Measuring the Performance of Shape Similarity Retrieval Methods

When datamining large databases with images or drawings of objects, there is a need to search for objects with a shape similarity. According to recent overview papers the issue of comparing different shape similarity retrieval methods and systems has largely been neglected in the research community due to the subjective character of such comparisons. Indeed, the evaluation of a shape similarity retrieval system can only be made with reference to a particular application, in our case retrieval of drawings of aluminum sections. In this paper we propose five different new performance measures for shape similarity retrieval. We have applied the methods on six different feature sets including skeleton, moment, Fourier, and fuzzy/symmetry features. The results clearly show that shape similarity retrieval is both application and representation dependent and can be evaluated by a number of independent methods.