CVPR 2021 Papers

Blind face restoration (BFR) from severely degraded face images in the wild is a very challenging problem. Due to the high illness of the problem and the complex unknown degradation, directly training a deep neural network (DNN) usually cannot lead to acceptable results. Existing generative adversarial network (GAN) based methods can produce better results but tend to generate over-smoothed restorations. In this work, we propose a new method by first learning a GAN for high-quality face image generation and embedding it into a U-shaped DNN as a prior decoder, then fine-tuning the GAN prior embedded DNN with a set of synthesized low-quality face images. The GAN blocks are designed to ensure that the latent code and noise input to the GAN can be respectively generated from the deep and shallow features of the DNN, controlling the global face structure, local face details and background of the reconstructed image. The proposed GAN prior embedded network (GPEN) is easy-to-implement, and it can generate visually photo-realistic results. Our experiments demonstrated that the proposed GPEN achieves significantly superior results to state-of-the-art BFR methods both quantitatively and qualitatively, especially for the restoration of severely degraded face images in the wild. The source code and models can be found at https://github.com/yangxy/GPEN.

Transfer learning with pre-training on large-scale datasets has played an increasingly significant role in computer vision and natural language processing recently. However, as there exist numerous application scenarios that have distinctive demands such as certain latency constraints and specialized data distributions, it is prohibitively expensive to take advantage of large-scale pre-training for per-task requirements. In this paper, we focus on the area of object detection and present a transfer learning system named GAIA, which could automatically and efficiently give birth to customized solutions according to heterogeneous downstream needs. GAIA is capable of providing powerful pre-trained weights, selecting models that conform to downstream demands such as latency constraints and specified data domains, and collecting relevant data for practitioners who have very few datapoints for their tasks. With GAIA, we achieve promising results on COCO, Objects365, Open Images, Caltech, CityPersons, and UODB which is a collection of datasets including KITTI, VOC, WiderFace, DOTA, Clipart, Comic, and more. Taking COCO as an example, GAIA is able to efficiently produce models covering a wide range of latency from 16ms to 53ms, and yields AP from 38.2 to 46.5 without whistles and bells. To benefit every practitioner in the community of object detection, we would release our pre-trained models and code.

Learning methods for relative camera pose estimation have been developed largely in isolation from classical geometric approaches. The question of how to integrate predictions from deep neural networks (DNNs) and solutions from geometric solvers, such as the 5-point algorithm, has as yet remained under-explored. In this paper, we present a novel framework that involves probabilistic fusion between the two families of predictions during network training, with a view to leveraging their complementary benefits in a learnable way. The fusion is achieved by learning the DNN uncertainty under explicit guidance by the geometric uncertainty, thereby learning to take into account the geometric solution in relation to the DNN prediction. Our network features a self-attention graph neural network, which drives the learning by enforcing strong interactions between different correspondences and potentially modeling complex relationships between points. We propose motion parmeterizations suitable for learning and show that our method achieves state-of-the-art performance on the challenging DeMoN and ScanNet datasets. While we focus on relative pose, we envision that our pipeline is broadly applicable for fusing classical geometry and deep learning.

Human volumetric capture is a long-standing topic in computer vision and computer graphics. Although high-quality results can be achieved using sophisticated off-line systems, real-time human volumetric capture of complex scenarios, especially using light-weight setups, remains challenging. In this paper, we propose a human volumetric capture method that combines temporal volumetric fusion and deep implicit functions. To achieve high-quality and temporal-continuous reconstruction, we propose dynamic sliding fusion to fuse neighboring depth observations together with topology consistency. Moreover, for detailed and complete surface generation, we propose detail-preserving deep implicit functions for RGBD input which can not only preserve the geometric details on the depth inputs but also generate more plausible texturing results. Results and experiments show that our method outperforms existing methods in terms of view sparsity, generalization capacity, reconstruction quality, and run-time efficiency.

Inferring 3D structure of a generic object from a 2D image is a long-standing objective of computer vision. Conventional approaches either learn completely from CAD-generated synthetic data, which have difficulty in inference from real images, or generate 2.5D depth image via intrinsic decomposition, which is limited compared to the full 3D reconstruction. One fundamental challenge lies in how to leverage numerous real 2D images without any 3D ground truth. To address this issue, we take an alternative approach with semi-supervised learning. That is, for a 2D image of a generic object, we decompose it into latent representations of category, shape and albedo, lighting and camera projection matrix, decode the representations to segmented 3D shape and albedo respectively, and fuse these components to render an image well approximating the input image. Using a category-adaptive 3D joint occupancy field (JOF), we show that the complete shape and albedo modeling enables us to leverage real 2D images in both modeling and model fitting. The effectiveness of our approach is demonstrated through superior 3D reconstruction from a single image, being either synthetic or real, and shape segmentation.

