CVPR 2021 Papers

Artistic style transfer is an image editing task that aims at repainting everyday photographs with learned artistic styles. Existing methods learn styles from either a single style example or a collection of artworks. Accordingly, the stylization results are either inferior in visual quality or limited in style controllability. To tackle this problem, we propose a novel Dual Style-Learning Artistic Style Transfer (DualAST) framework to learn simultaneously both the holistic artist-style (from a collection of artworks) and the specific artwork-style (from a single style image): the artist-style sets the tone (i.e., the overall feeling) for the stylized image, while the artwork-style determines the details of the stylized image, such as color and texture. Moreover, we introduce a Style-Control Block (SCB) to adjust the styles of generated images with a set of learnable style-control factors. We conduct extensive experiments to evaluate the performance of the proposed framework, the results of which confirm the superiority of our method.

We address the challenging problem of whole slide image (WSI) classification. WSIs have very high resolutions and usually lack localized annotations. WSI classification can be cast as a multiple instance learning (MIL) problem when only slide-level labels are available. We propose a MIL-based method for WSI classification and tumor detection that does not require localized annotations. Our method has three major components. First, we introduce a novel MIL aggregator that models the relations of the instances in a dual-stream architecture with trainable distance measurement. Second, since WSIs can produce large or unbalanced bags that hinder the training of MIL models, we propose to use self-supervised contrastive learning to extract good representations for MIL and alleviate the issue of prohibitive memory cost for large bags. Third, we adopt a pyramidal fusion mechanism for multiscale WSI features, and further improve the accuracy of classification and localization. Our model is evaluated on two representative WSI datasets. The classification accuracy of our model compares favorably to fully-supervised methods, with less than 2% accuracy gap across datasets. Our results also outperform all previous MIL-based methods. Additional benchmark results on standard MIL datasets further demonstrate the superior performance of our MIL aggregator on general MIL problems.

Remote photoplethysmography (rPPG) based physiological measurement has great application values in health monitoring, emotion analysis, etc. Existing methods mainly focus on how to enhance or extract the very weak blood volume pulse (BVP) signals from face videos, but seldom explicitly model the noises that dominate face video content. Thus, they may suffer from poor generalization ability in unseen scenarios. This paper proposes a novel adversarial learning approach for rPPG based physiological measurement by using Dual Generative Adversarial Networks (Dual-GAN) to model the BVP estimation and noise distribution jointly. The BVP-GAN aims to learn a noise-resistant mapping from input to ground-truth BVP, and the Noise-GAN aims to learn the noise distribution. The dual GANs can promote each other's capability, leading to improved feature disentanglement between BVP and noises. Besides, a plug-and-play block named ROI alignment and fusion (ROI-AF) block is proposed to alleviate the inconsistencies between different ROIs and exploit informative features from a wider receptive field in terms of ROIs. In comparison to state-of-the-art methods, our method achieves better performance in heart rate, heart rate variability, and respiration frequency estimation from face videos.

The dual-pixel (DP) hardware works by splitting each pixel in half and creating an image pair in a single snapshot. Several works estimate depth/inverse depth by treating the DP pair as a stereo pair. However, dual-pixel disparity only occurs in image regions with the defocus blur. The heavy defocus blur in DP pairs affects the performance of matching-based depth estimation approaches. Instead of removing the blur effect blindly, we study the formation of the DP pair which links the blur and the depth information. In this paper, we propose a mathematical DP model which can benefit depth estimation by the blur. These explorations motivate us to propose an end-to-end DDDNet (DP-based Depth and Deblur Network) to jointly estimate the depth and restore the image. Moreover, we define a reblur loss, which reflects the relationship of the DP image formation process with depth information, to regularise our depth estimate in training. To meet the requirement of a large amount of data for learning, we propose the first DP image simulator which allows us to create datasets with DP pairs from any existing RGBD dataset. As a side contribution, we collect a real dataset for further research. Extensive experimental evaluation on both synthetic and real datasets shows that our approach achieves competitive performance compared to state-of-the-art approaches.

