Matching 2.5D face scans to 3D models

The performance of face recognition systems that use two-dimensional images depends on factors such as lighting and subject's pose. We are developing a face recognition system that utilizes three-dimensional shape information to make the system more robust to arbitrary pose and lighting. For each subject, a 3D face model is constructed by integrating several 2.5D face scans which are captured from different views. 2.5D is a simplified 3D (x,y,z) surface representation that contains at most one depth value (z direction) for every point in the (x, y) plane. Two different modalities provided by the facial scan, namely, shape and texture, are utilized and integrated for face matching. The recognition engine consists of two components, surface matching and appearance-based matching. The surface matching component is based on a modified iterative closest point (ICP) algorithm. The candidate list from the gallery used for appearance matching is dynamically generated based on the output of the surface matching component, which reduces the complexity of the appearance-based matching stage. Three-dimensional models in the gallery are used to synthesize new appearance samples with pose and illumination variations and the synthesized face images are used in discriminant subspace analysis. The weighted sum rule is applied to combine the scores given by the two matching components. Experimental results are given for matching a database of 200 3D face models with 598 2.5D independent test scans acquired under different pose and some lighting and expression changes. These results show the feasibility of the proposed matching scheme.     

View-based active appearance models

We demonstrate that a small number of 2D linear statistical models are sufficient to capture the shape and appearance of a face from a wide range of viewpoints. Such models can be used to estimate head orientation and track faces through large angles. Given multiple images of the same face we can learn a coupled model describing the relationship between the frontal appearance and the profile of a face. This relationship can be used to predict new views of a face seen from one view and to constrain search algorithms which seek to locate a face in multiple views simultaneously.     

基于深度对齐网络的生成对抗网络伪造人脸检测

针对现有的生成对抗网络（GAN）伪造人脸图像检测方法在有角度及遮挡情况下存在的真实人脸误判问题,提出了一种基于深度对齐网络（DAN）的GAN伪造人脸图像检测方法。首先,基于DAN设计面部关键点提取网络,以提取真伪人脸关键点位置;然后,采用主成分分析（PCA）方法将每一组关键点映射到三维空间,从而减少冗余信息以及降低特征维度;最后,利用支持向量机（SVM）五折交叉验证对特征进行分类,并计算准确率。实验结果表明,该方法通过提高面部关键点定位准确度改善了由于定位误差引起的面部不协调问题,进而降低了真实人脸误判率。与VGG19、XceptionNet和Dlib-SVM方法相比,正脸情况下,该方法的ROC下面积（AUC）值提高了4.48到32.96个百分点,平均精度（AP）提高了4.26到33.12个百分点;有角度及遮挡人脸情况下,该方法的AUC值提高了10.56到30.75个百分点,AP提高了7.42到42.45个百分点。     

基于AAM和T型结构的人脸3D姿态估计

在分析已有的人脸姿态估计方法基础上,提出了一种基于主动表观模型(AAM)和T型结构的人脸3D姿态估计方法。对多姿态的人脸样本进行训练,得到多姿态的AAM模板集;利用训练得到的多姿态的AAM模板集进行最佳模板匹配,并对人脸的特征点进行精确定位;用人脸的双眼和嘴部构建T型模型,进行人脸3D姿态的参数估计。实验结果表明,该方法能适应较大的姿态旋转角度,并具有良好的姿态估计精度。     

Gabor texture in active appearance models

In computer vision applications, Active Appearance Models (AAMs) is usually used to model the shape and the gray-level appearance of an object of interest using statistical methods, such as PCA. However, intensity values used in standard AAMs cannot provide enough information for image alignment. In this paper, we firstly propose to utilize Gabor filters to represent the image texture. The benefit of Gabor-based representation is that it can express local structures of an image. As a result, this representation can lead to more accurate matching when condition changes. Given the problem of the excessive storage and computational complexity of the Gabor, three different Gabor-based image representations are used in AAMs: (1) GaborD is the sum of Gabor filter responses over directions, (2) GaborS is the sum of Gabor filter responses over scales, and (3) GaborSD is the sum of Gabor filter responses over scales and directions. Through a large number of experiments, we show that the proposed Gabor representations lead to more accurate and robust matching between model and images.     

