Comparative study of localization metrics for the evaluation of image interpretation systems

Image interpretation is particularly important in many real applications (video monitoring, biometrics, etc.). Due to the proliferation of image interpretation systems in the literature, their evaluation still remains a crucial stake. Among all the tasks in this field, the quality of object localization is often evaluated through an evaluation metric. We propose to review these techniques and study their reliability. We first propose a generic definition of a localization algorithm. Then, different state of the art techniques to evaluate image interpretation results are detailed. Secondly, we focus on metrics that enable us to evaluate localization results. We propose a general methodology to analyze the behavior of an evaluation metric, considered here as a black box (its definition is not even supposed to be known). We define the properties that these metrics should fulfill. We then perform a comparative study of 33 localization metrics from the state of the art. Experimental results conducted on a large and significant image database permit us to determine metrics that should be used in the future for the evaluation of object localization results.