We present a new generative autoencoder model with dual contradistinctive losses to improve generative autoencoder that performs simultaneous inference (reconstruction) and synthesis (sampling). Our model, named dual contradistinctive generative autoencoder (DC-VAE), integrates an instance-level discriminative loss (maintaining the instancelevel fidelity for the reconstruction / synthesis) with a set-level adversarial loss (encouraging the set-level fidelity for the reconstruction/synthesis), both being contradistinctive. Extensive experimental results by DC-VAE across different resolutions including 32x32, 64x64, 128x128, and 512x512 are reported. The two contradistinctive losses in VAE work harmoniously in DC-VAE leading to a significant qualitative and quantitative performance enhancement over the baseline VAEs without architectural changes. State-of-the-art or competitive results among generative autoencoders for image reconstruction, image synthesis, image interpolation, and representation learning are observed. DC-VAE is a general-purpose VAE model, applicable to a wide variety of downstream tasks in computer vision and machine learning.

Deep learning models are vulnerable to adversarial examples. As a more threatening type for practical deep learning systems, physical adversarial examples have received extensive research attention in recent years. However, without exploiting the intrinsic characteristics such as model-agnostic and human-specific patterns, existing works generate weak adversarial perturbations in the physical world, which fall short of attacking across different models and show visually suspicious appearance. Motivated by the viewpoint that attention reflects the intrinsic characteristics of the recognition process, this paper proposes the Dual Attention Suppression (DAS) attack to generate visually-natural physical adversarial camouflage with strong transferability by suppressing both model and human attention. As for attacking, we generate transferable adversarial camouflages by distracting the model-shared similar attention patterns from the target to non-target regions. Meanwhile, based on the fact that human visual attention always focuses on salient items (e.g., suspicious distortions), we evade the human-specific bottom-up attention to generate visually-natural camouflage which is correlated to the scenario context. We conduct extensive experiments in both the digital and physical world for classification and detection tasks on up to date models (e.g., Yolo-V5) and significantly demonstrate that our method outperforms state-of-the-art methods.

Gaze target detection aims to infer where each person in a scene is looking. Existing works focus on 2D gaze and 2D saliency, but fail to exploit 3D contexts. In this work, we propose a three-stage method to simulate the human gaze inference behavior in 3D space. In the first stage, we introduce a coarse-to-fine strategy to robustly estimate a 3D gaze orientation from the head. The predicted gaze is decomposed into a planar gaze on the image plane and a depth-channel gaze. In the second stage, we develop a Dual Attention Module (DAM), which takes the planar gaze to produce the filed of view and masks interfering objects regulated by depth information according to the depth-channel gaze. In the third stage, we use the generated dual attention as guidance to perform two sub-tasks: (1) identifying whether the gaze target is inside or out of the image; (2) locating the target if inside. Extensive experiments demonstrate that our approach performs favorably against state-of-the-art methods on GazeFollow and VideoAttentionTarget datasets.

Realistic simulators are critical for training and verifying robotics systems. While most of the contemporary simulators are hand-crafted, a scaleable way to build simulators is to use machine learning to learn how the environment behaves in response to an action, directly from data. In this work, we aim to learn to simulate a dynamic environment directly in pixel-space, by watching unannotated sequences of frames and their associated action pairs. We introduce a novel high-quality neural simulator referred to as DriveGAN that achieves controllability by disentangling different components without supervision. In addition to steering controls, it also includes controls for sampling features of a scene, such as the weather as well as the location of non-player objects. Since DriveGAN is a fully differentiable simulator, it further allows for re-simulation of a given video sequence, offering an agent to drive through a recorded scene again, possibly taking different actions. We train DriveGAN on multiple datasets, including 160 hours of real-world driving data. We showcase that our approach greatly surpasses the performance of previous data-driven simulators, and allows for new features not explored before.

