CVPR 2022 Papers

Recent high-performing Human-Object Interaction (HOI) detection techniques have been highly influenced by Transformer-based object detector (i.e., DETR). Nevertheless, most of them directly map parametric interaction queries into a set of HOI predictions through vanilla Transformer in a one-stage manner. This leaves rich inter- or intra-interaction structure under-exploited. In this work, we design a novel Transformer-style HOI detector, i.e., Structure-aware Transformer over Interaction Proposals (STIP), for HOI detection. Such design decomposes the process of HOI set prediction into two subsequent phases, i.e., an interaction proposal generation is first performed, and then followed by transforming the non-parametric interaction proposals into HOI predictions via a structure-aware Transformer. The structure-aware Transformer upgrades vanilla Transformer by encoding additionally the holistically semantic structure among interaction proposals as well as the locally spatial structure of human/object within each interaction proposal, so as to strengthen HOI predictions. Extensive experiments conducted on V-COCO and HICO-DET benchmarks have demonstrated the effectiveness of STIP, and superior results are reported when comparing with the state-of-the-art HOI detectors. Source code is available at https://github.com/zyong812/STIP.

By considering the spatial correspondence, dense self-supervised representation learning has achieved superior performance on various dense prediction tasks. However, the pixel-level correspondence tends to be noisy because of many similar misleading pixels, e.g., backgrounds. To address this issue, in this paper, we propose to explore set similarity (SetSim) for dense self-supervised representation learning. We generalize pixel-wise similarity learning to set-wise one to improve the robustness because sets contain more semantic and structure information. Specifically, by resorting to attentional features of views, we establish the corresponding set, thus filtering out noisy backgrounds that may cause incorrect correspondences. Meanwhile, these attentional features can keep the coherence of the same image across different views to alleviate semantic inconsistency. We further search the cross-view nearest neighbours of sets and employ the structured neighbourhood information to enhance the robustness. Empirical evaluations demonstrate that SetSim surpasses or is on par with state-of-the-art methods on object detection, keypoint detection, instance segmentation, and semantic segmentation.

Recently, contrastive learning-based image translation methods have been proposed, which contrasts different spatial locations to enhance the spatial correspondence. However, the methods often ignore the diverse semantic relation within the images. To address this, here we propose a novel semantic relation consistency (SRC) regularization along with the decoupled contrastive learning (DCL), which utilize the diverse semantics by focusing on the heterogeneous semantics between the image patches of a single image. To further improve the performance, we present a hard negative mining by exploiting the semantic relation. We verified our method for three tasks: single-modal and multi-modal image translations, and GAN compression task for image translation. Experimental results confirmed the state-of-art performance of our method in all the three tasks.

This paper investigates the geometric consistency for monocular 3D object detection, which suffers from the ill-posed depth estimation. We first conduct a thorough analysis to reveal how existing methods fail to consistently localize objects when different geometric shifts occur. In particular, we design a series of geometric manipulations to diagnose existing detectors and then illustrate their vulnerability to consistently associate the depth with object apparent sizes and positions. To alleviate this issue, we propose four geometry-aware data augmentation approaches to enhance the geometric consistency of the detectors. We first modify some commonly used data augmentation methods for 2D images so that they can maintain geometric consistency in 3D spaces. We demonstrate such modifications are important. In addition, we propose a 3D-specific image perturbation method that employs the camera movement. During the augmentation process, the camera system with the corresponding image is manipulated, while the geometric visual cues for depth recovery are preserved. We show that by using the geometric consistency constraints, the proposed augmentation techniques lead to improvements on the KITTI and nuScenes monocular 3D detection benchmarks with state-of-the-art results. In addition, we demonstrate that the augmentation methods are well suited for semi-supervised training and cross-dataset generalization.

Various facial manipulation techniques have drawn serious public concerns in morality, security, and privacy. Although existing face forgery classifiers achieve promising performance on detecting fake images, these methods are vulnerable to adversarial examples with injected imperceptible perturbations on the pixels. Meanwhile, many face forgery detectors always utilize the frequency diversity between real and fake faces as a crucial clue. In this paper, instead of injecting adversarial perturbations into the spatial domain, we propose a frequency adversarial attack method against face forgery detectors. Concretely, we apply discrete cosine transform (DCT) on the input images and introduce a fusion module to capture the salient region of adversary in the frequency domain. Compared with existing adversarial attacks (e.g. FGSM, PGD) in the spatial domain, our method is more imperceptible to human observers and does not degrade the visual quality of the original images. Moreover, inspired by the idea of meta-learning, we also propose a hybrid adversarial attack that performs attacks in both the spatial and frequency domains. Extensive experiments indicate that the proposed method fools not only the spatial-based detectors but also the state-of-the-art frequency-based detectors effectively. In addition, the proposed frequency attack enhances the transferability across face forgery detectors as black-box attacks.

