Empirical evaluation of dissimilarity measures for 3D object retrieval with application to multi-feature retrieval

A common approach for implementing content-based multimedia retrieval tasks resorts to extracting high-dimensional feature vectors from the multimedia objects. In combination with an appropriate dissimilarity function, such as the well-known L_p functions or statistical measures like χ², one can rank objects by dissimilarity with respect to a query. For many multimedia retrieval problems, a large number of feature extraction methods have been proposed and experimentally evaluated for their effectiveness. Much less work has been done to systematically study the impact of the choice of dissimilarity function on the retrieval effectiveness. Inspired by previous work which compared dissimilarity functions for image retrieval, we provide an extensive comparison of dissimilarity measures for 3D object retrieval. Our study is based on an encompassing set of feature extractors, dissimilarity measures and benchmark data sets. We identify the best performing dissimilarity measures and in turn identify dependencies between well-performing dissimilarity measures and types of 3D features. Based on these findings, we show that the effectiveness of 3D retrieval can be improved by a feature-dependent measure choice. In addition, we apply different normalization schemes to the dissimilarity distributions in order to show improved retrieval effectiveness for late fusion of multi-feature combination. Finally, we present preliminary findings on the correlation of rankings for dissimilarity measures, which could be exploited for further improvement of retrieval effectiveness for single features as well as combinations.