Artistic style transfer aims at migrating the style from an example image to a content image. Currently, optimization-based methods have achieved great stylization quality, but expensive time cost restricts their practical applications. Meanwhile, feed-forward methods still fail to synthesize complex style, especially when holistic global and local patterns exist. Inspired by the common painting process of drawing a draft and revising the details, we introduce a novel feed-forward method Laplacian Pyramid Network (LapStyle). LapStyle first transfers global style pattern in low-resolution via a Drafting Network. It then revises the local details in high-resolution via a Revision Network, which hallucinates a residual image according to the draft and the image textures extracted by Laplacian filtering. Higher resolution details can be easily generated by stacking Revision Networks with multiple Laplacian pyramid levels. The final stylized image is obtained by aggregating outputs of all pyramid levels. Experiments demonstrate that our method can synthesize high quality stylized images in real time, where holistic style patterns are properly transferred.

This paper presents a novel, simple yet robust self-representation method, i.e., Double Low-Rank Representation with Projection Distance penalty (DLRRPD) for clustering. With the learned optimal projected representations, DLRRPD is capable of obtaining an effective similarity graph to capture the multi-subspace structure. Besides the global low-rank constraint, the local geometrical structure is additionally exploited via a projection distance penalty in our DLRRPD, thus facilitating a more favorable graph. Moreover, to improve the robustness of DLRRPD to noises, we introduce a Laplacian rank constraint, which can further encourage the learned graph to be more discriminative for clustering tasks. Meanwhile, Frobenius norm (instead of the popularly used nuclear norm) is employed to enforce the graph to be more block-diagonal with lower complexity. Extensive experiments have been conducted on synthetic, real, and noisy data to show that the proposed method outperforms currently available alternatives by a margin of 1.0% 10.1%.

Domain adaptation methods face performance degradation in object detection, as the complexity of tasks require more about the transferability of the model. We propose a new perspective on how CNN models gain the transferability, viewing the weights of a model as a series of motion patterns. The directions of weights, and the gradients, can be divided into domain-specific and domain-invariant parts, and the goal of domain adaptation is to concentrate on the domain-invariant direction while eliminating the disturbance from domain-specific one. Current UDA object detection methods view the two directions as a whole while optimizing, which will cause domain-invariant direction mismatch even if the output features are perfectly aligned. In this paper, we propose the domain-specific suppression, an exemplary and generalizable constraint to the original convolution gradients in backpropagation to detach the two parts of directions and suppress the domain-specific one. We further validate our theoretical analysis and methods on several domain adaptive object detection tasks, including weather, camera configuration, and synthetic to real-world adaptation. Our experiment results show significant advance over the state-of-the-art methods in the UDA object detection field, performing a promotion of 10.2 12.2% mAP on all these domain adaptation scenarios.

The observation that computer vision methods overfit to dataset specifics has inspired diverse attempts to make object recognition models robust to domain shifts. However, similar work on domain-robust visual question answering methods is very limited. Domain adaptation for VQA differs from adaptation for object recognition due to additional complexity: VQA models handle multimodal inputs, methods contain multiple steps with diverse modules resulting in complex optimization, and answer spaces in different datasets are vastly different. To tackle these challenges, we first quantify domain shifts between popular VQA datasets, in both visual and textual space. To disentangle shifts between datasets arising from different modalities, we also construct synthetic shifts in the image and question domains separately. Second, we test the robustness of different families of VQA methods (classic two-stream, transformer, and neuro-symbolic methods) to these shifts. Third, we test the applicability of existing domain adaptation methods and devise a new one to bridge VQA domain gaps, adjusted to specific VQA models. To emulate the setting of real-world generalization, we focus on unsupervised domain adaptation and the open-ended classification task formulation.

