Automatic and robust semantic registration of 3D head scans

Abstract

A robust and fully automatic algorithm for semantic registration of laser scans of human faces is introduced. The scans are converted into a new semantically annotated topology such that topologically equivalent vertices have the same meaning. The algorithm works in three steps. First, landmark points are placed using a multi-stage variant of the iterative closest points scheme and an annotated reference mesh. Second, the scan is matched to the reference mesh globally by a nonlinear transformation. Finally, the transformed mesh is resampled in the reference mesh topology and the vertex locations are tracked back to the original scan. The use of the algorithm for the construction of a morphable head model is described. Moreover, a method for repairing defective head scans (i.e. scans with missing or corrupted vertex data) using the registration algorithm and the model is introduced and experimentally investigated. It is shown that a face can be reconstructed convincingly using as little as 20 percent of its vertices. (7 pages)