Hi!
What channels do you use during training?
Do you use only false color channels?
Skimmed through dataset, it seems that segmentation masks are annotated from false color images. Didn’t see any additional useful signal beyond false color channels. But might be wrong. What are your thought?