I trained MaskRCNN for building footprint extraction on imagery of cell size 0.5 and 0.75 with 12gb GPU. The results were really good except for large buildings. Now, I am working on imagery with cell size of 0.25 and the results are not good at all.
Any suggestion please.
Also, any possible way to improve shape of large commercial buildings. I have provided training samples for large buildings, still results are not good. Even the pre-trained model from Living Atlas doesn't seem to perform that well for large commercial buildings.