ICCV 2021 Papers

For face texture completion, previous methods typically use some complete textures captured by multiview imaging systems or 3D scanners for supervised learning. This paper deals with a new challenging problem -- learning to complete invisible texture in a single face image without using any complete texture. We simply leverage a large corpus of face images of different subjects (e.\,g., FFHQ) to train a texture completion model in an unsupervised manner. To achieve this, we propose DSD-GAN, a novel deep neural network based method that applies two discriminators in UV map space and image space. These two discriminators work in a complementary manner to learn both facial structures and texture details. We show that their combination is essential to obtain high-fidelity results. Despite the network never sees any complete facial appearance, it is able to generate compelling full textures from single images.

We propose a hierarchical graph neural network (GNN) model that learns how to cluster a set of images into an unknown number of identities using a training set of images annotated with labels belonging to a disjoint set of identities. Our hierarchical GNN uses a novel approach to merge connected components predicted at each level of the hierarchy to form a new graph at the next level. Unlike fully unsupervised hierarchical clustering, the choice of grouping and complexity criteria stems naturally from supervision in the training set. The resulting method, Hi-LANDER, achieves an average of 49% improvement in F-score and 7% increase in Normalized Mutual Information (NMI) relative to current GNN-based clustering algorithms. Additionally, state-of-the-art GNN-based methods rely on separate models to predict linkage probabilities and node densities as intermediate steps of the clustering process. In contrast, our unified framework achieves a three-fold decrease in computational cost. Our training and inference code are released.

Recent advances in differentiable rendering have sparked an interest in learning generative models of textured 3D meshes from image collections. These models natively disentangle pose and appearance, enable downstream applications in computer graphics, and improve the ability of generative models to understand the concept of image formation. Although there has been prior work on learning such models from collections of 2D images, these approaches require a delicate pose estimation step that exploits annotated keypoints, thereby restricting their applicability to a few specific datasets. In this work, we propose a GAN framework for generating textured triangle meshes without relying on such annotations. We show that the performance of our approach is on par with prior work that relies on ground-truth keypoints, and more importantly, we demonstrate the generality of our method by setting new baselines on a larger set of categories from ImageNet - for which keypoints are not available - without any class-specific hyperparameter tuning. We release our code at https://github.com/dariopavllo/textured-3d-gan

Learning from noisy data has attracted much attention, where most methods focus on label noise. In this work, we propose a new learning framework which simultaneously addresses three types of noise commonly seen in real-world data: label noise, out-of-distribution input, and input corruption. In contrast to most existing methods, we combat noise by learning robust representation. Specifically, we embed images into a low-dimensional subspace, and regularize the geometric structure of the subspace with robust contrastive learning, which includes an unsupervised consistency loss and a supervised mixup prototypical loss. We also propose a new noise cleaning method which leverages the learned representation to enforce a smoothness constraint on neighboring samples. Experiments on multiple benchmarks demonstrate state-of-the-art performance of our method and robustness of the learned representation. Code is available at https://github.com/salesforce/RRL/.

Deep learning based methods, especially convolutional neural networks (CNNs) have been successfully applied in the field of single image super-resolution (SISR). To obtain better fidelity and visual quality, most of existing networks are of heavy design with massive computation. However, the computation resources of modern mobile devices are limited, which cannot easily support the expensive cost. To this end, this paper explores a novel frequency-aware dynamic network for dividing the input into multiple parts according to its coefficients in the discrete cosine transform (DCT) domain. In practice, the high-frequency part will be processed using expensive operations and the lower-frequency part is assigned with cheap operations to relieve the computation burden. Since pixels or image patches belong to low-frequency areas contain relatively few textural details, this dynamic network will not affect the quality of resulting super-resolution images. In addition, we embed predictors into the proposed dynamic network to end-to-end fine-tune the handcrafted frequency-aware masks. Extensive experiments conducted on benchmark SISR models and datasets show that the frequency-aware dynamic network can be employed for various SISR neural architectures to obtain the better tradeoff between visual quality and computational complexity. For instance, we can reduce the FLOPs of SR models by approximate 50% while preserving the state-of-the-art SISR performance.