Existing cross-domain detection methods mostly study the domain shifts where differences between domains are often caused by external environment and perceivable for humans. However, in real-world scenarios (e.g., MRI medical diagnosis, X-ray security inspection), there still exists another type of shift, named endogenous shift, where the differences between domains are mainly caused by the intrinsic factors (e.g., imaging mechanisms, hardware components, etc.), and usually inconspicuous. This shift can also severely harm the cross-domain detection performance but has been rarely studied. To support this study, we contribute the first Endogenous Domain Shift (EDS) benchmark, X-ray security inspection, where the endogenous shifts among the domains are mainly caused by different X-ray machine types with different hardware parameters, wear degrees, etc. EDS consists of 14,219 images including 31,654 common instances from three domains (X-ray machines), with bounding-box annotations from 10 categories. To handle the endogenous shift, we further introduce the Perturbation Suppression Network (PSN), motivated by the fact that this shift is mainly caused by two types of perturbations: category-dependent and category-independent ones. PSN respectively exploits local prototype alignment and global adversarial learning mechanism to suppress these two types of perturbations. The comprehensive evaluation results show that PSN outperforms SOTA methods, serving a new perspective to the cross-domain research community.

Without access to the training data where a black-box victim model is deployed, training a surrogate model for black-box adversarial attack is still a struggle. In terms of data, we mainly identify three key measures for effective surrogate training in this paper. First, we show that leveraging the loss introduced in this paper to enlarge the inter-class similarity makes more sense than enlarging the inter-class diversity like existing methods. Next, unlike the approaches that expand the intra-class diversity in an implicit model-agnostic fashion, we propose a loss function specific to the surrogate model for our generator to enhance the intra-class diversity. Finally, in accordance with the in-depth observations for the methods based on proxy data, we argue that leveraging the proxy data is still an effective way for surrogate training. To this end, we propose a triple-player framework by introducing a discriminator into the traditional data-free framework. In this way, our method can be competitive when there are few semantic overlaps between the scarce proxy data (with the size between 1k and 5k) and the training data. We evaluate our method on a range of victim models and datasets. The extensive results witness the effectiveness of our method. Our source code is available at https://github.com/xuxiangsun/ST-Data.

Pose Guided Person Image Generation (PGPIG) is the task of transforming a person image from the source pose to a given target pose. Most of the existing methods only focus on the ill-posed source-to-target task and fail to capture reasonable texture mapping. To address this problem, we propose a novel Dual-task Pose Transformer Network (DPTN), which introduces an auxiliary task (i.e., source-tosource task) and exploits the dual-task correlation to promote the performance of PGPIG. The DPTN is of a Siamese structure, containing a source-to-source self-reconstruction branch, and a transformation branch for source-to-target generation. By sharing partial weights between them, the knowledge learned by the source-to-source task can effectively assist the source-to-target learning. Furthermore, we bridge the two branches with a proposed Pose Transformer Module (PTM) to adaptively explore the correlation between features from dual tasks. Such correlation can establish the fine-grained mapping of all the pixels between the sources and the targets, and promote the source texture transmission to enhance the details of the generated target images. Extensive experiments show that our DPTN outperforms state-of-the-arts in terms of both PSNR and LPIPS. In addition, our DPTN only contains 9.79 million parameters, which is significantly smaller than other approaches. Our code is available at: https://github.com/PangzeCheung/Dual-task-Pose-Transformer-Network.

Source-free domain adaptation (SFDA) newly emerges to transfer the relevant knowledge of a well-trained source model to an unlabeled target domain, which is critical in various privacy-preserving scenarios. Most existing methods focus on learning the domain-invariant representations depending solely on the target data, leading to the obtained representations are target-specific. In this way, they cannot fully address the distribution shift problem across domains. In contrast, we provide a fascinating insight: rather than attempting to learn domain-invariant representations, it is better to explore the domain-invariant parameters of the source model. The motivation behind this insight is clear: the domain-invariant representations are dominated by only partial parameters of an available deep source model. We devise the Domain-Invariant Parameter Exploring (DIPE) approach to capture such domain-invariant parameters in the source model to generate domain-invariant representations. A distinguishing method is developed correspondingly for two types of parameters, i.e., domain-invariant and domain-specific parameters, as well as an effective update strategy based on the clustering correction technique and a target hypothesis is proposed. Extensive experiments verify that DIPE successfully exceeds the current state-of-the-art models on many domain adaptation datasets.

