Combining Multiresolution Shape Descriptors for 3D Model Retrieval

Abstract

In this paper, we propose and evaluate a systematic approach for improving performance of 3D model retrieval by combining multiple shape descriptors. We explored two approaches for generating multiple, mutually independent, shape descriptors; (1) application of a (single-resolution) shape descriptor on a set of multiresolution shape models generated from a query 3D shape model, and (2) application of multiple, heterogeneous shape descriptors on the query 3D shape model. The shape descriptors are integrated via the linear combination of the distance values they produce, using either fixed or adaptive weights. Our experiment showed that both multiresolution and heterogeneous sets of shape descriptors are effective in improving retrieval performance. For example, by using the multiresolution approach, the R-precision of the SPRH shape descriptor by Wahl, et al, improved by 8%, from 29% to 37%. A combination of three heterogeneous shape descriptors achieved the R-precision of about 42%; this figure is about 5% better than the R-precision of 38% achieved by the Light Field Descriptor by Chen, et al., which is arguably the best single shape descriptor reported to date.