This paper presents a fully convolutional scene graph generation (FCSGG) model that detects objects and relations simultaneously. Most of the scene graph generation frameworks use a pre-trained two-stage object detector, like Faster R-CNN, and build scene graphs using bounding box features. Such pipeline usually has a large number of parameters and low inference speed. Unlike these approaches, FCSGG is a conceptually elegant and efficient bottom-up approach that encodes objects as bounding box center points, and relationships as 2D vector fields which are named as Relation Affinity Fields (RAFs). RAFs encode both semantic and spatial features, and explicitly represent the relationship between a pair of objects by the integral on a sub-region that points from subject to object. FCSGG only utilizes visual features and still generates strong results for scene graph generation. Comprehensive experiments on the Visual Genome dataset demonstrate the efficacy, efficiency, and generalizability of the proposed method. FCSGG achieves highly competitive results on recall and zero-shot recall with significantly reduced inference time.

In this paper, we present a conceptually simple, strong, and efficient framework for panoptic segmentation, called Panoptic FCN. Our approach aims to represent and predict foreground things and background stuff in a unified fully convolutional pipeline. In particular, Panoptic FCN encodes each object instance or stuff category into a specific kernel weight with the proposed kernel generator and produces the prediction by convolving the high-resolution feature directly. With this approach, instance-aware and semantically consistent properties for things and stuff can be respectively satisfied in a simple generate-kernel-then-segment workflow. Without extra boxes for localization or instance separation, the proposed approach outperforms previous box-based and -free models with high efficiency on COCO, Cityscapes, and Mapillary Vistas datasets with single scale input. Our code is made publicly available at https://github.com/Jia-Research-Lab/PanopticFCN.

Animal pose estimation is an important field that has received increasing attention in the recent years. The main challenge for this task is the lack of labeled data. Existing works circumvent this problem with pseudo labels generated from data of other easily accessible domains such as synthetic data. However, these pseudo labels are noisy even with consistency check or confidence-based filtering due to the domain shift in the data. To solve this problem, we design a multi-scale domain adaptation module (MDAM) to reduce the domain gap between the synthetic and real data. We further introduce an online coarse-to-fine pseudo label updating strategy. Specifically, we propose a self-distillation module in an inner coarse-update loop and a mean-teacher in an outer fine-update loop to generate new pseudo labels that gradually replace the old ones. Consequently, our model is able to learn from the old pseudo labels at the early stage, and gradually switch to the new pseudo labels to prevent overfitting in the later stage. We evaluate our approach on the TigDog and VisDA 2019 datasets, where we outperform existing approaches by a large margin. We also demonstrate the generalization ability of our model by testing extensively on both unseen domains and unseen animal categories. Our code is available at the project website.

Shadow removal is a computer-vision task that aims to restore the image content in shadow regions. While almost all recent shadow-removal methods require shadow-free images for training, in ECCV 2020 Le and Samaras introduces an innovative approach without this requirement by cropping patches with and without shadows from shadow images as training samples. However, it is still laborious and time-consuming to construct a large amount of such unpaired patches. In this paper, we propose a new G2R-ShadowNet which leverages shadow generation for weakly-supervised shadow removal by only using a set of shadow images and their corresponding shadow masks for training. The proposed G2R-ShadowNet consists of three sub-networks for shadow generation, shadow removal and refinement, respectively and they are jointly trained in an end-to-end fashion. In particular, the shadow generation sub-net stylises non-shadow regions to be shadow ones, leading to paired data for training the shadow-removal sub-net. Extensive experiments on the ISTD dataset and the Video Shadow Removal dataset show that the proposed G2R-ShadowNet achieves competitive performances against the current state of the arts and outperforms Le and Samaras' patch-based shadow-removal method.

Thanks to the rapid advances in the deep learning techniques and the wide availability of large-scale training sets, the performances of video saliency detection models have been improving steadily and significantly. However, the deep learning based visual-audio fixation prediction is still in its infancy. At present, only a few visual-audio sequences have been furnished with real fixations being recorded in the real visual-audio environment. Hence, it would be neither efficiency nor necessary to re-collect real fixations under the same visual-audio circumstance. To address the problem, this paper advocate a novel approach in a weakly-supervised manner to alleviating the demand of large-scale training sets for visual-audio model training. By using the video category tags only, we propose the selective class activation mapping (SCAM), which follows a coarse-to-fine strategy to select the most discriminative regions in the spatial-temporal-audio circumstance. Moreover, these regions exhibit high consistency with the real human-eye fixations, which could subsequently be employed as the pseudo GTs to train a new spatial-temporal-audio (STA) network. Without resorting to any real fixation, the performance of our STA network is comparable to that of the fully supervised ones.

