They propose to use different cases and deep network architectures to leverage OpenStreetMap data for semantic labeling of aerial and satellite images. Especially, the fusion-based architectures and coarse-to-fine segmentation to include the OpenStreetMap layer into convolutional networks. They used on two public datasets: ISPRS Potsdam and DFC2017. In the end, they show that OpenStreetMap data can efficiently be integrated into the vision-based deep learning models and that it significantly improves both the accuracy performance and the convergence speed of the networks.
Binary Vs. Signed Distance Transform (SDT)
- Binary: For each raster, they have an associated channel in the tensor which is a binary map denoting the presence of the raster class in the specified pixel.
- SDT: For each raster associated channel corresponding to the distance transform d, with d > 0 if the pixel is inside the class and d < 0 if not.
- For segmentation, we can only use roads, buildings and vegetation landuse OSM class.
- They use more recent OSM data than the ISPRS dataset (some buildings are on OSM and not on Potsdam).
- Using different channel with the OSM data can quickly bust all the memory.