Reeb Chart Unfolding Based 3D Shape Signatures

Abstract

This paper presents a novel surface parameterization based technique that addresses the pose insensitive shape signature problem for surface models of arbitrary genus. It is based on the key idea that two surface models are similar if the canonical mappings of their sub-parts introduce similar distortions. First, a Reeb graph of the shape is computed so as to segment it into charts of controlled topology, denoted as Reeb charts, that have either disk or annulus topology. Next, we define for each Reeb chart a straightforward mapping to the canonical planar domain. Then, we compute a stretching signature of the canonical mapping based on an area distortion evaluation. Finally, the input shape is represented by the set of the stretching signatures. An application to pose-insensitive shape similarity is proposed by comparing the signatures of the different Reeb charts. Promising experimental results are presented and compared to state-of-the-art techniques. The gain provided by this new signature as well as its interest for partial shape similarity are demonstrated.