Real-time 3D face tracking based on active appearance model constrained by depth data

Active Appearance Model (AAM) is an algorithm for fitting a generative model of object shape and appearance to an input image. AAM allows accurate, real-time tracking of human faces in 2D and can be extended to track faces in 3D by constraining its fitting with a linear 3D morphable model. Unfortunately, this AAM-based 3D tracking does not provide adequate accuracy and robustness, as we show in this paper. We introduce a new constraint into AAM fitting that uses depth data from a commodity RGBD camera (Kinect). This addition significantly reduces 3D tracking errors. We also describe how to initialize the 3D morphable face model used in our tracking algorithm by computing its face shape parameters of the user from a batch of tracked frames. The described face tracking algorithm is used in Microsoft's Kinect system.     

多视角2.5维人脸数据的自动配准与三维融合*

提出了一种有效的从三视角2.5维人脸图像到完整三维人脸模型的融合方法。首先用ICP(迭代最近点)方法对三视角人脸图像上手动选取的特征区域进行粗配准,然后用一种调整能量最优方法进行精确配准,最终合成完整的三维人脸模型。通过对融合结果进行相似度测量,实验说明了方法的有效性和优越性。     

2.5D face recognition using Patch Geodesic Moments

In this paper, we propose a novel Patch Geodesic Distance (PGD) to transform the texture map of an object through its shape data for robust 2.5D object recognition. Local geodesic paths within patches and global geodesic paths for patches are combined in a coarse to fine hierarchical computation of PGD for each surface point to tackle the missing data problem in 2.5D images. Shape adjusted texture patches are encoded into local patterns for similarity measurement between two 2.5D images with different viewing angles and/or shape deformations. An extensive experimental investigation is conducted on 2.5 face images using the publicly available BU-3DFE and Bosphorus databases covering face recognition under expression and pose changes. The performance of the proposed method is compared with that of three benchmark approaches. The experimental results demonstrate that the proposed method provides a very encouraging new solution for 2.5D object recognition.     

Towards a low bandwidth talking face using appearance models

This paper is motivated by the need to develop low bandwidth virtual humans capable of delivering audio-visual speech and sign language at a quality comparable to high bandwidth video. Using an appearance model combined with parameter compression significantly reduces the number of bits required for animating the face of a virtual human. A perceptual method is used to evaluate the quality of the synthesised sequences and it appears that 3.6 kb s⁻¹ can yield acceptable quality.     

A comparative study of facial appearance modeling methods for active appearance models

Active appearance models (AAMs) have been widely used in many face modeling and facial feature extraction methods. One of the problems of AAMs is that it is difficult to model a sufficiently wide range of human facial appearances, the pattern of intensities across a face image patch. Previous researches have used principal component analysis (PCA) for facial appearance modeling, but there has been little analysis and comparison between PCA and many other facial appearance modeling methods such as non-negative matrix factorization (NMF), local NMF (LNMF), and non-smooth NMF (ns-NMF). The main contribution of this paper is to find a suitable facial appearance modeling method for AAMs by a comparative study. In the experiments, PCA, NMF, LNMF, and ns-NMF were used to produce the appearance model of the AAMs and the root mean square (RMS) errors of the detected feature points were analyzed using the AR and BERC face databases. Experimental results showed that (1) if the appearance variations of testing face images were relatively non-sparser than those of training face images, the non-sparse methods (PCA, NMF) based AAMs outperformed the sparse methods (nsNMF, LNMF) based AAMs. (2) If the appearance variations of testing face images are relatively sparser than those of training face images, the sparse methods (nsNMF) based AAMs outperformed the non-sparse methods (PCA, NMF) based AAMs.     