For the single image rain removal (SIRR) task, the performance of deep learning (DL)-based methods is mainly affected by the designed deraining models and training datasets. Most of current state-of-the-art focus on constructing powerful deep models to obtain better deraining results. In this paper, to further improve the deraining performance, we novelly attempt to handle the SIRR task from the perspective of training datasets by exploring a more efficient way to synthesize rainy images. Specifically, we build a full Bayesian generative model for rainy image where the rain layer is parameterized as a generator with the input as some latent variables representing the physical structural rain factors, e.g., direction, scale, and thickness. To solve this model, we employ the variational inference framework to approximate the expected statistical distribution of rainy image in a data-driven manner. With the learned generator, we can automatically and sufficiently generate diverse and non-repetitive training pairs so as to efficiently enrich and augment the existing benchmark datasets. User study qualitatively and quantitatively evaluates the realism of generated rainy images. Comprehensive experiments substantiate that the proposed model can faithfully extract the complex rain distribution that not only helps significantly improve the deraining performance of current deep single image derainers, but also largely loosens the requirement of large training sample pre-collection for the SIRR task. Code is available in https://github.com/hongwang01/VRGNet.

We propose a method to detect and reconstruct multiple 3D objects from a single RGB image. The key idea is to optimize for detection, alignment and shape jointly over all objects in the RGB image, while focusing on realistic and physically plausible reconstructions. To this end, we propose a key-point detector that localizes objects as center points and directly predicts all object properties, including 9-DoF bounding boxes and 3D shapes, all in a single forward pass. The method formulates 3D shape reconstruction as a shape selection problem, i.e. it selects among exemplar shapes from a given database. This makes it agnostic to shape representations, which enables a lightweight reconstruction of realistic and visually-pleasing shapes based on CAD-models, while the training objective is formulated around point clouds and voxel representations. A collision-loss promotes non-intersecting objects, further increasing the reconstruction realism. Given the RGB image, the presented approach performs lightweight reconstruction in a single-stage, it is real-time capable, fully differentiable and end-to-end trainable. Our experiments compare multiple approaches for 9-DoF bounding box estimation, evaluate the novel shape-selection mechanism and compare to recent methods in terms of 3D bounding box estimation and 3D shape reconstruction quality.

Face forgery detection is raising ever-increasing interest in computer vision since facial manipulation technologies cause serious worries. Though recent works have reached sound achievements, there are still unignorable problems: a) learned features supervised by softmax loss are separable but not discriminative enough, since softmax loss does not explicitly encourage intra-class compactness and interclass separability; and b) fixed filter banks and hand-crafted features are insufficient to capture forgery patterns of frequency from diverse inputs. To compensate for such limitations, a novel frequency-aware discriminative feature learning framework is proposed in this paper. Specifically, we design a novel single-center loss (SCL) that only compresses intra-class variations of natural faces while boosting interclass differences in the embedding space. In such a case, the network can learn more discriminative features with less optimization difficulty. Besides, an adaptive frequency feature generation module is developed to mine frequency clues in a completely data-driven fashion. With the above two modules, the whole framework can learn more discriminative features in an end-to-end manner. Extensive experiments demonstrate the effectiveness and superiority of our framework on three versions of the FF++ dataset.

In this paper, we propose a conditional early exiting framework for efficient video recognition. While existing works focus on selecting a subset of salient frames to reduce the computation costs, we propose to use a simple sampling strategy combined with conditional early exiting to enable efficient recognition. Our model automatically learns to process fewer frames for simpler videos and more frames for complex ones. To achieve this, we employ a cascade of gating modules to automatically determine the earliest point in processing where an inference is sufficiently reliable. We generate on-the-fly supervision signals to the gates to provide a dynamic trade-off between accuracy and computational cost. Our proposed model outperforms competing methods on three large-scale video benchmarks. In particular, on ActivityNet1.3 and mini-kinetics, we outperform the state-of-the-art efficient video recognition methods with 1.3x and 2.1x less GFLOPs, respectively. Additionally, our method sets a new state of the art for efficient video understanding on the HVU benchmark.

One of the main challenges for arbitrary-shaped text detection is to design a good text instance representation that allows networks to learn diverse text geometry variances. Most of existing methods model text instances in image spatial domain via masks or contour point sequences in the Cartesian or the polar coordinate system. However, the mask representation might lead to expensive post-processing, while the point sequence one may have limited capability to model texts with highly-curved shapes. To tackle these problems, we model text instances in the Fourier domain and propose one novel Fourier Contour Embedding (FCE) method to represent arbitrary shaped text contours as compact signatures. We further construct FCENet with a backbone, feature pyramid networks (FPN) and a simple post-processing with the Inverse Fourier Transformation (IFT) and Non-Maximum Suppression (NMS). Different from previous methods, FCENet first predicts compact Fourier signatures of text instances, and then reconstructs text contours via IFT and NMS during test. Extensive experiments demonstrate that FCE is accurate and robust to fit contours of scene texts even with highly-curved shapes, and also validate the effectiveness and the good generalization of FCENet for arbitrary-shaped text detection. Furthermore, experimental results show that our FCENet is superior to the state-of-the-art (SOTA) methods on CTW1500 and Total-Text, especially on challenging highly-curved text subset.

