A Research On Non-rigid3D Model Matching

Non-rigid shapes exists in every place where we live, such as from microscopiccells to tissues and limbs of human body. Thus the analysis of non-rigid shapes hasbecome important in modern applications.3D model matching is the main componentsof the non-rigid shape analysis, and it is also very important in computer vision. Weneed to be related shapes together before performing any similarity analysis or targetdetected and recognition. In this paper we focus on3D model matching. A lot of methodhas been used in model matching. There are two main branch of3D model matching.Global matching method based on spatial representation, and local matching methodbased on descriptor. In this paper we improve the spatial representation method, andcombine a new distance metric method.In3D model matching, there are two main research directions. Embed the shapesinto new space. Retain effective shape information in the new model space, and to finda direct matching between the shapes in the new approximation space. These methodsmay introduce a new approximation error, and such error is often generated on the typeof matching results. Others adopt descriptor, these methods marked each vertex on thewhole shape. Each point on the shape carried different shape information. Then thevertices can be distinguished. Descriptor is hard to be convergence, and differencesbetween vertices are required high. When the differences between vertices are small, orgeometry structure are similar. This type of algorithm will produce errors, and even thematch fails. This paper analyzes these shortcomings，and improve the3D modelmatching algorithm.First, we propose a new sampling algorithm for3D model matching, which takeinto account not only the spatial distribution of base vertex, but also extract the saliencyof feature points, that is, extract feature points in different shapes consistent. In order toverify the performance of our method, we compare with some existing algorithms.Second, we use the spectral domain representation in our initialization procedure,and compute the cost affinity matrix in it. Then in the Euclidean space, we apply thediffusion distance, which can avoid topology noise, in our refinement process. Wecompare our method with generalized multidimensional scaling method. Our algorithmhas more performance and accuracy.