Recently, Convolutional Neural Network (CNN) based models have achieved great success in Single Image Super-Resolution (SISR). Owing to the strength of deep networks, these CNN learn an effective nonlinear mapping from low-resolution input image high-resolution target image, at cost requiring enormous parameters. This paper proposes a very model (up 52 convolutional layers) named Deep Recursive Residual (DRRN) that strives for yet concise networks. Specifically, residual learning is adopted,...

Recently, very deep convolutional neural networks (CNNs) have been attracting considerable attention in image restoration. However, as the depth grows, longterm dependency problem is rarely realized for these models, which results prior states/layers having little influence on subsequent ones. Motivated by fact that human thoughts persistency, we propose a persistent memory network (MemNet) introduces block, consisting of recursive unit and gate unit, to explicitly mine through an adaptive...

Face alignment, which fits a face model to an image and extracts the semantic meanings of facial pixels, has been important topic in CV community. However, most algorithms are designed for faces small medium poses (below 45), lacking ability align large up 90. The challenges three-fold: Firstly, commonly used landmark-based assumes that all landmarks visible is therefore not suitable profile views. Secondly, appearance varies more dramatically across poses, ranging from frontal view view....

The large pose discrepancy between two face images is one of the key challenges in recognition. Conventional approaches for pose-invariant recognition either perform frontalization on, or learn a representation from, non-frontal image. We argue that it more desirable to both tasks jointly allow them leverage each other. To this end, paper proposes Disentangled Representation learning-Generative Adversarial Network (DR-GAN) with three distinct novelties. First, encoder-decoder structure...

Face anti-spoofing is crucial to prevent face recognition systems from a security breach. Previous deep learning approaches formulate as binary classification problem. Many of them struggle grasp adequate spoofing cues and generalize poorly. In this paper, we argue the importance auxiliary supervision guide toward discriminative generalizable cues. A CNN-RNN model learned estimate depth with pixel-wise supervision, rPPG signals sequence-wise supervision. The estimated are fused distinguish...

Face Super-Resolution (SR) is a domain-specific superresolution problem. The facial prior knowledge can be leveraged to better super-resolve face images. We present novel deep end-to-end trainable Network (FSRNet), which makes use of the geometry prior, i.e., landmark heatmaps and parsing maps, very low-resolution (LR) images without well-aligned requirement. Specifically, we first construct coarse SR network recover high-resolution (HR) image. Then, HR image sent two branches: fine encoder...

Detecting manipulated facial images and videos is an increasingly important topic in digital media forensics. As advanced face synthesis manipulation methods are made available, new types of fake representations being created which have raised significant concerns for their use social media. Hence, it crucial to detect localize regions. Instead simply using multi-task learning simultaneously predict the mask (regions), we propose utilize attention mechanism process improve feature maps...

Understanding the world in 3D is a critical component of urban autonomous driving. Generally, combination expensive LiDAR sensors and stereo RGB imaging has been paramount for successful object detection algorithms, whereas monocular image-only methods experience drastically reduced performance. We propose to reduce gap by reformulating problem as standalone region proposal network. leverage geometric relationship 2D perspectives, allowing boxes utilize well-known powerful convolutional...

As an emerging topic in face recognition, designing margin-based loss functions can increase the feature margin between different classes for enhanced discriminability. More recently, idea of mining-based strategies is adopted to emphasize misclassified samples, achieving promising results. However, during entire training process, prior methods either do not explicitly sample based on its importance that renders hard samples fully exploited; or effects semi-hard/hard even at early stage may...

As a classic statistical model of 3D facial shape and texture, Morphable Model (3DMM) is widely used in analysis, e.g., fitting, image synthesis. Conventional 3DMM learned from set well-controlled 2D face images with associated scans, represented by two sets PCA basis functions. Due to the type amount training data, as well linear bases, representation power can be limited. To address these problems, this paper proposes an innovative framework learn nonlinear large unconstrained images,...

The face image is the most accessible biometric modality which used for highly accurate recognition systems, while it vulnerable to many different types of presentation attacks. Face anti-spoofing a very critical step before feeding systems. In this paper, we propose novel two-stream CNN-based approach anti-spoofing, by extracting local features and holistic depth maps from images. facilitate CNN discriminate spoof patches independent spatial areas. On other hand, map examine whether input...

