CVPR 2022 Papers

Deep image inpainting can inpaint a corrupted image using a feed-forward inference, but still fails to handle large missing area or complex semantics. Recently, GAN inversion based inpainting methods propose to leverage semantic information in pretrained generator (e.g., StyleGAN) to solve the above issues. Different from feed-forward methods, they seek for a closest latent code to the corrupted image and feed it to a pretrained generator. However, inferring the latent code is either time-consuming or inaccurate. In this paper, we develop a dual-path inpainting network with inversion path and feed-forward path, in which inversion path provides auxiliary information to help feed-forward path. We also design a novel deformable fusion module to align the feature maps in two paths. Experiments on FFHQ and LSUN demonstrate that our method is effective in solving the aforementioned problems while producing more realistic results than state-of-the-art methods.

The success of deep learning has enabled advances in multimodal tasks that require non-trivial fusion of multiple input domains. Although multimodal models have shown potential in many problems, their increased complexity makes them more vulnerable to attacks. A Backdoor (or Trojan) attack is a class of security vulnerability wherein an attacker embeds a malicious secret behavior into a network (e.g. targeted misclassification) that is activated when an attacker-specified trigger is added to an input. In this work, we show that multimodal networks are vulnerable to a novel type of attack that we refer to as Dual-Key Multimodal Backdoors. This attack exploits the complex fusion mechanisms used by state-of-the-art networks to embed backdoors that are both effective and stealthy. Instead of using a single trigger, the proposed attack embeds a trigger in each of the input modalities and activates the malicious behavior only when both the triggers are present. We present an extensive study of multimodal backdoors on the Visual Question Answering (VQA) task with multiple architectures and visual feature backbones. A major challenge in embedding backdoors in VQA models is that most models use visual features extracted from a fixed pretrained object detector. This is challenging for the attacker as the detector can distort or ignore the visual trigger entirely, which leads to models where backdoors are over-reliant on the language trigger. We tackle this problem by proposing a visual trigger optimization strategy designed for pretrained object detectors. Through this method, we create Dual-Key Backdoors with over a 98% attack success rate while only poisoning 1% of the training data. Finally, we release TrojVQA, a large collection of clean and trojan VQA models to enable research in defending against multimodal backdoors.

We propose the Dual-Generator (DG) network for large-pose face reenactment. Given a source face and a reference face as inputs, the DG network can generate an output face that has the same pose and expression as of the reference face, and has the same identity as of the source face. As most approaches do not particularly consider large-pose reenactment, the proposed approach addresses this issue by incorporating a 3D landmark detector into the framework and considering a loss function to capture visible local shape variation across large pose. The DG network consists of two modules, the ID-preserving Shape Generator (IDSG) and the Reenacted Face Generator (RFG). The IDSG encodes the 3D landmarks of the reference face into a reference landmark code, and encodes the source face into a source face code. The reference landmark code and the source face code are concatenated and decoded to a set of target landmarks that exhibits the pose and expression of the reference face and preserves the identity of the source face.

Contrastive learning (CL) is widely known to require many negative samples, 65536 in MoCo for instance, for which the performance of a dictionary-free framework is often inferior because the negative sample size (NSS) is limited by its mini-batch size (MBS). To decouple the NSS from the MBS, a dynamic dictionary has been adopted in a large volume of CL frameworks, among which arguably the most popular one is MoCo family. In essence, MoCo adopts a momentum-based queue dictionary, for which we perform a fine-grained analysis of its size and consistency. We point out that InfoNCE loss used in MoCo implicitly attract anchors to their corresponding positive sample with various strength of penalties and identify such inter-anchor hardness-awareness property as a major reason for the necessity of a large dictionary. Our findings motivate us to simplify MoCo v2 via the removal of its dictionary as well as momentum. Based on an InfoNCE with the proposed dual temperature, our simplified frameworks, SimMoCo and SimCo, outperform MoCo v2 by a visible margin. Moreover, our work bridges the gap between CL and non-CL frameworks, contributing to a more unified understanding of these two mainstream frameworks in SSL. Code is available at: https://bit.ly/3LkQbaT.

Recently, self-attention mechanisms have shown impressive performance in various NLP and CV tasks, which can help capture sequential characteristics and derive global information. In this work, we explore how to extend self-attention modules to better learn subtle feature embeddings for recognizing fine-grained objects, e.g., different bird species or person identities. To this end, we propose a dual cross-attention learning (DCAL) algorithm to coordinate with self-attention learning. First, we propose global-local cross-attention (GLCA) to enhance the interactions between global images and local high-response regions, which can help reinforce the spatial-wise discriminative clues for recognition. Second, we propose pair-wise cross-attention (PWCA) to establish the interactions between image pairs. PWCA can regularize the attention learning of an image by treating another image as distractor and will be removed during inference. We observe that DCAL can reduce misleading attentions and diffuse the attention response to discover more complementary parts for recognition. We conduct extensive evaluations on fine-grained visual categorization and object re-identification. Experiments demonstrate that DCAL performs on par with state-of-the-art methods and consistently improves multiple self-attention baselines, e.g., surpassing DeiT-Tiny and ViT-Base by 2.8% and 2.4% mAP on MSMT17, respectively.

