CVPR 2021 Papers

We define the concept of CompositeTasking as the fusion of multiple, spatially distributed tasks, for various aspects of image understanding. Learning to perform spatially distributed tasks is motivated by the frequent availability of only sparse labels across tasks, and the desire for a compact multi-tasking network. To facilitate CompositeTasking, we introduce a novel task conditioning model -- a single encoder-decoder network that performs multiple, spatially varying tasks at once. The proposed network takes an image and a set of pixel-wise dense task requests as inputs, and performs the requested prediction task for each pixel. Moreover, we also learn the composition of tasks that needs to be performed according to some CompositeTasking rules, which includes the decision of where to apply which task. It not only offers us a compact network for multi-tasking, but also allows for task-editing. Another strength of the proposed method is demonstrated by only having to supply sparse supervision per task. The obtained results are on par with our baselines that use dense supervision and a multi-headed multi-tasking design. The source code will be made publicly available at www.github.com/nikola3794/composite-tasking.

We show that explicit modeling of composition rules benefits image cropping. Image cropping is considered a promising way to automate aesthetic composition in professional photography. Existing efforts, however, only model such professional knowledge implicitly, e.g., by ranking from comparative candidates. Inspired by the observation that natural composition traits always follow a specific rule, we propose to learn such rules in a discriminative manner, and more importantly, to incorporate learned composition clues explicitly in the model. To this end, we introduce the concept of the key composition map (KCM) to encode the composition rules. The KCM can reveal the common laws hidden behind different composition rules and can inform the cropping model of what is important in composition. With the KCM, we present a novel cropping-by-composition paradigm and instantiate a network to implement composition-aware image cropping. Extensive experiments on two benchmarks justify that our approach enables effective, interpretable, and fast image cropping.

We study an unsupervised domain adaptation problem for the semantic labeling of 3D point clouds, with a particular focus on domain discrepancies induced by different LiDAR sensors. Based on the observation that sparse 3D point clouds are sampled from 3D surfaces, we take a Complete and Label approach to recover the underlying surfaces before passing them to a segmentation network. Specifically, we design a Sparse Voxel Completion Network (SVCN) to complete the 3D surfaces of a sparse point cloud. Unlike semantic labels, to obtain training pairs for SVCN requires no manual labeling. We also introduce local adversarial learning to model the surface prior. The recovered 3D surfaces serve as a canonical domain, from which semantic labels can transfer across different LiDAR sensors. Experiments and ablation studies with our new benchmark for cross-domain semantic labeling of LiDAR data show that the proposed approach provides 6.3-37.6% better performance than previous domain adaptation methods.

Knowledge distillation aims to transfer representation ability from a teacher model to a student model. Previous approaches focus on either individual representation distillation or inter-sample similarity preservation. While we argue that the inter-sample relation conveys abundant information and needs to be distilled in a more effective way. In this paper, we propose a novel knowledge distillation method, namely Complementary Relation Contrastive Distillation (CRCD), to transfer the structural knowledge from the teacher to the student. Specifically, we estimate the mutual relation in an anchor-based way and distill the anchor-student relation under the supervision of its corresponding anchor-teacher relation. To make it more robust, mutual relations are modeled by two complementary elements: the feature and its gradient. Furthermore, the low bound of mutual information between the anchor-teacher relation distribution and the anchor-student relation distribution is maximized via relation contrastive loss, which can distill both the sample representation and the inter-sample relations. Experiments on different benchmarks demonstrate the effectiveness of our proposed CRCD.

We tackle the problem of visual search under resource constraints. Existing systems use the same embedding model to compute representations (embeddings) for the query and gallery images. Such systems inherently face a hard accuracy-efficiency trade-off: the embedding model needs to be large enough to ensure high accuracy, yet small enough to enable query-embedding computation on resource-constrained platforms. This trade-off could be mitigated if gallery embeddings are generated from a large model and query embeddings are extracted using a compact model. The key to building such a system is to ensure representation compatibility between the query and gallery models. In this paper, we address two forms of compatibility: One enforced by modifying the parameters of each model that computes the embeddings. The other by modifying the architectures that compute the embeddings, leading to compatibility-aware neural architecture search (CMP-NAS). We test CMP-NAS on challenging retrieval tasks for fashion images (DeepFashion2), and face images (IJB-C). Compared to ordinary (homogeneous) visual search using the largest embedding model (paragon), CMP-NAS achieves 80-fold and 23-fold cost reduction while maintaining accuracy within 0.3% and 1.6% of the paragon on DeepFashion2 and IJB-C respectively.

