Measuring empirical discrepancy in image segmentation results

A methodology for comparison of boundary and segmentation images based on Precision-Recall graphs is presented in this study. The proposed methodology compares the location of edge pixels between an image under test and an ideal reference, in order to obtain a precise normalised similarity measure. This approach also deals with the case when multiple references are available using a merging procedure. Small displacement errors in edge pixel location are handled using a tolerance radius, which introduces the problem of multiple matching between test and reference edge pixels. This problem is addressed as a bipartite graph, solved by using the Hopcroft-Karp algorithm to obtain the maximum number of unique matchings. Experiments have been carried out in order to determine the performance of this evaluation approach.