Improving Descriptors for 3D Shape Matching

Abstract

This thesis presents novel descriptors for both rigid and non-rigid 3D shape matching with dentistry related applications. For rigid shape matching, a descriptor named Improved Spin Image (ISI) is proposed which is an improving version of the famous Spin Image. By encoding the signed angles, ISI becomes much more informative and robust to noise. Based on the rigid shape descriptor and a learning scheme, an efficient 3D dental identification method is introduced with both high accuracy and efficiency are achieved. For non-rigid shape matching, Symmetry Robust Descriptor (SRD) is proposed which is able to differentiate the intrinsic symmetric feature points on 3D geometric shapes. Symmetry flipping problem hinders the performance of most existing non-rigid shape matching algorithms. The proposed descriptor solves the problem through incorporating the orientation information. Based on the idea of SRD, a novel asymmetry evaluation metric is defined which can effectively detect and evaluate the mandibular asymmetry.