Weakly-supervised temporal action localization aims to localize actions in untrimmed videos with only video-level labels. Most existing methods address this problem with a "localization-by-classification" pipeline that localizes action regions based on snippet-wise classification sequences. Snippet-wise classifications are unfortunately error prone due to the sparsity of video-level labels. Inspired by recent success in unsupervised contrastive representation learning, we propose a novel denoised cross-video contrastive algorithm, aiming to enhance the feature discrimination ability of video snippets for accurate temporal action localization in the weakly-supervised setting. This is enabled by three key designs: 1) an effective pseudo-label denoising module to alleviate the side effects caused by noisy contrastive features, 2) an efficient region-level feature contrast strategy with a region-level memory bank to capture "global" contrast across the entire dataset, and 3) a diverse contrastive learning strategy to enable action-background separation as well as intra-class compactness & inter-class separability. Extensive experiments on THUMOS14 and ActivityNet v1.3 demonstrate the superior performance of our approach.

We endeavor on a rarely explored task named Insubstan-tial Object Detection (IOD), which aims to localize the object with following characteristics: (1) amorphous shape with indistinct boundary; (2) similarity to surroundings; (3) absence in color. Accordingly, it is far more challenging to distinguish insubstantial objects in a single static frame and the collaborative representation of spatial and tempo-ral information is crucial. Thus, we construct an IOD-Video dataset comprised of 600 videos (141,017 frames) covering various distances, sizes, visibility, and scenes captured by different spectral ranges. In addition, we develop a spatio-temporal aggregation framework for IOD, in which differ-ent backbones are deployed and a spatio-temporal aggregation loss (STAloss) is elaborately designed to leverage the consistency along the time axis. Experiments conducted on IOD-Video dataset demonstrate that spatio-temporal aggregation can significantly improve the performance of IOD. We hope our work will attract further researches into this valuable yet challenging task. The code will be available at: https://github.com/CalayZhou/IOD-Video.

We consider the object recognition problem in autonomous driving using automotive radar sensors. Comparing to Lidar sensors, radar is cost-effective and robust in all-weather conditions for perception in autonomous driving. However, radar signals suffer from low angular resolution and precision in recognizing surrounding objects. To enhance the capacity of automotive radar, in this work, we exploit the temporal information from successive ego-centric bird-eye-view radar image frames for radar object recognition. We leverage the consistency of an object's existence and attributes (size, orientation, etc.), and propose a temporal relational layer to explicitly model the relations between objects within successive radar images. In both object detection and multiple object tracking, we show the superiority of our method compared to several baseline approaches.

Previous LiDAR scene flow estimation methods, especially recurrent neural networks, usually suffer from structure distortion in challenging cases, such as sparse reflection and motion occlusions. In this paper, we propose a novel optimization method based on a recurrent neural network to predict LiDAR scene flow in a weakly supervised manner. Specifically, our neural recurrent network exploits direct rigidity constraints to preserve the geometric structure of the warped source scene during an iterative alignment procedure. An error awarded optimization strategy is proposed to update the LiDAR scene flow by minimizing the point measurement error instead of reconstructing the cost volume multiple times. Trained on two autonomous driving datasets, our network outperforms recent state-of-the-art networks on lidarKITTI by a large margin. The code and models will be available at https://github. com/gtdong-ustc/LiDARSceneFlow.

We present a novel self-distillation based self-supervised monocular depth estimation (SD-SSMDE) learning framework. In the first step, our network is trained in a self-supervised regime on high-resolution images with the photometric loss. The network is further used to generate pseudo depth labels for all the images in the training set. To improve the performance of our estimates, in the second step, we re-train the network with the scale invariant logarithmic loss supervised by pseudo labels. We resolve scale ambiguity and inter-frame scale consistency by introducing an automatically computed scale in our depth labels. To filter out noisy depth values, we devise a filtering scheme based on the 3D consistency between consecutive views. Extensive experiments demonstrate that each proposed component and the self-supervised learning framework improve the quality of the depth estimation over the baseline and achieve state-of-the-art results on the KITTI and Cityscapes datasets.

Explaining the generalization characteristics of deep learning is an emerging topic in advanced machine learning. There are several unanswered questions about how learning under stochastic optimization really works and why certain strategies are better than others. In this paper, we address the following question: can we probe intermediate layers of a deep neural network to identify and quantify the learning quality of each layer? With this question in mind, we propose new explainability metrics that measure the redundant information in a network's layers using a low-rank factorization framework and quantify a complexity measure that is highly correlated with the generalization performance of a given optimizer, network, and dataset. We subsequently exploit these metrics to augment the Stochastic Gradient Descent (SGD) optimizer by adaptively adjusting the learning rate in each layer to improve in generalization performance. Our augmented SGD -- dubbed RMSGD -- introduces minimal computational overhead compared to SOTA methods and outperforms them by exhibiting strong generalization characteristics across application, architecture, and dataset.