While significant progress has been made in garment transfer, one of the most applicable directions of human-centric image generation, existing works overlook the in-the-wild imagery, presenting severe garment-person misalignment as well as noticeable degradation in fine texture details. This paper, therefore, attends to virtual try-on in real-world scenes and brings essential improvements in authenticity and naturalness especially for loose garment (e.g., skirts, formal dresses), challenging poses (e.g., cross arms, bent legs), and cluttered backgrounds. Specifically, we find that the pixel flow excels at handling loose garments whereas the vertex flow is preferred for hard poses, and by combining their advantages we propose a novel generative network called wFlow that can effectively push up garment transfer to in-the-wild context. Moreover, former approaches require paired images for training. Instead, we cut down the laboriousness by working on a newly constructed large-scale video dataset named Dance50k with self-supervised cross-frame training and an online cycle optimization. The proposed Dance50k can boost real-world virtual dressing by covering a wide variety of garments under dancing poses. Extensive experiments demonstrate the superiority of our wFlow in generating realistic garment transfer results for in-the-wild images without resorting to expensive paired datasets.

Convolutional image deraining networks have achieved great success while suffering from tremendous computational and memory costs. Most model compression methods require original data for iterative fine-tuning, which is limited in real-world applications due to storage, privacy, and transmission constraints. We note that it is overstretched to fine-tune the compressed model using self-collected data, as it exhibits poor generalization over images with different degradation characteristics. To address this problem, we propose a novel data-free compression framework for deraining networks. It is based on our observation that deep degradation representations can be clustered by degradation characteristics (types of rain) while independent of image content. Therefore, in our framework, we "dream" diverse in-distribution degraded images using a deep inversion paradigm, thus leveraging them to distill the pruned model. Specifically, we preserve the performance of the pruned model in a dual-branch way. In one branch, we invert the pre-trained model (teacher) to reconstruct the degraded inputs that resemble the original distribution and employ the orthogonal regularization for deep features to yield degradation diversity. In the other branch, the pruned model (student) is distilled to fit the teacher's original statistical modeling on these dreamed inputs. Further, an adaptive pruning scheme is proposed to determine the hierarchical sparsity, which alleviates the regression drift of the initial pruned model. Experiments on various deraining datasets demonstrate that our method can reduce about 40% FLOPs of the state-of-the-art models while maintaining comparable performance without original data.

We introduce DoubleField, a novel framework combining the merits of both surface field and radiance field for high-fidelity human reconstruction and rendering. Within DoubleField, the surface field and radiance field are associated together by a shared feature embedding and a surface-guided sampling strategy. Moreover, a view-to-view transformer is introduced to fuse multi-view features and learn view-dependent features directly from high-resolution inputs. With the modeling power of DoubleField and the view-to-view transformer, our method significantly improves the reconstruction quality of both geometry and appearance, while supporting direct inference, scene-specific high-resolution finetuning, and fast rendering. The efficacy of DoubleField is validated by the quantitative evaluations on several datasets and the qualitative results in a real-world sparse multi-view system, showing its superior capability for high-quality human model reconstruction and photo-realistic free-viewpoint human rendering. Data and source code will be made public for the research purpose.

The human visual system is remarkable in learning new visual concepts from just a few examples. This is precisely the goal behind few-shot class incremental learning (FSCIL), where the emphasis is additionally placed on ensuring the model does not suffer from "forgetting". In this paper, we push the boundary further for FSCIL by addressing two key questions that bottleneck its ubiquitous application (i) can the model learn from diverse modalities other than just photo (as humans do), and (ii) what if photos are not readily accessible (due to ethical and privacy constraints). Our key innovation lies in advocating the use of sketches as a new modality for class support. The product is a "Doodle It Yourself" (DIY) FSCIL framework where the users can freely sketch a few examples of a novel class for the model to learn to recognise photos of that class. For that, we present a framework that infuses (i) gradient consensus for domain invariant learning, (ii) knowledge distillation for preserving old class information, and (iii) graph attention networks for message passing between old and novel classes. We experimentally show that sketches are better class support than text in the context of FSCIL, echoing findings elsewhere in the sketching literature.

