Supervised Learning of Similarity Measures for Content-Based 3D Model Retrieval

Abstract

In this paper we investigate on how the choice of similarity measures affects the performance of content-based 3D model retrieval (CB3DR) algorithms. In CB3DR, shape descriptors are used to provide a numerical representation of the salient features of the data, while similarity functions capture the high level semantic concepts. In the first part of the paper, we demonstrate experimentally that the Euclidean distance is not the optimal similarity function for 3D model classification and retrieval. Then, in the second part, we propose to use a supervised learning approach for automatic selection of the optimal similarity measure that achieves the best performance. Our experiments using the Princeton Shape Benchmark (PSB) show significant improvements in the retrieval performance.