Semantic binary segmentation of the aerial images on Inria dataset. I obtained the best accuracy on the test set with the Resnetxt50-Unet model. Here are the score specifications (as evaluated by the competition launcher):
IoU Acc.
bellingham | 0.705244658628 | 0.967294883333
bloomington | 0.661219544124 | 0.964045216667
innsbruck | 0.705316948749 | 0.960757862222
sfo | 0.741421801289 | 0.907259936667
tyrol-e | 0.75462637417 | 0.975925242222
Overall | 0.721548268122 | 0.955056628222