Evaluation quantitative de méthodes d'extraction de rues : choix d'une référence de comparaison

Abstract : In the context of an automatic street extraction from remotely sensed images, we are concerned about defining some quantitative criteria for evaluating the performance of algorithms. The aim is to obtain a measure of the quality of the results delivered by any method. This purpose needs the definition of a reference in order to compare it with extracted objects. In this paper, existing references in matter of pattern recognition from geographical scenes will be reviewed. Then a methodology for extracting a relative reference is defined when the only information available is the remotely sensed image. Related quantitative criteria are also given. The complete methodology is finally applied on a high resolution image from an urban scene.