Objective evaluation of segmentation masks in video sequences

In order to avoid the charge of visually evaluating the quality of segmentation masks, the present paper introduces a procedure for evaluating this quality by means of algorithmically computed figures of merit. Assuming the existence of a perfect (reference) mask, generated manually or with a reliable procedure over a test set, these figures of merit take into account visually desirable properties of a segmentation mask in order to provide the user with outputs that best qualify the spatial and temporal accuracy of the segmentation masks. For the sake of easy interpretation, results are presented on a PSNR-like logarithmic scale.