团队简介   研究论文   研究人员   发明专利   研究资助   联系我们
单张图片自动重建带几何细节的人脸形状

王 涵 1,2   夏时洪 1

1 (中国科学院计算技术研究所前瞻研究实验室/移动计算与新型终端北京市重点实验室 北京 100190) 
2 (中国科学院大学 北京 100049)

摘 要
         本文提供了一个从单张图片中自动快速重建高细节人脸模型的方法. 本方法的优点在于对生成结果由粗到 细进行逐步优化, 这样既降低了每一步的图像条件, 又可以尽可能挖掘图像信息来增加重建模型的逼真度. 首先, 本 文从单张照片检测到的特征点中, 通过多初值迭代方法优化求解对应的三维头部姿态和大尺度的人脸表情. 其次以 检测到的人脸特征点为依据对不准确的人脸表情进行矫正, 期间使用非刚性的迭代最近点方法对齐模型上的特征点 和图像特征点, 使用拉普拉斯坐标影响其余非特征点位置; 最后使用带有常量假设的明暗重建形状方法为人脸模型 重建细尺度的几何细节, 以增加重建模型的逼真度. 实验结果表明, 该方法能够在大幅度头部摆动的情况下, 生成准 确逼真带几何有细节的人脸模型. 此方法不需要对图片的摄像机进行标定, 也不需要预先训练或预设用户的特征模 型和混合形状, 同时不强制室内光照以及单调的背景环境, 因此该法适用于各种人脸图片的三维模型重建.
Construction of Facial Shape with Details from Single Image Title

Wang Han 1,2   Xia Shihong 1

1 (Beijing Key Lab of Mobile Computing and Pervasive Devices, Advanced Computing Research Laboratory, Institute Of Computing Technology Chinese Academy Of Sciences, Beijing, 100190) 
2 (University of Chinese Academy of Sciences, Beijing, 100049)

Abstract:
This paper proposes an automatic method on reconstructing detailed facial model with a single image. The virtue of this work is the idea of improving the geometry from coarse to fine. This method needs weaker conditions in image and digs more information from image to improve the reality of 3D facial geometry. After we detected feature points from image, this method is started by solving head pose and large-scale expression using different initials. Then, we correct the 3D shape by 2D feature points. This step is combined with non-rigid ICP method and a Laplacian deformation method. Finally, we generate a geometry of detail from the coarse shape using shape from shading method. In this step, our method provide a constant hypothesis to solve the boundary problem. We show results to demonstrate that the method can construct accuracy facial shape with details even when the head swing largely. This method does not need to calibrate the camera, neither needs the pre-process like user-training or some other manual work. Meanwhile, the method can achieve good result using a weak lighting condition or in a variegated background. This gives our method a wide range of applications.
Downloads

单张图片自动重建带几何细节的人脸形状.计算机辅助设计与图形学学报.(中国计算机图形学大会 2016) [PDF 2,836KB]