Supporting Effective and Efficient Three-dimensional Shape Retrival

In this paper, we discuss the search effectiveness and efficiency of content-based 3D shape search. The search effectiveness is characterized by the precision and recall of results; and the search efficiency is evaluated by the ratio of the number of the visited nodes in a search operation to the number of the nodes of an index. A real database of 113 3D shapes and synthetic databases created by random number generator are employed for the investigation. Four feature vectors are extracted to represent 3D shapes. The descending order of search effectiveness of using feature vectors are: principle moments, moment invariants, geometric parameters, and eigenvalues. We evaluated the performance of an R-tree based index to improve the search efficiency. The results show that R-tree index significantly improves the efficiency of both real and synthetic databases. The search efficiency decreases when the dimensionality of data record and the capacity of database node increase. However, based on the experiments of synthetic database, the efficiency is stable when the size of a database is relative large. Thus, it is scalable to the size of database. A multi-step refinement is proposed to improve search effectiveness. Based on the results of twenty six experiments with real database, the effectiveness of using multi-step refinement is 51% higher than that of one-shot search using principle moment as feature vector, although the difference becomes smaller when the number of retrieved shapes becomes larger.