Automatic Transfer of Landmarks on Human Skulls using GPU-based Non-rigid Registration

Abstract

In this work, we present a novel approach to automatically transfer landmarks from a template mesh of a human skull to other meshes obtained via 3D scanning. Since previous methods rely on user input or only work on a subset of the data, these algorithms are not suited for large databases. Our system is designed to work for arbitrary meshes of human skulls, i.e. having artifacts like incomplete geometry or being non-watertight. Since the input data has no common orientation, we first apply a rigid coarse registration followed by a refinement. Afterwards, the remaining geometric deviation is removed by non-rigidly deforming one mesh into the other. With this precise geometric mapping, arbitrary landmarks can be transferred easily. To ensure efficient computation, we use a highly optimized GPU implementation to solve arising optimization problems. We apply our method to a dataset consisting of 1200 models acquired via structured light scanning and evaluate its accuracy on a subset of these models.