Face alignment, which fits a face model to an image and extracts the semantic meanings of facial pixels, has been important topic in CV community. However, most algorithms are designed for faces small medium poses (below 45 degree), lacking ability align large up 90 degree. The challenges three-fold: Firstly, commonly used landmark-based assumes that all landmarks visible is therefore not suitable profile views. Secondly, appearance varies more dramatically across poses, ranging from frontal...

Despite recent advances in face recognition using deep learning, severe accuracy drops are observed for large pose variations unconstrained environments. Learning pose-invariant features is one solution, but needs expensively labeled large-scale data and carefully designed feature learning algorithms. In this work, we focus on frontalizing faces the wild under various head poses, including extreme profile view's. We propose a novel 3D Morphable Model (3DMM) conditioned Face Frontalization...

Demographic estimation entails automatic of age, gender and race a person from his face image, which has many potential applications ranging forensics to social media. Automatic demographic estimation, particularly age remains challenging problem because persons belonging the same group can be vastly different in their facial appearances due intrinsic extrinsic factors. In this paper, we present generic framework for (age, race) estimation. Given first extract informative features via...

Despite the large volume of face recognition datasets, there is a significant portion subjects, which samples are insufficient and thus under-represented. Ignoring such results in training data. Training with under-represented data leads to biased classifiers conventionally-trained deep networks. In this paper, we propose center-based feature transfer framework augment space subjects from regular that have sufficiently diverse samples. A Gaussian prior variance assumed across all ones...

RNN-based approaches have achieved outstanding performance on action recognition with skeleton inputs. Currently these methods limit their inputs to coordinates of joints and improve the accuracy mainly by extending RNN models spatial domains in various ways. While such explore relations between different parts directly from joint coordinates, we provide a simple universal modeling method perpendicular model enhancement. Specifically, select set geometric features, motivated evolution...

Recognition in low quality face datasets is challenging because facial attributes are obscured and degraded. Advances margin-based loss functions have resulted enhanced discriminability of faces the embedding space. Further, previous studies studied effect adaptive losses to assign more importance misclassified (hard) examples. In this work, we introduce another aspect adaptiveness function, namely image quality. We argue that strategy emphasize samples should be adjusted according their...

Large-pose face alignment is a very challenging problem in computer vision, which used as prerequisite for many important vision tasks, e.g, recognition and 3D reconstruction. Recently, there have been few attempts to solve this problem, but still more research needed achieve highly accurate results. In paper, we propose method large-pose images, by combining the powerful cascaded CNN regressor 3DMM. We formulate 3DMM fitting where camera projection matrix shape parameters are estimated...

Pedestrian detection is a critical problem in computer vision with significant impact on safety urban autonomous driving. In this work, we explore how semantic segmentation can be used to boost pedestrian accuracy while having little no network efficiency. We propose infusion enable joint supervision and detection. When placed properly, the additional helps guide features shared layers become more sophisticated helpful for downstream detector. Using approach, find weakly annotated boxes...

Gait, the walking pattern of individuals, is one most important biometrics modalities. Most existing gait recognition methods take silhouettes or articulated body models as features. These suffer from degraded performance when handling confounding variables, such clothing, carrying and view angle. To remedy this issue, we propose a novel AutoEncoder framework to explicitly disentangle pose appearance features RGB imagery LSTM-based integration over time produces feature. In addition, collect...

Face anti-spoofing is designed to keep face recognition systems from recognizing fake faces as the genuine users. While advanced methods are developed, new types of spoof attacks also being created and becoming a threat all existing systems. We define detection unknown Zero-Shot Anti-spoofing (ZSFA). Previous works ZSFA only study 1-2 attacks, such print/replay which limits insight this problem. In work, we expand problem wide range 13 including print attack, replay 3D mask so on. A novel...

Massive open online courses and other forms of remote education continue to increase in popularity reach. The ability efficiently proctor examinations is an important limiting factor the scalability this next stage education. Presently, human proctoring most common approach evaluation, by either requiring test taker visit examination center, or monitoring them visually acoustically during exams via a webcam. However, such methods are labor intensive costly. In paper, we present multimedia...

Embedding 3D morphable basis functions into deep neural networks opens great potential for models with better representation power. However, to faithfully learn those from an image collection, it requires strong regularization overcome ambiguities involved in the learning process. This critically prevents us high fidelity face which are needed represent images level of details. To address this problem, paper presents a novel approach additional proxies as means side-step regularizations,...