The present disclosure relates to processing images representing a real scene, the images comprising at least two images obtained from at least two sources of images, a sub-image of each of the at least two images representing an overlapping area of the real scene. A role map comprising a plurality of parts, all the points of each part of the role map being associated with a single identifier of a source of images, is generated during a training phase. To process the images, parts of the role map are selected as a function of an identifier of a considered source of images, and sub-images of an image obtained from the considered source of images are processed, the processed sub-images corresponding to the selected parts of the role map.