ICCV 2021 论文列表

Robust face reconstruction from monocular image in general lighting conditions is challenging. Methods combining deep neural network encoders with differentiable rendering have opened up the path for very fast monocular reconstruction of geometry, lighting and reflectance. They can also be trained in self-supervised manner for increased robustness and better generalization. However, their differentiable rasterization based image formation models, as well as underlying scene parameterization, limit them to Lambertian face reflectance and to poor shape details. More recently, ray tracing was introduced for monocular face reconstruction within a classic optimization-based framework and enables state-of-the art results. However optimization-based approaches are inherently slow and lack robustness. In this paper, we build our work on the aforementioned approaches and propose a new method that greatly improves reconstruction quality and robustness in general scenes. We achieve this by combining a CNN encoder with a differentiable ray tracer, which enables us to base the reconstruction on much more advanced personalized diffuse and specular albedos, a more sophisticated illumination model and a plausible representation of self-shadows. This enables to take a big leap forward in reconstruction quality of shape, appearance and lighting even in scenes with difficult illumination. With consistent face attributes reconstruction, our method leads to practical applications such as relighting and self-shadows removal. Compared to state-of-the-art methods, our results show improved accuracy and validity of the approach.

Training a single deep blind model to handle different quality factors for JPEG image artifacts removal has been attracting considerable attention due to its convenience for practical usage. However, existing deep blind methods usually directly reconstruct the image without predicting the quality factor, thus lacking the flexibility to control the output as the non-blind methods. To remedy this problem, in this paper, we propose a flexible blind convolutional neural network, namely FBCNN, that can predict the adjustable quality factor to control the trade-off between artifacts removal and details preservation. Specifically, FBCNN decouples the quality factor from the JPEG image via a decoupler module and then embeds the predicted quality factor into the subsequent reconstructor module through a quality factor attention block for flexible control. Besides, we find existing methods are prone to fail on non-aligned double JPEG images even with only a one-pixel shift, and we thus propose a double JPEG degradation model to augment the training data. Extensive experiments on single JPEG images, more general double JPEG images, and real-world JPEG images demonstrate that our proposed FBCNN achieves favorable performance against state-of-the-art methods in terms of both quantitative metrics and visual quality.

As billions of personal data being shared through social media and network, the data privacy and security have drawn an increasing attention. Several attempts have been made to alleviate the leakage of identity information from face photos, with the aid of, e.g., image obfuscation techniques. However, most of the present results are either perceptually unsatisfactory or ineffective against face recognition systems. Our goal in this paper is to develop a technique that can encrypt the personal photos such that they can protect users from unauthorized face recognition systems but remain visually identical to the original version for human beings. To achieve this, we propose a targeted identity-protection iterative method (TIP-IM) to generate adversarial identity masks which can be overlaid on facial images, such that the original identities can be concealed without sacrificing the visual quality. Extensive experiments demonstrate that TIP-IM provides 95%+ protection success rate against various state-of-the-art face recognition models under practical open-set test scenarios. Besides, we also show the practical and effective applicability of our method on a commercial API service.

We aim at improving the computational efficiency of graph convolutional networks (GCNs) for learning on point clouds. The basic graph convolution that is composed of a K-nearest neighbor (KNN) search and a multilayer perceptron (MLP) is examined. By mathematically analyzing the operations there, two findings to improve the efficiency of GCNs are obtained. (1) The local geometric structure information of 3D representations propagates smoothly across the GCN that relies on KNN search to gather neighborhood features. This motivates the simplification of multiple KNN searches in GCNs. (2) Shuffling the order of graph feature gathering and an MLP leads to equivalent or similar composite operations. Based on those findings, we optimize the computational procedure in GCNs. A series of experiments show that the optimized networks have reduced computational complexity, decreased memory consumption, and accelerated inference speed while maintaining comparable accuracy for learning on point clouds.

In this work, we address the problem of unsupervised domain adaptation for person re-ID where annotations are available for the source domain but not for target. Previous methods typically follow a two-stage optimization pipeline, where the network is first pre-trained on source and then fine-tuned on target with pseudo labels created by feature clustering. Such methods sustain two main limitations. (1) The label noise may hinder the learning of discriminative features for recognizing target classes. (2) The domain gap may hinder knowledge transferring from source to target. We propose three types of technical schemes to alleviate these issues. First, we propose a cluster-wise contrastive learning algorithm (CCL) by iterative optimization of feature learning and cluster refinery to learn noise-tolerant representations in the unsupervised manner. Second, we adopt a progressive domain adaptation (PDA) strategy to gradually mitigate the domain gap between source and target data. Third, we propose Fourier augmentation (FA) for further maximizing the class separability of re-ID models by imposing extra constraints in the Fourier space. We observe that these proposed schemes are capable of facilitating the learning of discriminative feature representations. Experiments demonstrate that our method consistently achieves notable improvements over the state-of-the-art unsupervised re-ID methods on multiple benchmarks, e.g., surpassing MMT largely by 8.1%, 9.9%, 11.4% and 11.1% mAP on the Market-to-Duke, Duke-to-Market, Market-to-MSMT and Duke-to-MSMT tasks, respectively.

