Measuring the quality of figure/ground segmentations

Figure/Ground segmentation is of great interest within the analysis of video streams. We propose a new, continuous evaluation measure for figure/ground segmentation algorithms which allows for assessing the quality of a segmentation. The evaluation approach is based on set similarities and logic-motivated considerations. Results obtained with the new measure are shown for several well-known algorithms. Compared to existing detection rate/false alarm considerations, the proposed measure reveals advantages and shortcommings of different algorithms much better and can therefore be used as a selection criterion.