From a simplified analysis of adaptive methods, we derive AvaGrad, a new optimizer which outperforms SGD on vision tasks when its adaptability is properly tuned. We observe that the power of our method is partially explained by a decoupling of learning rate and adaptability, greatly simplifying hyperparameter search. In light of this observation, we demonstrate that, against conventional wisdom, Adam can also outperform SGD on vision tasks, as long as the coupling between its learning rate and adaptability is taken into account. In practice, AvaGrad matches the best results, as measured by generalization accuracy, delivered by any existing optimizer (SGD or adaptive) across image classification (CIFAR, ImageNet) and character-level language modelling (Penn Treebank) tasks. When training GANs, AvaGrad improves upon existing optimizers.

In this paper, we investigate Universal Domain Adaptation (UniDA) problem, which aims to transfer the knowledge from source to target under unaligned label space. The main challenge of UniDA lies in how to separate common classes (i.e., classes shared across domains), from private classes (i.e., classes only exist in one domain). Previous works treat the private samples in the target as one generic class but ignore their intrinsic structure. Consequently, the resulting representations are not compact enough in the latent space and can be easily confused with common samples. To better exploit the intrinsic structure of the target domain, we propose Domain Consensus Clustering(DCC), which exploits the domain consensus knowledge to discover discriminative clusters on both common samples and private ones. Specifically, we draw the domain consensus knowledge from two aspects to facilitate the clustering and the private class discovery, i.e., the semantic-level consensus, which identifies the cycle-consistent clusters as the common classes, and the sample-level consensus, which utilizes the cross-domain classification agreement to determine the number of clusters and discover the private classes. Based on DCC, we are able to separate the private classes from the common ones, and differentiate the private classes themselves. Finally, we apply a class-aware alignment technique on identified common samples to minimize the distribution shift, and a prototypical regularizer to inspire discriminative target clusters. Experiments on four benchmarks demonstrate DCC significantly outperforms previous state-of-the-arts.

Domain adaptation (DA) aims to transfer knowledge from a label-rich but heterogeneous domain to a label-scare domain, which alleviates the labeling efforts and attracts considerable attention. Different from previous methods focusing on learning domain-invariant feature representations, some recent methods present generic semi-supervised learning (SSL) techniques and directly apply them to DA tasks, even achieving competitive performance. One of the most popular SSL techniques is pseudo-labeling that assigns pseudo labels for each unlabeled data via the classifier trained by labeled data. However, it ignores the distribution shift in DA problems and is inevitably biased to source data. To address this issue, we propose a new pseudo-labeling framework called Auxiliary Target Domain-Oriented Classifier (ATDOC). ATDOC alleviates the classifier bias by introducing an auxiliary classifier for target data only, to improve the quality of pseudo labels. Specifically, we employ the memory mechanism and develop two types of non-parametric classifiers, i.e. the nearest centroid classifier and neighborhood aggregation, without introducing any additional network parameters. Despite its simplicity in a pseudo classification objective, ATDOC with neighborhood aggregation significantly outperforms domain alignment techniques and prior SSL techniques on a large variety of DA benchmarks and even scare-labeled SSL tasks.

As airborne vehicles are becoming more autonomous and ubiquitous, it has become vital to develop the capability to detect the objects in their surroundings. This paper attempts to address the problem of drones detection from other flying drones. The erratic movement of the source and target drones, small size, arbitrary shape, large intensity variations, and occlusion make this problem quite challenging. In this scenario, region-proposal based methods are not able to capture sufficient discriminative foreground-background information. Also, due to the extremely small size and complex motion of the source and target drones, feature aggregation based methods are unable to perform well. To handle this, instead of using region-proposal based methods, we propose to use a two-stage segmentation-based approach employing spatio-temporal attention cues. During the first stage, given the overlapping frame regions, detailed contextual information is captured over convolution feature maps using pyramid pooling. After that pixel and channel-wise attention is enforced on the feature maps to ensure accurate drone localization. In the second stage, first stage detections are verified and new probable drone locations are explored. To discover new drone locations, motion boundaries are used. This is followed by tracking candidate drone detections for a few frames, cuboid formation, extraction of the 3D convolution feature map, and drones detection within each cuboid. The proposed approach is evaluated on two publicly available drone detection datasets and outperforms over several competitive baselines.

