange between different stages. To this end, we propose a two-faceted approach where the information is not only exchanged sequentially from early to late stages, but lateral connections between feature processing blocks also exist to avoid any loss of information. The resulting tightly interlinked multi-stage architecture, named as MPRNet, delivers strong performance gains on ten datasets across a range of tasks including image deraining, deblurring, and denoising. The source code and pre-trained models are available at https://github.com/swz30/MPRNet.

We address the problem of localizing a specific moment from an untrimmed video by a language sentence query. Generally, previous methods mainly exist two problems that are not fully solved: 1) How to effectively model the fine-grained visual-language alignment between video and language query? 2) How to accurately localize the moment in the original video length? In this paper, we streamline the temporal language localization as a novel multi-stage aggregated transformer network. Specifically, we first introduce a new visual-language transformer backbone, which enables iterations and alignments among all elements in visual and language sequences. Different from previous multi-modal transformers, our backbone keeps both structure unified and modality specific. Moreover, we also propose a multi-stage aggregation module topped on the transformer backbone. In this module, we compute three stage-specific representations corresponding to different moment stages respectively, i.e. starting, middle and ending stages, for each video element. Then for a moment candidate, we concatenate the starting/middle/ending representations of its starting/middle/ending elements respectively to form the final moment representation. Because the obtained moment representation captures the stage specific information, it is very discriminative for accurate localization. Extensive experiments on ActivityNet Captions and TACoS datasets demonstrate our proposed method achieves significant improvements compared with all other methods.

Multi-source unsupervised domain adaptation (MSDA) aims at adapting models trained on multiple labeled source domains to an unlabeled target domain. In this paper, we propose a novel multi-source domain adaptation framework based on collaborative learning for semantic segmentation. Firstly, a simple image translation method is introduced to align the pixel value distribution to reduce the gap between source domains and target domain to some extent. Then, to fully exploit the essential semantic information across source domains, we propose a collaborative learning method for domain adaptation without seeing any data from target domain. In addition, similar to the setting of unsupervised domain adaptation, unlabeled target domain data is leveraged to further improve the performance of domain adaptation. This is achieved by additionally constraining the outputs of multiple adaptation models with pseudo labels online generated by an ensembled model. Extensive experiments and ablation studies are conducted on the widely-used domain adaptation benchmark datasets in semantic segmentation. Our proposed method achieves 59.0% mIoU on the validation set of Cityscapes by training on the labeled Synscapes and GTA5 datasets and unlabeled training set of Cityscapes. It significantly outperforms all previous state-of-the-arts single-source and multi-source unsupervised domain adaptation methods.

Current developments in temporal event or action localization usually target actions captured by a single camera. However, extensive events or actions in the wild may be captured as a sequence of shots by multiple cameras at different positions. In this paper, we propose a new and challenging task called multi-shot temporal event localization, and accordingly, collect a large-scale dataset called MUlti-Shot EventS (MUSES). MUSES has 31,477 event instances for a total of 716 video hours. The core nature of MUSES is the frequent shot cuts, for an average of 19 shots per instance and 176 shots per video, which induces large intra-instance variations. Our comprehensive evaluations show that the state-of-the-art method in temporal action localization only achieves an mAP of 13.1% at IoU=0.5. As a minor contribution, we present a simple baseline approach for handling the intra-instance variations, which reports an mAP of 18.9% on MUSES and 56.9% on THUMOS14 at IoU=0.5. To facilitate research in this direction, we release the dataset and the project code at https://songbai.site/muses/.

In instance-level detection tasks (e.g., object detection), reducing input resolution is an easy option to improve runtime efficiency. However, this option severely hurts the detection performance. This paper focuses on boosting the performance of a low-resolution model, by distilling knowledge from a high/multi-resolution model. We first identify the challenge of applying knowledge distillation to teacher and student networks that act on different input resolutions. To tackle the challenge, we explore the idea of spatially aligning feature maps between models of different input resolutions, by shifting the position of the feature pyramid structure. With the alignment idea, we introduce aligned multi-scale training to train a multi-scale teacher that can distill its knowledge seamlessly to a low-resolution student. Furthermore, we propose cross feature-level fusion to dynamically fuse the multi-resolution features of the same teacher, to better guide the student. On several instance-level detection tasks and datasets, the low-resolution models trained via our approach perform competitively with high-resolution models trained via conventional multi-scale training, while outperforming the latter's low-resolution models by 2.1% to 3.6% in mAP.

