A Statistical Method for Selecting Pattern Descriptors of Textured 3D Models

This paper describes a similarity retrieval technique for 3D models with solid textures. Three dimensionally extended fractal based descriptors have been extracted from each 3D model in a database. Various sizes of fractal filters were used for extracting descriptors of the 3D models, and the descriptors were compared as indices of the 3D models. Since the filter size used in the system affects retrieval performances, it is important to determine the correct size. In our experiment, portions of the database were marked as a learning data set, and the relationship between the descriptors and filter sizes was analyzed by multiple regression analysis using the learning data set. Our experimental system reflects the analysis results for choosing optimal filter size at each query for maximizing similarity retrieval performance. This preliminary experimental retrieval system has been implemented for searching for similar 3D models with solid texture patterns. The experimental results of our approach showed retrieval performance improvements in terms of recall-precision