Unsupervised domain adaptation (UDA) is an important topic in the computer vision community. The key difficulty lies in defining a common property between the source and target domains so that the source-domain features can align with the target-domain semantics. In this paper, we present a simple and effective mechanism that regularizes cross-domain representation learning with a domain-agnostic prior (DAP) that constrains the features extracted from source and target domains to align with a domain-agnostic space. In practice, this is easily implemented as an extra loss term that requires a little extra costs. In the standard evaluation protocol of transferring synthesized data to real data, we validate the effectiveness of different types of DAP, especially one borrowed from a text embedding model that shows favorable performance beyond the state-of-the-art UDA approaches in terms of segmentation accuracy. Our research reveals that much room is left for designing better proxies for UDA.

With diverse presentation attacks emerging continually, generalizable face anti-spoofing (FAS) has drawn growing attention. Most existing methods implement domain generalization (DG) on the complete representations. However, different image statistics may have unique properties for the FAS tasks. In this work, we separate the complete representation into content and style ones. A novel Shuffled Style Assembly Network (SSAN) is proposed to extract and reassemble different content and style features for a stylized feature space. Then, to obtain a generalized representation, a contrastive learning strategy is developed to emphasize liveness-related style information while suppress the domain-specific one. Finally, the representations of the correct assemblies are used to distinguish between living and spoofing during the inferring. On the other hand, despite the decent performance, there still exists a gap between academia and industry, due to the difference in data quantity and distribution. Thus, a new large-scale benchmark for FAS is built up to further evaluate the performance of algorithms in reality. Both qualitative and quantitative results on existing and proposed benchmarks demonstrate the effectiveness of our methods. The codes will be available at https://github.com/wangzhuo2019/SSAN.

As a popular geometric representation, point clouds have attracted much attention in 3D vision, leading to many applications in autonomous driving and robotics. One important yet unsolved issue for learning on point cloud is that point clouds of the same object can have significant geometric variations if generated using different procedures or captured using different sensors. These inconsistencies induce domain gaps such that neural networks trained on one domain may fail to generalize on others. A typical technique to reduce the domain gap is to perform adversarial training so that point clouds in the feature space can align. However, adversarial training is easy to fall into degenerated local minima, resulting in negative adaptation gains. Here we propose a simple yet effective method for unsupervised domain adaptation on point clouds by employing a self-supervised task of learning geometry-aware implicits, which plays two critical roles in one shot. First, the geometric information in the point clouds is preserved through the implicit representations for downstream tasks. More importantly, the domain-specific variations can be effectively learned away in the implicit space. We also propose an adaptive strategy to compute unsigned distance fields for arbitrary point clouds due to the lack of shape models in practice. When combined with a task loss, the proposed outperforms state-of-the-art unsupervised domain adaptation methods that rely on adversarial domain alignment and more complicated self-supervised tasks. Our method is evaluated on both PointDA-10 and GraspNet datasets. Code and data are available at: https://github.com/Jhonve/ImplicitPCDA

E-commerce images are playing a central role in attracting people's attention when retailing and shopping online, and an accurate attention prediction is of significant importance for both customers and retailers, where its research is yet to start. In this paper, we establish the first dataset of saliency e-commerce images (SalECI), which allows for learning to predict saliency on the e-commerce images. We then provide specialized and thorough analysis by highlighting the distinct features of e-commerce images, e.g., non-locality and correlation to text regions. Correspondingly, taking advantages of the non-local and self-attention mechanisms, we propose a salient SWin-Transformer backbone, followed by a multi-task learning with saliency and text detection heads, where an information flow mechanism is proposed to further benefit both tasks. Experimental results have verified the state-of-the-art performances of our work in the e-commerce scenario.

As clean ImageNet accuracy nears its ceiling, the research community is increasingly more concerned about robust accuracy under distributional shifts. While a variety of methods have been proposed to robustify neural networks, these techniques often target models trained on ImageNet classification. At the same time, it is a common practice to use ImageNet pretrained backbones for downstream tasks such as object detection, semantic segmentation, and image classification from different domains. This raises a question: Can these robust image classifiers transfer robustness to downstream tasks? For object detection and semantic segmentation, we find that a vanilla Swin Transformer, a variant of Vision Transformer tailored for dense prediction tasks, transfers robustness better than Convolutional Neural Networks that are trained to be robust to the corrupted version of ImageNet. For CIFAR10 classification, we find that models that are robustified for ImageNet do not retain robustness when fully fine-tuned. These findings suggest that current robustification techniques tend to emphasize ImageNet evaluations. Moreover, network architecture is a strong source of robustness when we consider transfer learning.