We present a novel LSTM cell architecture capable of learning both intra- and inter-perspective relationships available in visual sequences captured from multiple perspectives. Our architecture adopts a novel recurrent joint learning strategy that uses additional gates and memories at the cell level. We demonstrate that by using the proposed cell to create a network, more effective and richer visual representations are learned for recognition tasks. We validate the performance of our proposed architecture in the context of two multi-perspective visual recognition tasks namely lip reading and face recognition. Three relevant datasets are considered and the results are compared against fusion strategies, other existing multi-input LSTM architectures, and alternative recognition solutions. The experiments show the superior performance of our solution over the considered benchmarks, both in terms of recognition accuracy and complexity. We make our code publicly available at: https://github.com/arsm/MPLSTM

We present a new end-to-end learning framework to obtain detailed and spatially coherent reconstructions of multiple people from a single image. Existing multi-person methods suffer from two main drawbacks: they are often model-based and therefore cannot capture accurate 3D models of people with loose clothing and hair; or they require manual intervention to resolve occlusions or interactions. Our method addresses both limitations by introducing the first end-to-end learning approach to perform model-free implicit reconstruction for realistic 3D capture of multiple clothed people in arbitrary poses (with occlusions) from a single image. Our network simultaneously estimates the 3D geometry of each person and their 6DOF spatial locations, to obtain a coherent multi-human reconstruction. In addition, we introduce a new synthetic dataset that depicts images with a varying number of inter-occluded humans in a variety of clothing and hair. We demonstrate robust, high-resolution reconstructions on images of multiple humans with complex occlusions, loose clothing and a large variety of poses, and scenes. Our quantitative evaluation on both synthetic and real world datasets demonstrates state-of-the-art performance with significant improvements in the accuracy and completeness of the reconstructions over competing approaches.

Deep neural networks trained with standard cross-entropy loss memorize noisy labels, which degrades their performance. Most research to mitigate this memorization proposes new robust classification loss functions. Conversely, we propose a Multi-Objective Interpolation Training (MOIT) approach that jointly exploits contrastive learning and classification to mutually help each other and boost performance against label noise. We show that standard supervised contrastive learning degrades in the presence of label noise and propose an interpolation training strategy to mitigate this behavior. We further propose a novel label noise detection method that exploits the robust feature representations learned via contrastive learning to estimate per-sample soft-labels whose disagreements with the original labels accurately identify noisy samples. This detection allows treating noisy samples as unlabeled and training a classifier in a semi-supervised manner to prevent noise memorization and improve representation learning. We further propose MOIT+, a refinement of MOIT by fine-tuning on detected clean samples. Hyperparameter and ablation studies verify the key components of our method. Experiments on synthetic and real-world noise benchmarks demonstrate that MOIT/MOIT+ achieves state-of-the-art results. Code is available at https://git.io/JI40X.

Given an untrimmed video and a query sentence, cross-modal video moment retrieval aims to rank a video moment from pre-segmented video moment candidates that best matches the query sentence. Pioneering work typically learns the representations of the textual and visual content separately and then obtains the interactions or alignments between different modalities. However, the task of cross-modal video moment retrieval is not yet thoroughly addressed as it needs to further identify the fine-grained differences of video moment candidates with high repeatability and similarity. Moveover, the relation among objects in both video and query sentence is intuitive and efficient for understanding semantics but is rarely considered. Toward this end, we contribute a multi-modal relational graph to capture the interactions among objects from the visual and textual content to identify the differences among similar video moment candidates. Specifically, we first introduce a visual relational graph and a textual relational graph to form relation-aware representations via message propagation. Thereafter, a multi-task pre-training is designed to capture domain-specific knowledge about objects and relations, enhancing the structured visual representation after explicitly defined relation. Finally, the graph matching and boundary regression are employed to perform the cross-modal retrieval. We conduct extensive experiments on two datasets about daily activities and cooking activities, demonstrating significant improvements over state-of-the-art solutions.

