Access Type

Open Access Dissertation

Date of Award

January 2011

Degree Type


Degree Name



Computer Science

First Advisor



Shape analysis is a fundamental aspect of many problems in computer graphics and computer vision, including shape matching, shape registration, object recognition and classification. Since the SIFT achieves excellent matching results in 2D image domain, it inspires us to convert the 3D shape analysis to 2D image analysis using geometric maps. However, the major disadvantage of geometric maps is that it introduces inevitable, large distortions when mapping large, complex and topologically complicated surfaces to a canonical domain. It is demanded for the researchers to construct the scale space directly on the 3D shape.

To address these research issues, in this dissertation, in order to find the multiscale processing for the 3D shape, we start with shape vector image diffusion framework using the geometric mapping. Subsequently, we investigate the shape spectrum field by introducing the implementation and application of Laplacian shape spectrum. In order to construct the scale space on 3D shape directly, we present a novel idea to solve the diffusion equation using the manifold harmonics in the spectral point of view. Not only confined on the mesh, by using the point-based manifold harmonics, we rigorously derive our solution from the diffusion equation which is the essential of the scale space processing on the manifold. Built upon the point-based manifold harmonics transform, we generalize the diffusion function directly on the point clouds to create the scale space. In virtue of the multiscale structure from the scale space, we can detect the feature points and construct the descriptor based on the local neighborhood. As a result, multiscale shape analysis directly on the 3D shape can be achieved.