Gradient compression has been widely adopted in data-parallel distributed training of deep neural networks to reduce communication overhead. Some literatures have demonstrated that large gradients are more important than small ones because they contain more information, such as Top-k compressor. Other mainstream methods, like random-k compressor and gradient quantization, usually treat all gradients equally. Different from all of them, we regard large and small gradients selection as the exploitation and exploration of gradient information, respectively. And we find taking both of them into consideration is the key to boost the final accuracy. So, we propose a novel gradient compressor: Gradient Sampling with Bayes Prior in this paper. Specifically, we sample important/large gradients based on the global gradient distribution, which is periodically updated across multiple workers. Then we introduce Bayes Prior into distribution model to further explore the gradients. We prove the convergence of our method for smooth non-convex problems in the distributed system. Compared with methods that running after high compression ratio at the expense of accuracy, we pursue no loss of accuracy and the actual acceleration benefit in practice. Experimental comparisons on a variety of computer vision tasks (e.g. image classification and object detection) and backbones (ResNet, MobileNetV2, InceptionV3 and AlexNet) show that our approach outperforms the state-of-the-art techniques in terms of both speed and accuracy, with the limitation of 100* compression ratio.

The goal of cross-view image based geo-localization is to determine the location of a given street view image by matching it against a collection of geo-tagged satellite images. This task is notoriously challenging due to the drastic viewpoint and appearance differences between the two domains. We show that we can address this discrepancy explicitly by learning to synthesize realistic street views from satellite inputs. Following this observation, we propose a novel multi-task architecture in which image synthesis and retrieval are considered jointly. The rationale behind this is that we can bias our network to learn latent feature representations that are useful for retrieval if we utilize them to generate images across the two input domains. To the best of our knowledge, ours is the first approach that creates realistic street views from satellite images and localizes the corresponding query street view simultaneously in an end-to-end manner. In our experiments, we obtain state-of-the-art performance on the CVUSA and CVACT benchmarks. Finally, we show compelling qualitative results for satellite-to-street view synthesis.

Generation of stroke-based non-photorealistic imagery, is an important problem in the computer vision community. As an endeavor in this direction, substantial recent research efforts have been focused on teaching machines "how to paint", in a manner similar to a human painter. However, the applicability of previous methods has been limited to datasets with little variation in position, scale and saliency of the foreground object. As a consequence, we find that these methods struggle to cover the granularity and diversity possessed by real world images. To this end, we propose a Semantic Guidance pipeline with 1) a bi-level painting procedure for learning the distinction between foreground and background brush strokes at training time. 2) We also introduce invariance to the position and scale of the foreground object through a neural alignment model, which combines object localization and spatial transformer networks in an end to end manner, to zoom into a particular semantic instance. 3) The distinguishing features of the in-focus object are then amplified by maximizing a novel guided backpropagation based focus reward. The proposed agent does not require any supervision on human stroke-data and successfully handles variations in foreground object attributes, thus, producing much higher quality canvases for the CUB-200 Birds and Stanford Cars-196 datasets. Finally, we demonstrate the further efficacy of our method on complex datasets with multiple foreground object instances by evaluating an extension of our method on the challenging Virtual-KITTI dataset. Source code and models are available at https://github.com/1jsingh/semantic-guidance.

Most works on person re-identification (ReID) take advantage of large backbone networks such as ResNet, which are designed for image classification instead of ReID, for feature extraction. However, these backbones may not be computationally efficient or the most suitable architectures for ReID. In this work, we aim to design a lightweight and suitable network for ReID. To this end, we propose a novel search space called Combined Depth Space (CDS), based on which we search for an efficient network architecture, which we call CDNet, via a differentiable architecture search algorithm. Through the use of the combined basic building blocks in CDS, CDNet tends to focus on combined pattern information that is typically found in images of pedestrians. We then propose a low-cost search strategy named the Top-k Sample Search strategy to make full use of the search space and avoid trapping in local optimal result. Furthermore, an effective Fine-grained Balance Neck (FBLNeck), which is removable at the inference time, is presented to balance the effects of triplet loss and softmax loss during the training process. Extensive experiments show that our CDNet ( 1.8 M parameters) has comparable performance with state-of-the-art lightweight networks.

