Researches On Shape Classification Algorithm Of Incomplete Non-rigid 3D Models

With the development of 3D scan skills,3D models are growing rapidly on the Internet in recent ten years. How to classify and integrate the existing 3D models correctly and effectively becomes a research hotspot, and receives more and more attention by researchers. However, most of the existing shape classification approaches are designed for rigid models, and not suitable for non-rigid models with plentiful pose changes. Moreover, it is more challenging for the shape analysis of incomplete 3D models because of the missing data.The shape features based on diffusion geometry are analyzed in this thesis and are then applied to incomplete non-rigid 3D shape classification. A classification algorithm based on persistent heat signature (PHS) and a classification algorithm based on sparse representation are then proposed. The main research work of this thesis is as follows:(1) By comparison and analysis of the existing shape features for non-rigid 3D modes, this thesis uses shape features based on diffusion geometry to describe the vertex characteristics of 3D models. According to the comparison of the heat kernel signature (HKS)and wave kernel signature (WKS) of the incomplete models, we find the impact on HKS caused by the missing data is less. Therefore, HKS is more suitable for the classification of incomplete models.(2) A shape classification algorithm based on PHS is implemented. Firstly, we calculate the Persistence of HKS at a certain scale to select the maximum point of local HKS as feature points. By comparing the multi-scale HKS of the feature points, the similarity between the models is measured. The final classification results are then obtained by combining with the nearest neighbor classification algorithm. Experiments show that this method extends the scope of HKS and improves the classification accuracy for incomplete models to some extent. However, the extraction of the feature points will be affected when the missing part of the model is more, thereby affecting the classification accuracy.(3) A shape classification algorithm based on sparse representation is proposed. Firstly, the largest connected component is obtained. The HKS descriptors of the largest connected component are then calculated and those descriptors of the boundary vertices and their 1-ring neighbors are excluded. For shape classifications, the dictionary is learned for each class based on the sparse representation theory. For a test model, each dictionary is utilized to sparsely represent its descriptor set, and the most appropriate dictionary is then determined by the representation error, the model is finally classified according to this dictionary. This method reduces the impact of unconnected parts and boundary regions when computing the HKS and extends the scope of HKS applied for non-rigid models. Experimental results show the proposed method has good classification accuracy for the incomplete non-rigid 3D models.