The problem of grounding VQA tasks has seen an increased attention in the research community recently, with most attempts usually focusing on solving this task by using pretrained object detectors. However, pre-trained object detectors require bounding box annotations for detecting relevant objects in the vocabulary, which may not always be feasible for real-life large-scale applications. In this paper, we focus on a more relaxed setting: the grounding of relevant visual entities in a weakly supervised manner by training on the VQA task alone. To address this problem, we propose a visual capsule module with a query-based selection mechanism of capsule features, that allows the model to focus on relevant regions based on the textual cues about visual information in the question. We show that integrating the proposed capsule module in existing VQA systems significantly improves their performance on the weakly supervised grounding task. Overall, we demonstrate the effectiveness of our approach on two state-of-the-art VQA systems, stacked NMN and MAC, on the CLEVR-Answers benchmark, our new evaluation set based on CLEVR scenes with ground truth bounding boxes for objects that are relevant for the correct answer, as well as on GQA, a real world VQA dataset with compositional questions. We show that the systems with the proposed capsule module consistently outperform the respective baseline systems in terms of answer grounding, while achieving comparable performance on VQA task.

Single-view 3D object reconstruction has seen much progress, yet methods still struggle generalizing to novel shapes unseen during training. Common approaches predominantly rely on learned global shape priors and, hence, disregard detailed local observations. In this work, we address this issue by learning a hierarchy of priors at different levels of locality from ground truth input depth maps. We argue that exploiting local priors allows our method to efficiently use input observations, thus improving generalization in visible areas of novel shapes. At the same time, the combination of local and global priors enables meaningful hallucination of unobserved parts resulting in consistent 3D shapes. We show that the hierarchical approach generalizes much better than the global approach. It generalizes not only between different instances of a class but also across classes and to unseen arrangements of objects.

The rapid progress of photorealistic synthesis techniques has reached at a critical point where the boundary between real and manipulated images starts to blur. Thus, benchmarking and advancing digital forgery analysis have become a pressing issue. However, existing face forgery datasets either have limited diversity or only support coarse-grained analysis. To counter this emerging threat, we construct the ForgeryNet dataset, an extremely large face forgery dataset with unified annotations in image- and video-level data across four tasks: 1) Image Forgery Classification, including two-way (real / fake), three-way (real / fake with identity-replaced forgery approaches / fake with identity-remained forgery approaches), and n-way (real and 15 respective forgery approaches) classification. 2) Spatial Forgery Localization, which segments the manipulated area of fake images compared to their corresponding source real images. 3) Video Forgery Classification, which re-defines the video-level forgery classification with manipulated frames in random positions. This task is important because attackers in real world are free to manipulate any target frame. and 4) Temporal Forgery Localization, to localize the temporal segments which are manipulated. ForgeryNet is by far the largest publicly available deep face forgery dataset in terms of data-scale (2.9 million images, 221,247 videos), manipulations (7 image-level approaches, 8 video-level approaches), perturbations (36 independent and more mixed perturbations) and annotations (6.3 million classification labels, 2.9 million manipulated area annotations and 221,247 temporal forgery segment labels). We perform extensive benchmarking and studies of existing face forensics methods and obtain several valuable observations.We hope that the scale, quality, and variety of ForgeryNet dataset will foster further research and innovation in the area of face forgery classification, spatial and temporal forgery localization etc.

Forecasting Parapapillary atrophy (PPA), i.e., a symptom related to most irreversible eye diseases, provides an alarm for implementing an intervention to slow down the disease progression at early stage. A key question for this forecast is: how to fully utilize the historical data (e.g., retinal image) up to the current stage for future disease prediction? In this paper, we provide an answer with a novel framework, namely Disease Forecast via Progression Learning (DFPL), which exploits the irreversibility prior (i.e., cannot be reversed once diagnosed). Specifically, based on this prior, we decompose two factors that contribute to the prediction of the future disease: i) the current disease label given the data (retinal image, clinical attributes) at present and ii) the future disease label given the progression of the retinal images that from the current to the future. To model these two factors, we introduce the current and progression predictors in DFPL, respectively. In order to account for the degree of progression of the disease, we propose a temporal generative model to accurately generate the future image and compare it with the current one to get a residual image. The generative model is implemented by a recurrent neural network, in order to exploit the dependency of the historical data. To verify our approach, we apply it to a PPA in-house dataset and it yields a significant improvement (e.g., 4.48% of accuracy; 3.45% of AUC) over others. Besides, our generative model can accurately localize the disease-related regions.