How should representations from complementary sensors be integrated for autonomous driving? Geometry-based sensor fusion has shown great promise for perception tasks such as object detection and motion forecasting. However, for the actual driving task, the global context of the 3D scene is key, e.g. a change in traffic light state can affect the behavior of a vehicle geometrically distant from that traffic light. Geometry alone may therefore be insufficient for effectively fusing representations in end-to-end driving models. In this work, we demonstrate that imitation learning policies based on existing sensor fusion methods under-perform in the presence of a high density of dynamic agents and complex scenarios, which require global contextual reasoning, such as handling traffic oncoming from multiple directions at uncontrolled intersections. Therefore, we propose TransFuser, a novel Multi-Modal Fusion Transformer, to integrate image and LiDAR representations using attention. We experimentally validate the efficacy of our approach in urban settings involving complex scenarios using the CARLA urban driving simulator. Our approach achieves state-of-the-art driving performance while reducing collisions by 76% compared to geometry-based fusion.

Predicting all applicable labels for a given image is known as multi-label classification. Compared to the standard multi-class case (where each image has only one label), it is considerably more challenging to annotate training data for multi-label classification. When the number of potential labels is large, human annotators find it difficult to mention all applicable labels for each training image. Furthermore, in some settings detection is intrinsically difficult e.g. finding small object instances in high resolution images. As a result, multi-label training data is often plagued by false negatives. We consider the hardest version of this problem, where annotators provide only one relevant label for each image. As a result, training sets will have only one positive label per image and no confirmed negatives. We explore this special case of learning from missing labels across four different multi-label image classification datasets for both linear classifiers and end-to-end fine-tuned deep networks. We extend existing multi-label losses to this setting and propose novel variants that constrain the number of expected positive labels during training. Surprisingly, we show that in some cases it is possible to approach the performance of fully labeled classifiers despite training with significantly fewer confirmed labels.

Multi-label activity recognition is designed for recognizing multiple activities that are performed simultaneously or sequentially in each video. Most recent activity recognition networks focus on single-activities, that assume only one activity in each video. These networks extract shared features for all the activities, which are not designed for multi-label activities. We introduce an approach to multi-label activity recognition that extracts independent feature descriptors for each activity and learns activity correlations. This structure can be trained end-to-end and plugged into any existing network structures for video classification. Our method outperformed state-of-the-art approaches on four multi-label activity recognition datasets. To better understand the activity-specific features that the system generated, we visualized these activity-specific features in the Charades dataset. The code will be released later.

Fast and accurate reconstruction of magnetic resonance (MR) images from under-sampled data is important in many clinical applications. In recent years, deep learning-based methods have been shown to produce superior performance on MR image reconstruction. However, these methods require large amounts of data which is difficult to collect and share due to the high cost of acquisition and medical data privacy regulations. In order to overcome this challenge, we propose a federated learning (FL) based solution in which we take advantage of the MR data available at different institutions while preserving patients' privacy. However, the generalizability of models trained with the FL setting can still be suboptimal due to domain shift, which results from the data collected at multiple institutions with different sensors, disease types, and acquisition protocols, etc. With the motivation of circumventing this challenge, we propose a cross-site modeling for MR image reconstruction in which the learned intermediate latent features among different source sites are aligned with the distribution of the latent features at the target site. Extensive experiments are conducted to provide various insights about FL for MR image reconstruction. Experimental results demonstrate that the proposed framework is a promising direction to utilize multi-institutional data without compromising patients' privacy for achieving improved MR image reconstruction. Our code is available at https://github.com/guopengf/FL-MRCM

Existing deep deraining models are mainly learned via directly minimizing the statistical differences between rainy images and rain-free ground truths. They emphasize learning a mapping from rainy images to rain-free images with supervision. Despite the demonstrated success, these methods do not perform well on restoring the fine-grained local details or removing blurry rainy traces. In this work, we aim to exploit the intrinsic priors of rainy images and develop intrinsic loss functions to facilitate training deraining networks, which decompose a rainy image into a rain-free background layer and a rainy layer containing intact rain streaks. To this end, we introduce the quasi-sparsity prior to train network so as to generate two sparse layers with intact textures of different objects. Then we explore the low-value prior to compensate sparsity, forcing all rain streaks to enter into one layer while non-rain contents into another layer to restore image details. We introduce a multi-decoding structure to specially supervise the generation of multi-type deraining features. This helps to learn the most contributory features to deraining in respective spaces. Moreover, our model stabilizes the feature values from multi-spaces via information sharing to alleviate potential artifacts, which also accelerates the running speed. Extensive experiments show that the proposed deraining method outperforms the state-of-the-art approaches in terms of effectiveness and efficiency.