Training sample re-weighting is an effective approach for tackling data biases such as imbalanced and corrupted labels. Recent methods develop learning-based algorithms to learn sample re-weighting strategies jointly with model training based on the frameworks of reinforcement learning and meta learning. However, depending on additional unbiased reward data is limiting their general applicability. Furthermore, existing learning-based sample re-weighting methods require nested optimizations of models and weighting parameters, which requires expensive second-order computation. This paper addresses these two problems and presents a novel learning-based fast sample re-weighting (FSR) method that does not require additional reward data. The method is based on two key ideas: learning from history to build proxy reward data and feature sharing to reduce the optimization cost. Our experiments show the proposed method achieves competitive results compared to state of the arts on label noise robustness and long-tailed recognition, and does so while achieving significantly improved training efficiency. The source code is publicly available at https://github.com/google-research/google-research/tree/master/ieg.

Faces manifest large variations in many aspects, such as identity, expression, pose, and face styling. Therefore, it is a great challenge to disentangle and extract these characteristics from facial images, especially in an unsupervised manner. In this work, we introduce cycle-consistency in facial characteristics as free supervisory signal to learn facial representations from unlabeled facial images. The learning is realized by superimposing the facial motion cycle-consistency and identity cycle-consistency constraints. The main idea of the facial motion cycle-consistency is that, given a face with expression, we can perform de-expression to a neutral face via the removal of facial motion and further perform re-expression to reconstruct back to the original face. The main idea of the identity cycle-consistency is to exploit both de-identity into mean face by depriving the given neutral face of its identity via feature re-normalization and re-identity into neutral face by adding the personal attributes to the mean face. At training time, our model learns to disentangle two distinct facial representations to be useful for performing cycle-consistent face reconstruction. At test time, we use the linear protocol scheme for evaluating facial representations on various tasks, including facial expression recognition and head pose regression. We also can directly apply the learnt facial representations to person recognition, frontalization and image-to-image translation. Our experiments show that the results of our approach is competitive with those of existing methods, demonstrating the rich and unique information embedded in the disentangled representations. Code is available at https://github.com/JiaRenChang/FaceCycle.

We present a novel framework to learn to convert the per-pixel photometric information at each view into spatially distinctive and view-invariant low-level features, which can be plugged into existing multi-view stereo pipeline for enhanced 3D reconstruction. Both the illumination conditions during acquisition and the subsequent per-pixel feature transform can be jointly optimized in a differentiable fashion. Our framework automatically adapts to and makes efficient use of the geometric information available in different forms of input data. High-quality 3D reconstructions of a variety of challenging objects are demonstrated on the data captured with an illumination multiplexing device, as well as a point light. Our results compare favorably with state-of-the-art techniques.

In this paper, we tackle the problem of dense light field (LF) reconstruction from sparsely-sampled ones with wide baselines and propose a learnable model, namely dynamic interpolation, to replace the commonly-used geometry warping operation. Specifically, with the estimated geometric relation between input views, we first construct a lightweight neural network to dynamically learn weights for interpolating neighbouring pixels from input views to synthesize each pixel of novel views independently. In contrast to the fixed and content-independent weights employed in the geometry warping operation, the learned interpolation weights implicitly incorporate the correspondences between the source and novel views and adapt to different image content information. Then, we recover the spatial correlation between the independently synthesized pixels of each novel view by referring to that of input views using a geometry-based spatial refinement module. We also constrain the angular correlation between the novel views through a disparity-oriented LF structure loss. Experimental results on LF datasets with wide baselines show that the reconstructed LFs achieve much higher PSNR/SSIM and preserve the LF parallax structure better than state-of-the-art methods. The source code is publicly available at https://github.com/MantangGuo/DI4SLF.

Deep learning (DL)-based methods have achieved great success in solving the ill-posed JPEG compression artifacts removal problem. However, as most DL architectures are designed to directly learn pixel-level mapping relationships, they largely ignore semantic-level information and lack sufficient interpretability. To address the above issues, in this work, we propose an interpretable deep network to learn both pixel-level regressive prior and semantic-level discriminative prior. Specifically, we design a variational model to formulate the image de-blocking problem and propose two prior terms for the image content and gradient, respectively. The content-relevant prior is formulated as a DL-based image-to-image regressor to perform as a de-blocker from the pixel-level. The gradient-relevant prior serves as a DL-based classifier to distinguish whether the image is compressed from the semantic-level. To effectively solve the variational model, we design an alternating minimization algorithm and unfold it into a deep network architecture. In this way, not only the interpretability of the deep network is increased, but also the dual priors can be well estimated from training samples. By integrating the two priors into a single framework, the image de-blocking problem can be well-constrained, leading to a better performance. Experiments on benchmarks and real-world use cases demonstrate the superiority of our method to the existing state-of-the-art approaches.

