Successive convex optimization for point set registration

Abstract

We propose a new robust point set registration technique utilizing the global information of the two sets, rather than looking locally, for constructing the point-wise correspondence and estimating the transformation in between. Then, a multi-scale scheme further extends the algorithm to handling nonlinear diffeomorphic transformations. The algorithm is tested on 3D geometric and protein structure data sets to demonstrate its robustness to severe distortions.