Research On Correspondence And Segmentation Of 3D Shapes

In computer graphics, 3d scanning and modeling technology is developing rapidly. As a result, the scanning result appears to be a 3D dynamic model instead of a static model. In consequence, how to understand and analyze 3D dynamic model remains a hot research topic. This thesis focuses on the correspondence and segmentation of 3D dynamic model. The main research work includes:1. Existing approaches will cause inconsistent segmentations for 3D animated meshes. In view of this, we propose an algorithm of segmenting 3D animated meshes based on 3D data correspondence. As a result, it can obtain highly consistent segmentations with user’s marks. This algorithm only requires user to annotate the first frame. According to the corresponding relationship, the user’s marks can be easily transformed to other frames. Finally, based on directional angle, the algorithm refines cutting boundary to get expected segmenting results. Experimental results show that the algorithm can effectively segment artificial 3D animation and multi-view reconstruction animation. Furthermore, the proposed algorithm can effectively improve the segmentation quality over the existing algorithm.2. This paper addressed to the problem that poses changing would lead to an inconsistent segmentation. It proposes a new algorithm driven by semantic ontology. Firstly, it builds a segmenting ontology used for segmenting process. The defined ontology builds a mapping between semantic label and local geometrical features. Secondly, the hierarchical segmentations of dynamic shapes will be consistent with the help of semantic label. Finally, the segmentation boundary is extracted by poisson equation. This method makes the boundary smooth and consistent. In experiment, the segmenting results of different dynamic models show that theproposed algorithm is very stable. Furthermore, it can get better results than existing approaches. 3. This paper proposed 3D model progressive correspondence with saliency constraint. Different from existing algorithms, the proposed method recognizes and matches salient point prior to corresponding process. As a result, it can resolve the mismatching problem deduced by isometric of feature points. Then, sparse correspondences can be obtained with the constraint of salient point. Finally, dense correspondences are obtained using Kuhn-Munkres. In experiment, the algorithm’s robustness is verified. Moreover, it can greatly improve corresponding accuracy over the existing algorithm.