An automatic non-rigid registration method for dense surface models

In this paper, we present an automatic registration method to align the dense points of 3D surfaces, for example, faces. It combines global and local deformations. An iterative closest point (ICP)-based deformation, which is global, is applied first to roughly align the two surfaces. Afterwards, the two surfaces will match in some regions, but not the whole. So, a sequence of compact support radial basis functions (CSRBF)-based deformations, which are local, will be then performed on those mismatched regions as adjustments. The CSRBF-based deformation is landmark-based. The landmarks are automatically defined using the bilateral correspondences of the two surfaces, with the aim to reduce the differences between the two correspondences through the adjustments. As a result, an optimal alignment of the two surfaces will be achieved finally when a unique correspondence is established. The results of our proposed method and those obtained by ICP and TPS are compared. Improvements on accuracy of registration can be easily seen from the comparison.