A hybrid deformation model for the non-rigid registration of 3d scans

Abstract

3D model matching (non-rigid registration) is one of the most important problems in computer graphics. Of particular interest is the non-rigid registration of 3D facial scans, which are then used in animations. There are a lot of non-rigid registration algorithms, but most of them cannot function adequately when the input models differ greatly. This happens, for example, when it is necessary to register a neutral face and a screaming face. In this article we suggest an iterative algorithm for finding the closest pair of points modified in order to provide for the highly accurate registration of 3D models which differ greatly. The key components of the proposed algorithm are a hybrid deformation model based on an embedded deformation graph which can capture small details, and a blendshape allowing a single mesh to deform to achieve numerous pre-defined shapes.