Find dense correspondence between high resolution non-rigid 3D human faces

Abstract

It’s a very complex problem to achieve dense correspondence between high resolution 3D human faces. Solving the problem can contribute to a variety of computer vision tasks. This paper proposed an automatic method to find dense correspondence between different high resolution non-rigid 3D human faces. The main idea of this method is to use the correspondent facial feature points to generate Möbius transformations and using these Möbius transformations to achieve sparse correspondence between 3D faces. The texture and shape information of 3D face are used to locate the facial feature points. TPS (Thin-Plate Spline) transformation is used to represent the deformation of 3D faces, the TPS control points are selected from the sparse correspondence set. After performing TPS warping, for every vertex of the warped reference 3D face, we project them to every triangle face of the sample 3D face and use the closest projections to define the new mesh vertices of the sample 3D face. The sample 3D face with new mesh shares the same connectivity with the reference 3D face, thus the dense correspondence between the reference 3D face and the sample 3D face with new mesh is achieved. The experiments on BJUT-3D face databases show that our method achieves better performance than existing methods.