The geographic data in OpenStreetMap (OSM), described through geo- metry and semantics, is an essential source of information for relief maps production for the visually impaired. However, at the crossroads scale, the geometric details re- quired for modeling some components are not available. The use of aerial imagery is a complementary source of information, but involves advanced image processing. In this article, we propose an approach based on deep learning (conditional Generative Ad- versarial Networks), by enhancing the information of orthophotographs with semantic and geometric data from OSM. In order to measure the influence of this enrichment, we present the results of two series of learning, with and without enrichment.