In image retrieval, learning local features with deep convolutional networks has been demonstrated effective to improve the performance. To discriminate deep local features, some research efforts turn to attention learning. However, existing attention-based methods only generate a single attention map for each image, which limits the exploration of diverse visual patterns. To this end, we propose a novel deep local feature learning architecture to simultaneously focus on multiple discriminative local patterns in an image. In our framework, we first adaptively reorganize the channels of activation maps for multiple heads. For each head, a new dynamic attention module is designed to learn the potential attentions. The whole architecture is trained as metric learning of weighted-sum-pooled global image features, with only image-level relevance label. After the architecture training, for each database image, we select local features based on their multi-head dynamic attentions, which are further indexed for efficient retrieval. Extensive experiments show the proposed method outperforms the state-of-the-art methods on the Revisited Oxford and Paris datasets. Besides, it typically achieves competitive results even using local features with lower dimensions.

Learning transferable and domain adaptive feature representations from videos is important for video-relevant tasks such as action recognition. Existing video domain adaptation methods mainly rely on adversarial feature alignment, which has been derived from the RGB image space. However, video data is usually associated with multi-modal information, e.g., RGB and optical flow, and thus it remains a challenge to design a better method that considers the crossmodal inputs under the cross-domain adaptation setting. To this end, we propose a unified framework for video domain adaptation, which simultaneously regularizes cross-modal and cross-domain feature representations. Specifically, we treat each modality in a domain as a view and leverage the contrastive learning technique with properly designed sampling strategies. As a result, our objectives regularize feature spaces, which originally lack the connection across modalities or have less alignment across domains. We conduct experiments on domain adaptive action recognition benchmark datasets, i.e., UCF, HMDB and EPIC-Kitchens, and demonstrate the effectiveness of our individual components against state-of-the-art algorithms.

An important scenario for image quality assessment (IQA) is to evaluate image restoration (IR) algorithms. The state-of-the-art approaches adopt a full-reference paradigm that compares restored images with their corresponding pristine-quality images. However, pristine-quality images are usually unavailable in blind image restoration tasks and real-world scenarios. In this paper, we propose a practical solution named degraded-reference IQA (DR-IQA), which exploits the inputs of IR models, degraded images, as references. Specifically, we extract reference information from degraded images by distilling knowledge from pristine-quality images. The distillation is achieved through learning a reference space, where various degraded images are encouraged to share the same feature statistics with pristine-quality images. And the reference space is optimized to capture deep image priors that are useful for quality assessment. Note that pristine-quality images are only used during training. Our work provides a powerful and differentiable metric for blind IRs, especially for GAN-based methods. Extensive experiments show that our results can even be close to the performance of full-reference settings.

Achieving backward compatibility when rolling out new models can highly reduce costs or even bypass feature re-encoding of existing gallery images for in-production visual retrieval systems. Previous related works usually leverage losses used in knowledge distillation which can cause performance degradations or not guarantee compatibility. To address these issues, we propose a general framework called Learning Compatible Embeddings (LCE) which is applicable for both cross model compatibility and compatible training in direct/forward/backward manners. Our compatibility is achieved by aligning class centers between models directly or via a transformation, and restricting more compact intra-class distributions for the new model. Experiments are conducted in extensive scenarios such as changes of training dataset, loss functions, network architectures as well as feature dimensions, and demonstrate that LCE efficiently enables model compatibility with marginal sacrifices of accuracies.

