What Could We Learn from Many Datasets in Remote Sensing Roof Semantic Segmentation?

Deep learning models trained for roof semantic segmentation from remote sensing images may suffer from low per-formance when applied in new cities as roof appearance may be different depending on the geographic areas. However, we claim in this paper that models learned on many different cities have better performances on a new city than a model learned on few data from this targeted city. This result may interest the community for industrial purpose: annotating few tiles of an area for simple supervised training is worthless in our experiments.