Data often has many semantic attributes that are causally associated with each other. But do attribute-specific learned representations of data also respect the same causal relations? We answer this question in three steps. First, we introduce NCINet, an approach for observational causal discovery from high-dimensional data. It is trained purely on synthetically generated representations and can be applied to real representations, and is specifically designed to mitigate the domain gap between the two. Second, we apply NCINet to identify the causal relations between image representations of different pairs of attributes with known and unknown causal relations between the labels. For this purpose, we consider image representations learned for predicting attributes on the 3D Shapes, CelebA, and the CASIA-WebFace datasets, which we annotate with multiple multi-class attributes. Third, we analyze the effect on the underlying causal relation between learned representations induced by various design choices in representation learning. Our experiments indicate that (1) NCINet significantly outperforms existing observational causal discovery approaches for estimating the causal relation between pairs of random samples, both in the presence and absence of an unobserved confounder, (2) under controlled scenarios, learned representations can indeed satisfy the underlying causal relations between their respective labels, and (3) the causal relations are positively correlated with the predictive capability of the representations. Code and annotations are available at: https://github.com/human-analysis/causal-relations-between-representations.

It is a mystery which input features contribute to a neural network's output. Various explanations methods are proposed in the literature to shed light on the problem. One peculiar observation is that these explanations point to different features as being important. The phenomenon raises the question, which explanation to trust? We propose a framework for evaluating the explanations using the neural network model itself. The framework leverages the network to generate input features that impose a particular behavior on the output. Using the generated features, we devise controlled experimental setups to evaluate whether an explanation method conforms to an axiom. Thus we propose an empirical framework for axiomatic evaluation of explanation methods. We evaluate well-known and promising explanation solutions using the proposed framework. The framework provides a toolset to reveal properties and drawbacks within existing and future explanation solutions.

In response to the explosively-increasing requirement of annotated data, Novel Class Discovery (NCD) has emerged as a promising alternative to automatically recognize unknown classes without any annotation. To this end, a model makes use of a base set to learn basic semantic discriminability that can be transferred to recognize novel classes. Most existing works handle the base and novel sets using separate objectives within a two-stage training paradigm. Despite showing competitive performance on novel classes, they fail to generalize to recognizing samples from both base and novel sets. In this paper, we focus on this generalized setting of NCD (GNCD), and propose to divide and conquer it with two groups of Compositional Experts (ComEx). Each group of experts is designed to characterize the whole dataset in a comprehensive yet complementary fashion. With their union, we can solve GNCD in an efficient end-to-end manner. We further look into the drawback in current NCD methods, and propose to strengthen ComEx with global-to-local and local-to-local regularization. ComEx is evaluated on four popular benchmarks, showing clear superiority towards the goal of GNCD.

We address the problem of generating a 360-degree image from a single image with a narrow field of view by estimating its surroundings. Previous methods suffered from overfitting to the training resolution and deterministic generation. This paper proposes a completion method using a transformer for scene modeling and novel methods to improve the properties of a 360-degree image on the output image. Specifically, we use CompletionNets with a transformer to perform diverse completions and AdjustmentNet to match color, stitching, and resolution with an input image, enabling inference at any resolution. To improve the properties of a 360-degree image on an output image, we also propose WS-perceptual loss and circular inference. Thorough experiments show that our method outperforms state-of-the-art (SOTA) methods both qualitatively and quantitatively. For example, compared to SOTA methods, our method completes images 16 times larger in resolution and achieves 1.7 times lower Frechet inception distance (FID). Furthermore, we propose a pipeline that uses the completion results for lighting and background of 3DCG scenes. Our plausible background completion enables perceptually natural results in the application of inserting virtual objects with specular surfaces.

In this paper, we present a novel Distribution-Aware Single-stage (DAS) model for tackling the challenging multi-person 3D pose estimation problem. Different from existing top-down and bottom-up methods, the proposed DAS model simultaneously localizes person positions and their corresponding body joints in the 3D camera space in a one-pass manner. This leads to a simplified pipeline with enhanced efficiency. In addition, DAS learns the true distribution of body joints for the regression of their positions, rather than making a simple Laplacian or Gaussian assumption as previous works. This provides valuable priors for model prediction and thus boosts the regression-based scheme to achieve competitive performance with volumetric-base ones. Moreover, DAS exploits a recursive update strategy for progressively approaching to regression target, alleviating the optimization difficulty and further lifting the regression performance. DAS is implemented with a fully Convolutional Neural Network and end-to-end learnable. Comprehensive experiments on benchmarks CMU Panoptic and MuPoTS-3D demonstrate the superior efficiency of the proposed DAS model, specifically 1.5x speedup over previous best model, and its stat-of-the-art accuracy for multi-person 3D pose estimation.