3D pose estimation has attracted increasing attention with the availability of high-quality benchmark datasets. However, prior works show that deep learning models tend to learn spurious correlations, which fail to generalize beyond the specific dataset they are trained on. In this work, we take a step towards training robust models for cross-domain pose estimation task, which brings together ideas from causal representation learning and generative adversarial networks. Specifically, this paper introduces a novel framework for causal representation learning which explicitly exploits the causal structure of the task. We consider changing domain as interventions on images under the data-generation process and steer the generative model to produce counterfactual features. This help the model learn transferable and causal relations across different domains. Our framework is able to learn with various types of unlabeled datasets. We demonstrate the efficacy of our proposed method on both human and hand pose estimation task. The experiment results show the proposed approach achieves state-of-the-art performance on most datasets for both domain adaptation and domain generalization settings.

In this paper, we focus on recognizing 3D shapes from arbitrary views, i.e., arbitrary numbers and positions of viewpoints. It is a challenging and realistic setting for view-based 3D shape recognition. We propose a canonical view representation to tackle this challenge. We first transform the original features of arbitrary views to a fixed number of view features, dubbed canonical view representation, by aligning the arbitrary view features to a set of learnable reference view features using optimal transport. In this way, each 3D shape with arbitrary views is represented by a fixed number of canonical view features, which are further aggregated to generate a rich and robust 3D shape representation for shape recognition. We also propose a canonical view feature separation constraint to enforce that the view features in canonical view representation can be embedded into scattered points in a Euclidean space. Experiments on the ModelNet40, ScanObjectNN, and RGBD datasets show that our method achieves competitive results under the fixed viewpoint settings, and significantly outperforms the applicable methods under the arbitrary view setting.

We propose a novel framework for fine-grained object recognition that learns to recover object variation in 3D space from a single image, trained on an image collection without using any ground-truth 3D annotation. We accomplish this by representing an object as a composition of 3D shape and its appearance, while eliminating the effect of camera viewpoint, in a canonical configuration. Unlike conventional methods modeling spatial variation in 2D images only, our method is capable of reconfiguring the appearance feature in a canonical 3D space, thus enabling the subsequent object classifier to be invariant under 3D geometric variation. Our representation also allows us to go beyond existing methods, by incorporating 3D shape variation as an additional cue for object recognition. To learn the model without ground-truth 3D annotation, we deploy a differentiable renderer in an analysis-by-synthesis framework. By incorporating 3D shape and appearance jointly in a deep representation, our method learns the discriminative representation of the object and achieves competitive performance on fine-grained image recognition and vehicle re-identification. We also demonstrate that the performance of 3D shape reconstruction is improved by learning fine-grained shape deformation in a boosting manner.

Deep learning algorithms mine knowledge from the training data and thus would likely inherit the dataset's bias information. As a result, the obtained model would generalize poorly and even mislead the decision process in real-life applications. We propose to remove the bias information misused by the target task with a cross-sample adversarial debiasing (CSAD) method. CSAD explicitly extracts target and bias features disentangled from the latent representation generated by a feature extractor and then learns to discover and remove the correlation between the target and bias features. The correlation measurement plays a critical role in adversarial debiasing and is conducted by a cross-sample neural mutual information estimator. Moreover, we propose joint content and local structural representation learning to boost mutual information estimation for better performance. We conduct thorough experiments on publicly available datasets to validate the advantages of the proposed method over state-of-the-art approaches.

In many computer vision problems, it is desired to learn the effective visual data similarity such that the prediction accuracy can be enhanced. Deep Metric Learning (DML) methods have been actively studied to measure the data similarity. Pair-based and proxy-based losses are the two major paradigms in DML. However, pair-wise methods involve expensive training costs, while proxy-based methods are less accurate in characterizing the relationships between data points. In this paper, we provide a hybrid grouplet paradigm, which inherits the accurate pair-wise relationship in pair-based methods and the efficient training in proxy-based methods. Our method also equips a non-Euclidean space to DML, which employs a hierarchical representation manifold. More specifically, we propose a unified graph perspective --- different DML methods learn different local connecting patterns between data points. Based on the graph interpretation, we construct a flexible subset of data points, dubbed grouplet. Our grouplet doesn't require explicit pair-wise relationships, instead, we encode the data relationships in an optimal transport problem regarding the proxies, and solve this problem via a differentiable implicit layer to automatically determine the relationships. Extensive experimental results show that our method significantly outperforms state-of-the-art baselines on several benchmarks. The ablation studies also verify the effectiveness of our method.