Multi-Scale Criteria for the Evaluation of Image Segmentation Algorithms

This paper deals with evaluation of image segmentation methods. We start with a state-of-the art of the evaluation criteria, involving a reference segmentation or not. Based on an analysis of the main existing criteria, we propose new criteria, when no ground-truth is available. These criteria, based on an energetic formalism, take into account both the complexity of the segmented image, through the boundary length and the goodness-offit of an underlying model with the initial data. The main interest is not to fix the level of detail expected by the evaluation criterion but to leave it to the user’s choice, according to his purpose. These evaluation criteria are thus multi-scale criteria. Various forms of energy formulation are experimentally compared. Then after having chosen a particular energy, the performances of this new criterion is compared to the main existing criteria.