New criteria for evaluating image segmentation results

This paper is dedicated to evaluation criteria for image segmentation when there is no ground truth. New criteria based on a formulation of the image segmentation as a piecewise modeling problem are proposed. These criteria take into account both the complexity of the segmented image, through the total boundary length and the goodness-of-fit through a distance between model and initial image. They allow to rank segmentation results or human segmentations, according to an expected level of detail. These new evaluation criteria are compared to the most used evaluation criteria both on results of segmentation algorithms and on manual segmentation achieved by humans.