Graph Edit Distance (GED) is a popular similarity measurement for pairwise graphs and it also refers to the recovery of the edit path from the source graph to the target graph. Traditional A* algorithm suffers scalability issues due to its exhaustive nature, whose search heuristics heavily rely on human prior knowledge. This paper presents a hybrid approach by combing the interpretability of traditional search-based techniques for producing the edit path, as well as the efficiency and adaptivity of deep embedding models to achieve a cost-effective GED solver. Inspired by dynamic programming, node-level embedding is designated in a dynamic reuse fashion and suboptimal branches are encouraged to be pruned. To this end, our method can be readily integrated into A* procedure in a dynamic fashion, as well as significantly reduce the computational burden with a learned heuristic. Experimental results on different graph datasets show that our approach can remarkably ease the search process of A* without sacrificing much accuracy. To our best knowledge, this work is also the first deep learning-based GED method for recovering the edit path.

Instance segmentation, the task of identifying and separating each individual object of interest in the image, is one of the actively studied research topics in computer vision. Although many feed-forward networks produce high-quality binary segmentation on different types of images, their final result heavily relies on the post-processing step, which separates instances from the binary mask. In comparison, the existing iterative methods extract a single object at a time using discriminative knowledge-based properties (e.g., shapes, boundaries, etc.) without relying on post-processing. However, they do not scale well with a large number of objects. To exploit the advantages of conventional sequential segmentation methods without impairing the scalability, we propose a novel iterative deep reinforcement learning agent that learns how to differentiate multiple objects in parallel. By constructing a relational graph between pixels, we design a reward function that encourages separating pixels of different objects and grouping pixels that belong to the same instance. We demonstrate that the proposed method can efficiently perform instance segmentation of many objects without heavy post-processing.

Language-queried video actor segmentation aims to predict the pixel-level mask of the actor which performs the actions described by a natural language query in the target frames. Existing methods adopt 3D CNNs over the video clip as a general encoder to extract a mixed spatio-temporal feature for the target frame. Though 3D convolutions are amenable to recognizing which actor is performing the queried actions, it also inevitably introduces misaligned spatial information from adjacent frames, which confuses features of the target frame and yields inaccurate segmentation. Therefore, we propose a collaborative spatial-temporal encoder-decoder framework which contains a 3D temporal encoder over the video clip to recognize the queried actions, and a 2D spatial encoder over the target frame to accurately segment the queried actors. In the decoder, a Language-Guided Feature Selection (LGFS) module is proposed to flexibly integrate spatial and temporal features from the two encoders. We also propose a Cross-Modal Adaptive Modulation (CMAM) module to dynamically recombine spatial- and temporal-relevant linguistic features for multimodal feature interaction in each stage of the two encoders. Our method achieves new state-of-the-art performance on two popular benchmarks with less computational overhead than previous approaches.

Conventional stereo suffers from a fundamental trade-off between imaging volume and signal-to-noise ratio (SNR) -- due to the conflicting impact of aperture size on both these variables. Inspired by the extended depth of field cameras, we propose a novel end-to-end learning-based technique to overcome this limitation, by introducing a phase mask at the aperture plane of the cameras in a stereo imaging system. The phase mask creates a depth-dependent point spread function, allowing us to recover sharp image texture and stereo correspondence over a significantly extended depth of field (EDOF) than conventional stereo. The phase mask pattern, the EDOF image reconstruction, and the stereo disparity estimation are all trained together using an end-to-end learned deep neural network. We perform theoretical analysis and characterization of the proposed approach and show a 6x increase in volume that can be imaged in simulation. We also build an experimental prototype and validate the approach using real-world results acquired using this prototype system.

Person re-identification (Re-ID) is to retrieve a particular person captured by different cameras, which is of great significance for security surveillance and pedestrian behavior analysis. However, due to the large intra-class variation of a person across cameras, e.g., occlusions, illuminations, viewpoints, and poses, Re-ID is still a challenging task in the field of computer vision. In this paper, to attack the issues concerning with intra-class variation, we propose a coarse-to-fine Re-ID framework with the incorporation of auxiliary-domain classification (ADC) and second-order information bottleneck (2O-IB). In particular, as an auxiliary task, ADC is introduced to extract the coarse-grained essential features to distinguish a person from miscellaneous backgrounds, which leads to the effective coarse- and fine-grained feature representations for Re-ID. On the other hand, to cope with the redundancy, irrelevance, and noise contained in the Re-ID features caused by intra-class variations, we integrate 2O-IB into the network to compress and optimize the features, without increasing additional computation overhead during inference. Experimental results demonstrate that our proposed method significantly reduces the neural network output variance of intra-class person images and achieves the superior performance to state-of-the-art methods.

