Probability Density Based 3D Model Retrieval

With the rapid development of 3D data acquiring technology, shape modeling technology and computer hardware technology, 3D models are becoming more and more useful in different fields. At the same time, the designing idea has gradually transformed from the question of how to build a new 3D model to the question of how to search for a similar 3D model which already exists and then make full use of the 3D model. Based on this background, more and more researchers are becoming interested in the technology of 3D model retrieval. Each of the already existed 3D model retrieval technologies, such as text based retrieval technology, or content based retrieval technology, which includes directly feature based retrieval method and shape descriptor based retrieval method, has its own advantage. However, these 3D model retrieval technologies share a problem of high computational complexity and poor retrieval performance.Based on the starting point that similar shapes induce similar feature distributions, and feature's probability density could describe the feature distributions, this paper uses the probability density of features for 3D model retrieval, which named probability density based basic retrieval method. This method uses the Gaussian kernel density of the 3D model's surface features as the shape descriptor and calculates the Minkowski distance between two 3D models'shape descriptors. This method doesn't need to split the 3D model, either construct topological diagrams, therefore, It has the advantage of easy implementation.Based on the analysis of the disadvantage of basic retrieval method, this paper proposes an improved retrieval method, which named the probability density based hierarchical retrieval method. The basic idea of hierarchical retrieval method is to use relatively lower level features to retrieve the 3D model firstly, and then, use relatively higher level features for further retrieval. Due to the reason that not all the 3D models in database have to calculate the relatively higher level features'probability density, thus the computational complexity could be reduced, and the retrieval performance is better.By using a subset of the Princeton Shape Benchmark as the testing database, the experiments'results show that the technology of probability density based 3D model retrieval has a better retrieval performance over the traditional retrieval technologies, especially when the recall option is low. At the mean time, compared with the probability density based basic retrieval method, the probability density based hierarchical retrieval method is better at both precision and computational complexity.To conclude, probability density based 3D model retrieval technology, including the basic retrieval method and the hierarchical retrieval method, is a supplementary for the traditional 3D model retrieval technologies, and can make its contribution to the research of 3D model retrieval technologies with its good retrieval performance.