超球流形约束下的3DMM分步全局优化  

Multi-stage Global Optimization for 3DMM under Hypersphere Manifold Constraint

作　　者：高巍[1,2] 沈晔湖[1] 貊睿[1] 魏磊[1] 朱怡[1] 彭振云[1] 

机构地区：[1]中国科学院苏州纳米技术与纳米仿生研究所系统集成研究部,苏州215123 [2]中国科学院研究生院,北京100049

出　　处：《计算机辅助设计与图形学学报》2012年第8期1065-1075,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：江苏省科技支撑计划社会发展项目(09EK021001);中国科学院苏州纳米技术与纳米仿生研究所科技计划项目(08CA011001)

摘　　要：通过综合运用人脸空间的超球流形约束、基于梯度的启发式全局优化、光照的球面谐波描述以及凸包可见点集的直接消隐方法,提出一种三维可形变模型的图像匹配方法.首先通过形状超球流形约束下的全局优化算法求解摄像机参数和形状参数,然后使用以上参数和凸包点集的直接消隐方法确定物像点对应关系,最后根据物像点对应关系由反射率超球流形约束下的全局优化算法求解光照参数和反射率参数.定量的对比实验结果表明,该方法无需借助分区域拟合、人为估计参数值、层次匹配策略或复杂的特征组合,即可由单幅图像恢复三维可形变模型(3DMM)的全部参数.A method for fitting 3D morphable model to images is proposed by combining hypersphere manifold constraint of face space, gradient-based heuristic global optimization, spherical harmonic representation of illumination, and direct visibility of convex hull points. First, camera and shape parameters are solved by the global optimization algorithm under hypersphere manifold constraint of shape. Second, correspondences of object points and image points are determined by these parameters and the direct visibility of convex hull points. Finally, according to these correspondences, illumination and albedo parameters are computed by the global optimization algorithm under hypersphere manifold constraint of albedo. Quantitative experiment results show that the proposed method can recovers all parameters of 3DMM from a single image without segmented fitting, artificial estimation of parameters, coarse-to-fine strategy or complex combination of features.

关 键 词：三维可形变模型 全局优化 超球流形 球面谐波 人脸空间 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]