基于图割及动态片结构的3维人脸多视图体重建

Multi-view stereo reconstruction of volumetric human face via graph-cuts and dynamic patch structure

针对3维人脸重建问题提出了一种新颖的多视图体重建方法,以解决目前3维人脸重建方法只适用于小样本集合,大范围推广时精度难以保证的弱点。该方法创新之处在于将基于特征点匹配的重建方法与立体重建方法结合引入到图割优化框架,并应用于3维人脸重建。本文两个重要改进工作是设计动态片结构描述来进行颜色一致性估计以及设计新的动态图结构以去除半个体素尺寸的重建误差。实验中分别采用8张、16张和30张存在亮度变化的人脸多视角图像验证算法。实验结果逼真,同时避免了传统重建方法结果受限于样本集分布的问题。

In this paper, a novel approach is proposed to reconstruct human face model from muhi-view images with volumetric scene representation. This method focuses on improving the reconstruction results without the limit of the representative ability of 3D facial model samples. Our main contribution is to combine feature points based reconstruction and volumetric based reconstruction into the framework of Graph Cuts when applying facial reconstruction. Two considerations contribute to the improvement of final results. First, a dynamic patch based photo consistency estimation is designed to get the value of the photo-consistency constrain. Second, a dynamic weighted graph is constructed in order to avoid half voxel size error. We evaluate the performance on a real dataset, in which multi-view face images （8, 16, about 30） are captured by moving digital camera under changed illumination. The results indicate our approach have significant improvements.