Comparing Dissimilarity Measures

Abstract

In Chapter 1 we surveyed distribution-based dissimilarity measures and discussed their properties, and in Chapter 2 we presented the EMD, which we hypothesized to be a good choice for image retrieval applications. In Chapters 3 and 4 we showed how these measures can be instantiated to compare color and texture distributions In this chapter we compare the results of the different dissimilarity measures when used for color- and texture-based image retrieval. The main difficulty in such a comparison is establishing ground truth that will help to determine if a returned image is relevant or not. To this end, we create databases where the ground truth is known, and use them to conduct experiments that evaluate the performance of the dissimilarity measures.

Doubt, the essential preliminary of all improvement and discovery, must accompany the stages of man’s onward progress. The faculty of doubting and questioning, without which those of comparison and judgment would be useless, is itself a divine prerogative of the reason.

—Albert Pike, 1809–1891