Small urban areas in developing countries have very little resources, especially for keep up to date maps that serves the management of the territory by public authorities. Roads are probably one of the most rapidly, but important and expensive to maintain elements in infrastructure. Use of the most available and affordable aerial imagery and the sounded deep learning technology could help in this matter. This work presents a pipeline using an Image Translation approach instead of a traditional image segmentation architecture for the extraction of urban roads directly from high resolution drone orthomosaics is presented. The results show that applying overlapping masks, multi-scale re-training and data augmentation besides skeletonization of inferred tiles output clean vector roads.