A template fitting, non-rigid registration algorithm applied to craniofacial depth scans

Abstract

As faster and cheaper 3D acquisition systems emerge, so does the need to find methods that are able to deal with their inherent noise and lack of structure. In the particular case of facial and craniofacial scans, the available databases present abrupt changes on resolution, significant noise and holes. Aside from defects on the regularity of the meshes, there is not a vertex-to-vertex correspondence between scans of different subjects. Correspondence has traditionally been a crucial part for building morphable models and more recently, for building structured databases for machine learning applications. In this thesis, we present a template-fitting, non-rigid registration algorithm for morphing a clean and smooth template towards facial and craniofacials depth scans. Our approach employs an iterative closest point algorithm that minimizes the distance between correspondent points at the same time as imposing constraints on the deformations to preserve the smoothness and cleanness of the original mesh. We tested our framework with two modern datasets with positive results in capturing the user identity and dealing with missing parts, noise and low resolution meshes. Meshes processed with the same template also presented good correspondence between vertices of the same indexes. Finally, our algorithm shows comparable performance to state of the art approaches at minimizing the per-vertex difference between the morphed template and the scanned faces.