just putting this here so that we dont forget, this is early 2025 stuff
I think the conclusion of previous conversations was that we compute a tiled dataset (orthophoto 1024x1024x3 + label 1024x1024x1) on either the storage or compute server and than offer to download that. Multiple options:
dummy_datset for testing
orig full_resolution dataset based on metadata based filter
all downsampled to 10cm patches
The code in torchgeo then reads from this downloaded repo.
just putting this here so that we dont forget, this is early 2025 stuff
I think the conclusion of previous conversations was that we compute a tiled dataset (orthophoto 1024x1024x3 + label 1024x1024x1) on either the storage or compute server and than offer to download that. Multiple options:
The code in torchgeo then reads from this downloaded repo.