基于改进三维形变模型的三维人脸重建和密集人脸对齐方法

针对现在广泛使用的三维形变模型表达能力不够,导致重建出的三维人脸模型泛化性能不佳的问题,提出了一种在姿态、表情和光照未知的条件下的基于单张人脸图片的三维人脸重建和密集人脸对齐的新方法。首先,通过卷积神经网络对现有的三维形变模型进行改进,以提高三维人脸模型的表达能力;然后,基于人脸光滑性和图像相似性,在特征点和像素层面提出新的损失函数,并使用弱监督学习训练卷积神经网络模型;最后,通过训练出的网络模型进行三维人脸重建和密集人脸对齐。实验结果表明,对于三维人脸重建任务,所提模型在AFLW2000-3D上实现了2.25的归一化平均误差;对于密集人脸对齐任务,所提模型在AFLW2000-3D和AFLW-LFPA上分别实现了3.80和3.34的归一化平均误差。与原始使用三维形变模型的方法相比,所提模型在三维人脸重建和密集人脸对齐上的归一化平均误差分别降低了7.4%和7.8%。针对不同光照环境以及角度的人脸图片,该网络模型的重建准确,鲁棒性好,且具有较高的三维人脸重建和密集人脸对齐质量。     

基于改进三维形变模型的三维人脸重建和密集人脸对齐方法

针对现在广泛使用的三维形变模型表达能力不够，导致重建出的三维人脸模型泛化性能不佳的问题，提出了一种在姿态、表情和光照未知的条件下的基于单张人脸图片的三维人脸重建和密集人脸对齐的新方法。首先，通过卷积神经网络对现有的三维形变模型进行改进，以提高三维人脸模型的表达能力；然后，基于人脸光滑性和图像相似性，在特征点和像素层面提出新的损失函数，并使用弱监督学习训练卷积神经网络模型；最后，通过训练出的网络模型进行三维人脸重建和密集人脸对齐。实验结果表明，对于三维人脸重建任务，所提模型在AFLW2000-3D上实现了2.25的归一化平均误差；对于密集人脸对齐任务，所提模型在AFLW2000-3D和AFLW-LFPA上分别实现了3.80和3.34的归一化平均误差。与原始使用三维形变模型的方法相比，所提模型在三维人脸重建和密集人脸对齐上的归一化平均误差分别降低了7.4%和7.8%。针对不同光照环境以及角度的人脸图片，该网络模型的重建准确，鲁棒性好，且具有较高的三维人脸重建和密集人脸对齐质量。     

Dense 3D face alignment from 2D video for real-time use

To enable real-time, person-independent 3D registration from 2D video, we developed a 3D cascade regression approach in which facial landmarks remain invariant across pose over a range of approximately 60°. From a single 2D image of a person's face, a dense 3D shape is registered in real time for each frame. The algorithm utilizes a fast cascade regression framework trained on high-resolution 3D face-scans of posed and spontaneous emotion expression. The algorithm first estimates the location of a dense set of landmarks and their visibility, then reconstructs face shapes by fitting a part-based 3D model. Because no assumptions are required about illumination or surface properties, the method can be applied to a wide range of imaging conditions that include 2D video and uncalibrated multi-view video. The method has been validated in a battery of experiments that evaluate its precision of 3D reconstruction, extension to multi-view reconstruction, temporal integration for videos and 3D head-pose estimation. Experimental findings strongly support the validity of real-time, 3D registration and reconstruction from 2D video. The software is available online at http://zface.org.     

Data driven image models through continuous joint alignment

This paper presents a family of techniques that we call congealing for modeling image classes from data. The idea is to start with a set of images and make them appear as similar as possible by removing variability along the known axes of variation. This technique can be used to eliminate "nuisance" variables such as affine deformations from handwritten digits or unwanted bias fields from magnetic resonance images. In addition to separating and modeling the latent images - i.e., the images without the nuisance variables - we can model the nuisance variables themselves, leading to factorized generative image models. When nuisance variable distributions are shared between classes, one can share the knowledge learned in one task with another task, leading to efficient learning. We demonstrate this process by building a handwritten digit classifier from just a single example of each class. In addition to applications in handwritten character recognition, we describe in detail the application of bias removal from magnetic resonance images. Unlike previous methods, we use a separate, nonparametric model for the intensity values at each pixel. This allows us to leverage the data from the MR images of different patients to remove bias from each other. Only very weak assumptions are made about the distributions of intensity values in the images. In addition to the digit and MR applications, we discuss a number of other uses of congealing and describe experiments about the robustness and consistency of the method.     

