Anatomic Segmentation of Statistical Shape Models

Shape segmentation is one of the fundamental tools in shape processing and provides the starting point for building part based shape models. A multitude of methods for segmenting static and dynamic shapes has been developed over the last decades based on various intrinsic and extrinsic geometric features. Interestingly, so far no method considering information from an available statistical shape model seems to exist. In this work a segmentation based on statistical covariance analysis is derived. Its results show meaningful parts in the sense that all points in a part share similar (co-)variability, i.e. behave similar according to the model, while different parts show distinct variation patterns. We show initial results in two application domains where this is a desired property: In morphometrics independent modules are sought relating to an underlying independent evolutionary development; in motion analysis correlated movement is relevant for motion understanding and compression.