基于多视角的高精度三维人脸重建

High-Quality 3D Face Reconstruction from Multi-View Images

提出了一种多阶段优化的方法来解决基于多视角图片在未知姿态、表情以及光照条件下的高精度三维人脸重建问题.首先,通过重新渲染合成的方法将参数化模型拟合到输入的多视角图片,然后在纹理域上求解一个光流问题来获取不同视角之间的对应关系.通过对应关系可以恢复出人脸的点云,并利用基于明暗恢复几何的方法来恢复人脸细节.在真实数据以及合成数据下的实验结果表明,文中方法能够恢复出带有几何细节的高精度的三维人脸模型,并且提高了现有方法的重建精度.

This paper presents a method to reconstruct high-quality 3D face model from multi-view images with unknown poses,expressions,and illumination conditions.We adopt a multi-stage based optimization method to solve this challenging problem.Specifically,we first fit the input multi-view images with a parametric model by an analysis-by-synthesis method.Next,dense correspondence between input facial images is constructed by solving an optical flow problem on the albedo domain,and a point cloud with person-specified geometry characteristics is then recovered based on the reliable dense correspondence.In the final,fine-scale details are reconstructed by using a multi-view shape from shading method.Experiments on synthetic and real data demonstrate that our proposed method is able to reconstruct accurate 3D face models with fine geometric details,and the quantification studies show that our method is better than state-of-the-art methods.