NEW APPROACH TO THE CONSISTENT LABELING PROBLEM.

The problem of defining a goodness measure for object labeling problems is addressed. The optimal probabilistic measure is derived but shown to be impractical for realistic problems. An alternative consistency measure is suggested, which is based on the accuracy with which each object can estimate its label. This measure is shown to be a generalization of the crisp idea of consistency and can be computed using statistical information about the problem. Results are shown on a triangle labeling example and on gray-level picture segmentation and border detection.