With the recent progress in Generative Adversarial Networks (GANs), it is imperative for media and visual forensics to develop detectors which can identify and attribute images to the model generating them. Existing works have shown to attribute images to their corresponding GAN sources with high accuracy. However, these works are limited to a closed set scenario, failing to generalize to GANs unseen during train time and are therefore, not scalable with a steady influx of new GANs. We present an iterative algorithm for discovering images generated from previously unseen GANs by exploiting the fact that all GANs leave distinct fingerprints on their generated images. Our algorithm consists of multiple components including network training, out-of-distribution detection, clustering, merge and refine steps. Through extensive experiments, we show that our algorithm discovers unseen GANs with high accuracy and also generalizes to GANs trained on unseen real datasets. We additionally apply our algorithm to attribution and discovery of GANs in an online fashion as well as to the more standard task of real/fake detection. Our experiments demonstrate the effectiveness of our approach to discover new GANs and can be used in an open-world setup.

The wide-angle lens gains increasing attention since it can capture a wide field-of-view scene (FoV). However, the obtained image is contaminated with radial distortion, making the scene not realistic. Previous distortion rectification methods rectify the image in a rectangle or invagination, failing to display the complete content and regular shape simultaneously. In this paper, we rethink the representation of rectification results and present a Rectification OutPainting (ROP) method, aiming to extrapolate the coherent semantics to the blank area and create a wider FoV beyond the original wide-angle lens. To address the specific challenges such as the variable painting region and curve boundary, a rectification module is designed to rectify the image with geometry supervision, and the extrapolated results are generated using a dual conditional expansion strategy. In terms of the spatially discounted correlation, a curve-aware correlation measurement is proposed to focus on the generated region to enforce the local consistency. To our knowledge, we are the first to tackle the challenging rectification via outpainting, and our curve-aware strategy can reach a rectification construction with complete content and regular shape. Extensive experiments well demonstrate the superiority of our ROP over other state-of-the-art solutions.

Increasing demands for understanding the internal behavior of convolutional neural networks (CNNs) have led to remarkable improvements in explanation methods. Particularly, several class activation mapping (CAM) based methods, which generate visual explanation maps by a linear combination of activation maps from CNNs, have been proposed. However, the majority of the methods lack a clear theoretical basis on how they assign the coefficients of the linear combination. In this paper, we revisit the intrinsic linearity of CAM with respect to the activation maps; we construct an explanation model of CNN as a linear function of binary variables that denote the existence of the corresponding activation maps. With this approach, the explanation model can be determined by additive feature attribution methods in an analytic manner. We then demonstrate the adequacy of SHAP values, which is a unique solution for the explanation model with a set of desirable properties, as the coefficients of CAM. Since the exact SHAP values are unattainable, we introduce an efficient approximation method, LIFT-CAM, based on DeepLIFT. Our proposed LIFT-CAM can estimate the SHAP values of the activation maps with high speed and accuracy. Furthermore, it greatly outperforms other previous CAM-based methods in both qualitative and quantitative aspects.

In this work, we study the ambiguity problem in the task of unsupervised 3D human pose estimation from 2D counterpart. On one hand, without explicit annotation, the scale of 3D pose is difficult to be accurately captured (scale ambiguity). On the other hand, one 2D pose might correspond to multiple 3D gestures, where the lifting procedure is inherently ambiguous (pose ambiguity). Previous methods generally use temporal constraints (e.g., constant bone length and motion smoothness) to alleviate the above issues. However, these methods commonly enforce the outputs to fulfill multiple training objectives simultaneously, which often lead to sub-optimal results. In contrast to the majority of previous works, we propose to split the whole problem into two sub-tasks, i.e., optimizing 2D input poses via a scale estimation module and then mapping optimized 2D pose to 3D counterpart via a pose lifting module. Furthermore, two temporal constraints are proposed to alleviate the scale and pose ambiguity respectively. These two modules are optimized via a iterative training scheme with corresponding temporal constraints, which effectively reduce the learning difficulty and lead to better performance. Results on the Human3.6M dataset demonstrate that our approach improves upon the prior art by 23.1% and also outperforms several weakly supervised approaches that rely on 3D annotations. Our project is available at https://sites.google.com/view/ambiguity-aware-hpe.