Interpretability is an important property for visual models as it helps researchers and users understand the internal mechanism of a complex model. However, generating semantic explanations about the learned representation is challenging without direct supervision to produce such explanations. We propose a general framework, Latent Visual Semantic Explainer (LaViSE), to teach any existing convolutional neural network to generate text descriptions about its own latent representations at the filter level. Our method constructs a mapping between the visual and semantic spaces using generic image datasets, using images and category names. It then transfers the mapping to the target domain which does not have semantic labels. The proposed framework employs a modular structure and enables to analyze any trained network whether or not its original training data is available. We show that our method can generate novel descriptions for learned filters beyond the set of categories defined in the training dataset and perform an extensive evaluation on multiple datasets. We also demonstrate a novel application of our method for unsupervised dataset bias analysis which allows us to automatically discover hidden biases in datasets or compare different subsets without using additional labels. The dataset and code are made public to facilitate further research.

Modern object detectors have achieved impressive progress under the close-set setup. However, open-set object detection (OSOD) remains challenging since objects of unknown categories are often misclassified to existing known classes. In this work, we propose to identify unknown objects by separating high/low-density regions in the latent space, based on the consensus that unknown objects are usually distributed in low-density latent regions. As traditional threshold-based methods only maintain limited low-density regions, which cannot cover all unknown objects, we present a novel Open-set Detector (OpenDet) with expanded low-density regions. To this aim, we equip OpenDet with two learners, Contrastive Feature Learner (CFL) and Unknown Probability Learner (UPL). CFL performs instance-level contrastive learning to encourage compact features of known classes, leaving more low-density regions for unknown classes; UPL optimizes unknown probability based on the uncertainty of predictions, which further divides more low-density regions around the cluster of known classes. Thus, unknown objects in low-density regions can be easily identified with the learned unknown probability. Extensive experiments demonstrate that our method can significantly improve the OSOD performance, e.g., OpenDet reduces the Absolute Open-Set Errors by 25%-35% on six OSOD benchmarks. Code is available at: https://github.com/csuhan/opendet2.

Fine-tuning pre-trained models for downstream tasks is mainstream in deep learning. However, the pre-trained models are limited to be fine-tuned by data from a specific modality. For example, as a visual model, DenseNet cannot directly take the textual data as its input. Hence, although the large pre-trained models such as DenseNet or BERT have a great potential for the downstream recognition tasks, they have weaknesses in leveraging multimodal information, which is a new trend of deep learning. This work focuses on fine-tuning pre-trained unimodal models with multimodal inputs of image-text pairs and expanding them for image-text multimodal recognition. To this end, we propose the Multimodal Information Injection Plug-in (MI2P) which is attached to different layers of the unimodal models (e.g., DenseNet and BERT). The proposed MI2P unit provides the path to integrate the information of other modalities into the unimodal models. Specifically, MI2P performs cross-modal feature transformation by learning the fine-grained correlations between the visual and textual features. Through the proposed MI2P unit, we can inject the language information into the vision backbone by attending the word-wise textual features to different visual channels, as well as inject the visual information into the language backbone by attending the channel-wise visual features to different textual words. Armed with the MI2P attachments, the pre-trained unimodal models can be expanded to process multimodal data without the need to change the network structures.

Synthesis of ergodic, stationary visual patterns is widely applicable in texturing, shape modeling, and digital content creation. The wide applicability of this technique thus requires the pattern synthesis approaches to be scalable, diverse, and authentic. In this paper, we propose an exemplar-based visual pattern synthesis framework that aims to model the inner statistics of visual patterns and generate new, versatile patterns that meet the aforementioned requirements. To this end, we propose an implicit network based on generative adversarial network (GAN) and periodic encoding, thus calling our network the Implicit Periodic Field Network (IPFN). The design of IPFN ensures scalability: the implicit formulation directly maps the input coordinates to features, which enables synthesis of arbitrary size and is computationally efficient for 3D shape synthesis. Learning with a periodic encoding scheme encourages diversity: the network is constrained to model the inner statistics of the exemplar based on spatial latent codes in a periodic field. Coupled with continuously designed GAN training procedures, IPFN is shown to synthesize tileable patterns with smooth transitions and local variations. Last but not least, thanks to both the adversarial training technique and the encoded Fourier features, IPFN learns high-frequency functions that produce authentic, high-quality results. To validate our approach, we present novel experimental results on various applications in 2D texture synthesis and 3D shape synthesis.