3D model retrieval using distance filter

Abstract

3D model retrieval is a very important research topic both in computer vision and computer graphics field. Most previous work is focusing on defining robust shape descriptor in order to get better retrieving accuracy. Unlike these work, we propose distance filter to improve the retrieving accuracy when using the existing shape descriptors. The main idea of distance filter is to eliminate distance noise in the process of evaluating similarity between models. As a result, the computed similarity between 3D shapes will be more robust under different variation. The implementation of distance filter can be efficiently done by using filter theory, which is widely used in signal processing. Moreover, the definition of distance filter is independent of any shape descriptors. Therefore, it can be used in conjunction with any 3D shape descriptors. To verify the effect of distance filter, we carry an experimental analysis by using two public 3D testing databases and two different kinds of shape descriptors. All results show distance filter can greatly improve retrieving accuracy whether single or hybrid shape descriptors.