Unsupervised domain adaptation (UDA) in semantic segmentation is a fundamental yet promising task relieving the need for laborious annotation works. However, the domain shifts/discrepancies problem in this task compromise the final segmentation performance. Based on our observation, the main causes of the domain shifts are differences in imaging conditions, called image-level domain shifts, and differences in object category configurations called category-level domain shifts. In this paper, we propose a novel UDA pipeline that unifies image-level alignment and category-level feature distribution regularization in a coarse-to-fine manner. Specifically, on the coarse side, we propose a photometric alignment module that aligns an image in the source domain with a reference image from the target domain using a set of image-level operators; on the fine side, we propose a category-oriented triplet loss that imposes a soft constraint to regularize category centers in the source domain and a self-supervised consistency regularization method in the target domain. Experimental results show that our proposed pipeline improves the generalization capability of the final segmentation model and significantly outperforms all previous state-of-the-arts.

In this paper, we introduce 'Coarse-Fine Networks', a two-stream architecture which benefits from different abstractions of temporal resolution to learn better video representations for long-term motion. Traditional Video models process inputs at one (or few) fixed temporal resolution without any dynamic frame selection. However, we argue that, processing multiple temporal resolutions of the input and doing so dynamically by learning to estimate the importance of each frame can largely improve video representations, specially in the domain of temporal activity localization. To this end, we propose (1) 'Grid Pool', a learned temporal downsampling layer to extract coarse features, and, (2) 'Multi-stage Fusion', a spatio-temporal attention mechanism to fuse a fine-grained context with the coarse features. We show that our method outperforms the state-of-the-arts for action detection in public datasets including Charades with a significantly reduced compute and memory footprint. The code is available at https://github.com/kkahatapitiya/Coarse-Fine-Networks.

We tackle the task of image retrieval with text feedback, where a reference image and modifier text are combined to identify the desired target image. We focus on designing an image-text compositor, i.e., integrating multi-modal inputs to produce a representation similar to that of the target image. In our algorithm, Content-Style Modulation (CoSMo), we approach this challenge by introducing two modules based on deep neural networks: the content and style modulators. The content modulator performs local updates to the reference image feature after normalizing the style of the image, where a disentangled multi-modal non-local block is employed to achieve the desired content modifications. Then, the style modulator reintroduces global style information to the updated feature. We provide an in-depth view of our algorithm and its design choices, and show that it accomplishes outstanding performance on multiple image-text retrieval benchmarks. Our code can be found at: https://github.com/postBG/CosMo.pytorch

CoMoGAN is a continuous GAN relying on the unsupervised reorganization of the target data on a functional manifold. To that matter, we introduce a new Functional Instance Normalization layer and residual mechanism, which together disentangle image content from position on target manifold. We rely on naive physics-inspired models to guide the training while allowing private model/translations features. CoMoGAN can be used with any GAN backbone and allows new types of image translation, such as cyclic image translation like timelapse generation, or detached linear translation. On all datasets, it outperforms the literature. Our code is available in this page: https://github.com/cv-rits/CoMoGAN.

Weakly-supervised temporal action localization (WS-TAL) aims to localize actions in untrimmed videos with only video-level labels. Most existing models follow the "localization by classification" procedure: locate temporal regions contributing most to the video-level classification. Generally, they process each snippet (or frame) individually and thus overlook the fruitful temporal context relation. Here arises the single snippet cheating issue: "hard" snippets are too vague to be classified. In this paper, we argue that learning by comparing helps identify these hard snippets and we propose to utilize snippet Contrastive learning to Localize Actions, CoLA for short. Specifically, we propose a Snippet Contrast (SniCo) Loss to refine the hard snippet representation in feature space, which guides the network to perceive precise temporal boundaries and avoid the temporal interval interruption. Besides, since it is infeasible to access frame-level annotations, we introduce a Hard Snippet Mining algorithm to locate the potential hard snippets. Substantial analyses verify that this mining strategy efficaciously captures the hard snippets and SniCo Loss leads to more informative feature representation. Extensive experiments show that CoLA achieves state-of-the-art results on THUMOS'14 and ActivityNet v1.2 datasets.