3D hand-mesh reconstruction from RGB images facilitates many applications, including augmented reality (AR). However, this requires not only real-time speed and accurate hand pose and shape but also plausible mesh-image alignment. While existing works already achieve promising results, meeting all three requirements is very challenging. This paper presents a novel pipeline by decoupling the hand-mesh reconstruction task into three stages: a joint stage to predict hand joints and segmentation; a mesh stage to predict a rough hand mesh; and a refine stage to fine-tune it with an offset mesh for mesh-image alignment. With careful design in the network structure and in the loss functions, we can promote high-quality finger-level mesh-image alignment and drive the models together to deliver real-time predictions. Extensive quantitative and qualitative results on benchmark datasets demonstrate that the quality of our results outperforms the state-of-the-art methods on hand-mesh/pose precision and hand-image alignment. In the end, we also showcase several real-time AR scenarios.

A large body of recent work has identified transformations in the latent spaces of generative adversarial networks (GANs) that consistently and interpretably transform generated images. But existing techniques for identifying these transformations rely on either a fixed vocabulary of pre-specified visual concepts, or on unsupervised disentanglement techniques whose alignment with human judgments about perceptual salience is unknown. This paper introduces a new method for building open-ended vocabularies of primitive visual concepts represented in a GAN's latent space. Our approach is built from three components: (1) automatic identification of perceptually salient directions based on their layer selectivity; (2) human annotation of these directions with free-form, compositional natural language descriptions; and (3) decomposition of these annotations into a visual concept vocabulary, consisting of distilled directions labeled with single words. Experiments show that concepts learned with our approach are reliable and composable--generalizing across classes, contexts, and observers, and enabling fine-grained manipulation of image style and content.

Recent image generation models show remarkable generation performance. However, they mirror strong location preference in datasets, which we call spatial bias. Therefore, generators render poor samples at unseen locations and scales. We argue that the generators rely on their implicit positional encoding to render spatial content. From our observations, the generator's implicit positional encoding is translation-variant, making the generator spatially biased. To address this issue, we propose injecting explicit positional encoding at each scale of the generator. By learning the spatially unbiased generator, we facilitate the robust use of generators in multiple tasks, such as GAN inversion, multi-scale generation, generation of arbitrary sizes and aspect ratios. Furthermore, we show that our method can also be applied to denoising diffusion probabilistic models.

Recovering the 3D structure of an object from a single image is a challenging task due to its ill-posed nature. One approach is to utilize the plentiful photos of the same object category to learn a strong 3D shape prior for the object. This approach has successfully been demonstrated by a recent work of Wu et al. (2020), which obtained impressive 3D reconstruction networks with unsupervised learning. However, their algorithm is only applicable to symmetric objects. In this paper, we eliminate the symmetry requirement with a novel unsupervised algorithm that can learn a 3D reconstruction network from a multi-image dataset. Our algorithm is more general and covers the symmetry-required scenario as a special case. Besides, we employ a novel albedo loss that improves the reconstructed details and realisticity. Our method surpasses the previous work in both quality and robustness, as shown in experiments on datasets of various structures, including single-view, multi-view, image-collection, and video sets.

In recent years, many dexterous robotic hands have been designed to assist or replace human hands in executing various tasks. But how to teach them to perform dexterous operations like human hands is still a challenging task. In this paper, we propose a grasp synthesis framework to make robots grasp and manipulate objects like human beings. We first build a dataset by accurately segmenting the functional areas of the object and annotating semantic touch code for each functional area to guide the dexterous hand to complete the functional grasp and post-grasp manipulation. This dataset contains 18 categories of 129 objects selected from four datasets, and 15 people participated in data annotation. Then we carefully design four loss functions to constrain the network, which successfully generates the functional grasp of dexterous hand under the guidance of semantic touch code. The thorough experiments in synthetic data show our model can robustly generate functional grasp, even for objects that the model has not see before.

High-fidelity face digitization solutions often combine multi-view stereo (MVS) techniques for 3D reconstruction and a non-rigid registration step to establish dense correspondence across identities and expressions. A common problem is the need for manual clean-up after the MVS step, as 3D scans are typically affected by noise and outliers and contain hairy surface regions that need to be cleaned up by artists. Furthermore, mesh registration tends to fail for extreme facial expressions. Most learning-based methods use an underlying 3D morphable model (3DMM) to ensure robustness, but this limits the output accuracy for extreme facial expressions. In addition, the global bottleneck of regression architectures cannot produce meshes that tightly fit the ground truth surfaces. We propose ToFu, Topological consistent Face from multi-view, a geometry inference framework that can produce topologically consistent meshes across facial identities and expressions using a volumetric representation instead of an explicit underlying 3DMM. Our novel progressive mesh generation network embeds the topological structure of the face in a feature volume, sampled from geometry-aware local features. A coarse-to-fine architecture facilitates dense and accurate facial mesh predictions in a consistent mesh topology. ToFu further captures displacement maps for pore-level geometric details and facilitates high-quality rendering in the form of albedo and specular reflectance maps. These high-quality assets are readily usable by production studios for avatar creation, animation and physically-based skin rendering. We demonstrate state-of-the-art geometric and correspondence accuracy, while only taking 0.385 seconds to compute a mesh with 10K vertices, which is three orders of magnitude faster than traditional techniques. The code and the model are available for research purposes at https://tianyeli.github.io/tofu.

