Abstract
Recent surge in interest for 3D shape similarity search and retrieval of 3D shape models has produced many 3D shape similarity comparison methods. Yet, the performance of each shape comparison methods is not sufficient, and the performance varies depending on the kind of shape sought for. In this paper, we propose a systematic approach that combines multiple distances obtained from multiple shape features to improve overall 3D model retrieval system performance. We linearly combine normalized distances by using (1) static weights, and (2) dynamic weights that are determined using a pre-classified training database. As the shape features to be combined, we used (1) four basic shape features (D2, AAD, SPRH, and LFD), as well as (2) their respective multiresolution versions produced by using Ohbuchi and Takei's method. Experimental evaluations showed that the combination of multiresolution SPRH shape features outperforms the LFD, arguably the best performing shape feature today. Furthermore, some combinations of multiple shape features outperformed the LFD by a large margin.
