DL in Remote Sensing: Segmenting (buildings from) SpaceNet 7 (SN7) satellite images (AWS dataset) with fastai UNet with python.

Since the images are very high-resolution images, first the training input images and the corresponding ground-truth are split into tiles and trained (for ~50 epochs) with fastai UNet (a model built using pytorch underneath), in order to prevent it to be extremely computionally expensive. The segmentation mask for the held-out images are computed first by predicting those masks for the tiles and then combining the tile masks.

свернуть