Due to the intensive cost of labor and expertise in annotating 3D medical images at a voxel level, most benchmark datasets are equipped with the annotations of only one type of organs and/or tumors, resulting in the so-called partially labeling issue. To address this issue, we propose a dynamic on-demand network (DoDNet) that learns to segment multiple organs and tumors on partially labeled datasets. DoDNet consists of a shared encoder-decoder architecture, a task encoding module, a controller for dynamic filter generation, and a single but dynamic segmentation head. The information of current segmentation task is encoded as a task-aware prior to tell the model what the task is expected to achieve. Different from existing approaches which fix kernels after training, the kernels in dynamic head are generated adaptively by the controller, conditioned on both input image and assigned task. Thus, DoDNet is able to segment multiple organs and tumors, as done by multiple networks or a multi-head network, in a much efficient and flexible manner. We created a large-scale partially labeled dataset called MOTS and demonstrated the superior performance of our DoDNet over other competitors on seven organ and tumor segmentation tasks. We also transferred the weights pre-trained on MOTS to a downstream multi-organ segmentation task and achieved state-of-the-art performance. This study provides a general 3D medical image segmentation model that has been pre-trained on a large-scale partially labeled dataset and can be extended (after fine-tuning) to downstream volumetric medical data segmentation tasks. Code and models are available at https://git.io/DoDNet.

Trajectory prediction is a safety-critical tool for autonomous vehicles to plan and execute actions. Our work addresses two key challenges in trajectory prediction, learning multimodal outputs, and better predictions by imposing constraints using driving knowledge. Recent methods have achieved strong performances using Multi-Choice Learning objectives like winner-takes-all (WTA) or best-of-many. But the impact of those methods in learning diverse hypotheses is under-studied as such objectives highly depend on their initialization for diversity. As our first contribution, we propose a novel Divide-And-Conquer (DAC) approach that acts as a better initialization technique to WTA objective, resulting in diverse outputs without any spurious modes. Our second contribution is a novel trajectory prediction framework called ALAN that uses existing lane centerlines as anchors to provide trajectories constrained to the input lanes. Our framework provides multi-agent trajectory outputs in a forward pass by capturing interactions through hypercolumn descriptors and incorporating scene information in the form of rasterized images and per-agent lane anchors. Experiments on synthetic and real data show that the proposed DAC captures the data distribution better compare to other WTA family of objectives. Further, we show that our ALAN approach provides on par or better performance with SOTA methods evaluated on Nuscenes urban driving benchmark.

Quantization has emerged as one of the most prevalent approaches to compress and accelerate neural networks. Recently, data-free quantization has been widely studied as a practical and promising solution. It synthesizes data for calibrating the quantized model according to the batch normalization (BN) statistics of FP32 ones and significantly relieves the heavy dependency on real training data in traditional quantization methods. Unfortunately, we find that in practice, the synthetic data identically constrained by BN statistics suffers serious homogenization at both distribution level and sample level and further causes a significant performance drop of the quantized model. We propose Diverse Sample Generation (DSG) scheme to mitigate the adverse effects caused by homogenization. Specifically, we slack the alignment of feature statistics in the BN layer to relax the constraint at the distribution level and design a layerwise enhancement to reinforce specific layers for different data samples. Our DSG scheme is versatile and even able to be applied to the state-of-the-art post-training quantization method like AdaRound. We evaluate the DSG scheme on the large-scale image classification task and consistently obtain significant improvements over various network architectures and quantization methods, especially when quantized to lower bits (e.g., up to 22% improvement on W4A4). Moreover, benefiting from the enhanced diversity, models calibrated by synthetic data perform close to those calibrated by real data and even outperform them on W4A4.