Improved Method For Shape Correspondence Based On Heat Kernel Signature

With the improvement of 3D scanning technology,3d data processing and shape analysis have attracted extensive attention,and shape correspondence has become a hot issue in research.Shape correspondence has applications in shape deformation,texture mapping,statistical shape modeling,etc.Shape correspondence refers to finding the point-to-point match between two shapes,mainly including rigid and non-rigid matching.Rigid matching research is quite mature,and the non-rigid matching is a difficult problem.Even the approximate isometric matching can cause some problems.For intrinsic symmetric shapes,symmetry flips is easy to occur in matching.In this paper,we mainly to solve the symmetric flips in shape correspondence.Based on the Novel Method for 3D Symmetric Shape Matching Based on Shape Embedding,this paper proposes an improved method.Experiments show that the selection of feature points,the classification of sampling points and the initial matching of the original method all affect the final matching results.This paper mainly improves these three aspects.In this paper,we use the farthest point sampling on the basis of the heat kernel signature sampling,and arrange the feature points in descending order according to the average geodesic distance to fuse the redundant points to obtain the stable and representative feature points;we combine the heat kernel signature at different times and the average geodesic distance to increase the classification accuracy of feature points;in order to improve the initial matching accuracy,we use the one point isometric matching algorithm to fix asymmetrical point matching,and then propagate the symmetric point matching.Finally,the original matching method is adjusted to solve the symmetric flips.Several experiments is conducted with the TOSCA database to compare the state-of-art method,the results show that the proposed method effectively solves the problem of symmetric flips,greatly improves the accuracy,and can be applied to a wider range of models.