Evaluation metric for image understanding

摘要

We propose in this paper a new evaluation metric that enables to quantify the quality of an image interpretation result. This metric takes into account the a priori knowledge used by the interpretation algorithm and the ground truth associated with the original image. We combine two metrics that evaluate the localization and recognition results of each detected object. We show that the proposed metric fulfills some theoretical properties and has a correct behavior face to empirical experiments on an image benchmark database. We think that this metric could be a reliable reference for image and video understanding competitions.