Multi-view face segmentation using fusion of statistical shape and appearance models

This paper demonstrates how a weighted fusion of multiple Active Shape (ASM) or Active Appearance (AAM) models can be utilized to perform multi-view facial segmentation with only a limited number of views available for training the models. The idea is to construct models only from frontal and profile views and subsequently fuse these models with adequate weights to segment any facial view. This reduces the problem of multi-view facial segmentation to that of weight estimation, the algorithm for which is proposed as well. The evaluation is performed on a set of 280 landmarked static face images corresponding to seven different rotation angles and on several video sequences of the AV@CAR database. The evaluation demonstrates that the estimation of the weights does not have to be very accurate in the case of ASM, while in the case of AAM the influence of correct weight estimation is more critical. The segmentation with the proposed weight estimation method produced accurate segmentations in 91% of 280 testing images with the median point-to-point error varying from two to eight pixels (1.8–7.2% of average inter-eye distance).     

Fast and reliable active appearance model search for 3-D face tracking

This paper addresses the three-dimensional (3-D) tracking of pose and animation of the human face in monocular image sequences using active appearance models. The major problem of the classical appearance-based adaptation is the high computational time resulting from the inclusion of a synthesis step in the iterative optimization. Whenever the dimension of the face space is large, a real-time performance cannot be achieved. In this paper, we aim at designing a fast and stable active appearance model search for 3-D face tracking. The main contribution is a search algorithm whose CPU-time is not dependent on the dimension of the face space. Using this algorithm, we show that both the CPU-time and the likelihood of a nonaccurate tracking are reduced. Experiments evaluating the effectiveness of the proposed algorithm are reported, as well as method comparison and tracking synthetic and real image sequences.     

基于改进主动表观模型算法的人脸特征定位

人脸特征点的精确定位一直是人脸图像处理的重要研究内容,特征点定位精确与否直接影响后续工作结果的好坏。在基于反向组合AAM（Active Appearance Models）人脸特征点定位算法的基础上,提出结合特征点局部纹理模型来对AAM初始形状参数做最优化以及对AAM匹配模板升级的改进。改进的算法采用特征点局部纹理模型和AAM全局纹理模型结合的方法来最优化AAM初始形状参数,并在此前提下对AAM匹配模板进行升级,使其更接近待匹配图像的信息。在精确的匹配模板和优化的初始形状参数下,匹配的最终精度会得到提升。实验和理论证明,改进后的算法比传统反向组合AAM算法以及现有改进的PAAM（Progressive AAM）算法以及简单的结合ASM和AAM的改进算法都有更好的特征点定位精度。     

Automatic construction of active appearance models as an image coding problem

The automatic construction of active appearance models (AAMs) is usually posed as finding the location of the base mesh vertices in the input training images. In this paper, we repose the problem as an energy-minimizing image coding problem and propose an efficient gradient-descent algorithm to solve it.     

3D face reconstruction and dense alignment with a new generated dataset

Face alignment and reconstruction are classical problems in the computer vision field, one of the greatest difficulties of which is the limited number of facial images with landmark points. The 300 W-LP dataset is the most commonly used for the existing methods of single-view 3D Morphable Model (3DMM)-based reconstruction; however, the model performance is limited by the small variety of facial images in this dataset. In this work, a 3D facial image dataset with landmark points generated by the rotate-and-render method is proposed. The key innovation of the proposed method is that the back-and-forth rotation of faces in 3D space and then re-rendering them to the 2D plane can provide strong self-supervision. The recent advances in 3D face modeling and high-resolution generative adversarial networks (GANs) are leveraged to constitute the blocks. To obtain more precise facial landmark points, the 3D dense face alignment (3DDFA) model is used to label the generated images and filter the landmark points. Finally, the 3DDFA model is retrained using the proposed dataset, and an improved result is achieved.