Researches On3D Shape Segmentation Based On Diffusion Geometry

In recent years,3D models have been widely used in many industries, such as entertainment, medical image, movie, animation, computer aided design and3D printing. Moreover, the demand for3D models is increasing dramatically. For dealing with a large number of3D models automatically, researchers attach more attention to the research of3D shape analysis.3D shape segmentation is one of important and fundamental research problems in3D shape analysis. However, most existing shape segmentation approaches are designed for rigid objects, and not suitable for non-rigid objects with plentiful shape changes which are common in life. Non-rigid shape segmentation is more challenging than rigid shape segmentation, and therefore segmentation algorithms are expected to maintain consistency under various shape changes of non-rigid objects.This paper firstly analyzes shape features based on diffusion geometry, and then heat kernel and wave kernel signature are applied in non-rigid3D shape segmentation, finally a critical feature point detection algorithm and a shape segmentation approach based on hybrid features are proposed. The main research work of this paper is as follow:(1) Existing3D shape segmentation approaches are analyzed and summarized, especially non-rigid3D shape segmentation approaches based on diffusion geometry. Furthermore, the advantages and characteristics of shape descriptors based on diffusion geometry are also investigated to meet the need of non-rigid3D shape segmentation.(2) A Heat-Mapping algorithm for non-rigid shape segmentation is implemented. In this algorithm, heat mean signature (HMS) based on heat diffusion are firstly utilized to detect heat centers, and then these heat centers are taken as initial cluster centers to drive vertex clustering, segmentation results are finally achieved.(3) In order to solve the instability of heat center detection in Heat-Mapping algorithm, this paper presents a critical point detection algorithm based on wave kernel signature (WKS), which takes the vertices with local extremes of WKS on mesh surface as critical points.(4) For purpose of being adaptive to various postures of non-rigid objects, a non-rigid3D shape segmentation approach based on diffusion geometry is proposed. The critical points detected by the above detection algorithm are taken as initial cluster centers, and K-means clustering approach is performed to get segmentation results. The extensive experiments show that the proposed approach can get consistent segmentation results for a non-rigid3D object under various postures, and it is robust to noises and holes as well.