@sgugger I have been messing around taking a stab at trying to fit this into the fastai framework. I’m building off of @Hadus’s solution above. Do you think the best way is to use a random seed for matching up the SegmentationTileLabelList and SegmentationTileItemList? Or if we return all segments by adding an axis and then handle downstream further? What about for the test dataset where we just want to scan across and then recombine and show the full stitched together image?
I’m fairly close on having my pipeline working by just preslicing the images and saving to disk so you load in that way initially. Then for test doing the slicing and then stitch back together outside of the fastai framework, but I think it would be really slick and help a lot of people if we try to get it integrated within.