If an image tells a story, the image caption is the briefest narrator. Generally, a scene graph prefers to be an omniscient "generalist", while the image caption is more willing to be a "specialist", which outlines the gist. Lots of previous studies have found that a scene graph is not as practical as expected unless it can reduce the trivial contents and noises. In this respect, the image caption is a good tutor. To this end, we let the scene graph borrow the ability from the image caption so that it can be a specialist on the basis of remaining all-around, resulting in the so-called Topic Scene Graph. What an image caption pays attention to is distilled and passed to the scene graph for estimating the importance of partial objects, relationships, and events. Specifically, during the caption generation, the attention about individual objects in each time step is collected, pooled, and assembled to obtain the attention about relationships, which serves as weak supervision for regularizing the estimated importance scores of relationships. In addition, as this attention distillation process provides an opportunity for combining the generation of image caption and scene graph together, we further transform the scene graph into linguistic form with rich and free-form expressions by sharing a single generation model with image caption. Experiments show that attention distillation brings significant improvements in mining important relationships without strong supervision, and the topic scene graph shows great potential in subsequent applications.

Transformers, which are popular for language modeling, have been explored for solving vision tasks recently, e.g., the Vision Transformer (ViT) for image classification. The ViT model splits each image into a sequence of tokens with fixed length and then applies multiple Transformer layers to model their global relation for classification. However, ViT achieves inferior performance to CNNs when trained from scratch on a midsize dataset like ImageNet. We find it is because: 1) the simple tokenization of input images fails to model the important local structure such as edges and lines among neighboring pixels, leading to low training sample efficiency; 2) the redundant attention backbone design of ViT leads to limited feature richness for fixed computation budgets and limited training samples. To overcome such limitations, we propose a new Tokens-To-Token Vision Transformer (T2T-ViT), which incorporates 1) a layer-wise Tokens-to-Token (T2T) transformation to progressively structurize the image to tokens by recursively aggregating neighboring Tokens into one Token (Tokens-to-Token), such that local structure represented by surrounding tokens can be modeled and tokens length can be reduced; 2) an efficient backbone with a deep-narrow structure for vision transformer motivated by CNN architecture design after empirical study. Notably, T2T-ViT reduces the parameter count and MACs of vanilla ViT by half, while achieving more than 3.0% improvement when trained from scratch on ImageNet. It also outperforms ResNets and achieves comparable performance with MobileNets by directly training on ImageNet. For example, T2T-ViT with comparable size to ResNet50 (21.5M parameters) can achieve 83.3% top1 accuracy in image resolution 384x384 on ImageNet.

Human pose estimation deeply relies on visual clues and anatomical constraints between parts to locate keypoints. Most existing CNN-based methods do well in visual representation, however, lacking in the ability to explicitly learn the constraint relationships between keypoints. In this paper, we propose a novel approach based on Token representation for human Pose estimation (TokenPose). In detail, each keypoint is explicitly embedded as a token to simultaneously learn constraint relationships and appearance cues from images. Extensive experiments show that the small and large TokenPose models are on par with state-of-the-art CNN-based counterparts while being more lightweight. Specifically, our TokenPose-S and TokenPose-L achieve 72.5 AP and 75.8 AP on COCO validation dataset respectively, with significant reduction in parameters and GFLOPs. Code is publicly available at https://github.com/leeyegy/TokenPose.

Top-k multi-label learning, which returns the top-k predicted labels from an input, has many practical applications such as image annotation, document analysis, and web search engine. However, the vulnerabilities of such algorithms with regards to dedicated adversarial perturbation attacks have not been extensively studied previously. In this work, we develop methods to create adversarial perturbations that can be used to attack top-k multi-label learning-based image annotation systems (T_kML-AP). Our methods explicitly consider the top-k ranking relation and are based on novel loss functions. Experimental evaluations on large-scale benchmark datasets including PASCAL VOC and MS COCO demonstrate the effectiveness of our methods in reducing the performance of state-of-the-art top-k multi-label learning methods, under both untargeted and targeted attacks.