Face forgery by deepfake is widely spread over the internet and has raised severe societal concerns. Recently, how to detect such forgery contents has become a hot research topic and many deepfake detection methods have been proposed. Most of them model deepfake detection as a vanilla binary classification problem, i.e, first use a backbone network to extract a global feature and then feed it into a binary classifier (real/fake). But since the difference between the real and fake images in this task is often subtle and local, we argue this vanilla solution is not optimal. In this paper, we instead formulate deepfake detection as a fine-grained classification problem and propose a new multi-attentional deepfake detection network. Specifically, it consists of three key components: 1) multiple spatial attention heads to make the network attend to different local parts; 2) textural feature enhancement block to zoom in the subtle artifacts in shallow features; 3) aggregate the low-level textural feature and high-level semantic features guided by the attention maps. Moreover, to address the learning difficulty of this network, we further introduce a new regional independence loss and an attention guided data augmentation strategy. Through extensive experiments on different datasets, we demonstrate the superiority of our method over the vanilla binary classifier counterparts, and achieve state-of-the-art performance.

This paper tackles video prediction from a new dimension of predicting spacetime-varying motions that are incessantly changing across both space and time. Prior methods mainly capture the temporal state transitions but overlook the complex spatiotemporal variations of the motion itself, making them difficult to adapt to ever-changing motions. We observe that physical world motions can be decomposed into transient variation and motion trend, while the latter can be regarded as the accumulation of previous motions. Thus, simultaneously capturing the transient variation and the motion trend is the key to make spacetime-varying motions more predictable. Based on these observations, we propose the MotionRNN framework, which can capture the complex variations within motions and adapt to spacetime-varying scenarios. MotionRNN has two main contributions. The first is that we design the MotionGRU unit, which can model the transient variation and motion trend in a unified way. The second is that we apply the MotionGRU to RNN-based predictive models and indicate a new flexible video prediction architecture with a Motion Highway that can significantly improve the ability to predict changeable motions and avoid motion vanishing for stacked multiple-layer predictive models. With high flexibility, this framework can adapt to a series of models for deterministic spatiotemporal prediction. Our MotionRNN can yield significant improvements on three challenging benchmarks for video prediction with spacetime-varying motions.

We propose novel motion representations for animating articulated objects consisting of distinct parts. In a completely unsupervised manner, our method identifies object parts, tracks them in a driving video, and infers their motions by considering their principal axes. In contrast to the previous keypoint-based works, our method extracts meaningful and consistent regions, describing locations, shape, and pose. The regions correspond to semantically relevant and distinct object parts, that are more easily detected in frames of the driving video. To force decoupling of foreground from background, we model non-object related global motion with an additional affine transformation. To facilitate animation and prevent the leakage of the shape of the driving object, we disentangle shape and pose of objects in the region space. Our model can animate a variety of objects, surpassing previous methods by a large margin on existing benchmarks. We present a challenging new benchmark with high-resolution videos and show that the improvement is particularly pronounced when articulated objects are considered, reaching 96.6% user preference vs. the state of the art.

A fundamental challenge faced by existing Fine-Grained Sketch-Based Image Retrieval (FG-SBIR) models is the data scarcity -- model performances are largely bottlenecked by the lack of sketch-photo pairs. Whilst the number of photos can be easily scaled, each corresponding sketch still needs to be individually produced. In this paper, we aim to mitigate such an upper-bound on sketch data, and study whether unlabelled photos alone (of which they are many) can be cultivated for performance gain. In particular, we introduce a novel semi-supervised framework for cross-modal retrieval that can additionally leverage large-scale unlabelled photos to account for data scarcity. At the center of our semi-supervision design is a sequential photo-to-sketch generation model that aims to generate paired sketches for unlabelled photos. Importantly, we further introduce a discriminator-guided mechanism to guide against unfaithful generation, together with a distillation loss-based regularizer to provide tolerance against noisy training samples. Last but not least, we treat generation and retrieval as two conjugate problems, where a joint learning procedure is devised for each module to mutually benefit from each other. Extensive experiments show that our semi-supervised model yields a significant performance boost over the state-of-the-art supervised alternatives, as well as existing methods that can exploit unlabelled photos for FG-SBIR.