Shape Retrieval with Qualitative Relations: The Influence of Part-Order and Approximation Precision on Retrieval Performance and Computational Effort

Abstract

Manifold approaches exist in the field of similarity-based shape retrieval. Although many of them achieve good results in reference tests, there has been less focus on systematically examining the factors influencing both retrieval performance and computational effort. Such an investigation, however, is important for the structured development and improvement of shape descriptors. This paper contributes a thorough investigation of the influence of the shape part-order and approximation precision. Firstly, two shape descriptors based on qualitative spatial relations are introduced and evaluated. These descriptors are particularly suited for the intended investigation because their only distinction is that one of them preserves the part-order, the other abandons it. Secondly, the recall and precision values are related to the degree of approximation in three-dimensional recall-precision-approximation diagrams. This helps choose an appropriate approximation precision. Finally, it turns out that remarkable retrieval results can be achieved even if only qualitative position information is considered.