Uncommon mask input

[june94]

1. As I understood it right, input mask (for training) should have labels/classes as pixels, for example, 0-6 pixel range if i have 6 classes + background.
2. Occasionally, my custom dataset has masks labeled with ade20k palette. But without any specifications about it (whole pipeline is rather simple), training doesn't crash and also labeling seems to be right.

Questions:
Am I right about 1)? Is 2) behaviour normal or it is just a coincidence? Where could I find the method where gt_semantic_seg pixels are converted to class labels if there any?
(inspection of LoadAnnotations(), forward, loss calculations and hooks does nothing in my case)

Thank you in advance!

[MengzhangLI]

(1) In most cases, Yes. If your label 0 in annotation is background but meta info is not, like ADE20K. You should use reduce_zero_label, here is doc.

(2) Check out your customized dataset class in ./mmseg/datasets/xxx.py, you should define class and palette information like this. If you want to debug, you could set breakpoint in custom.py to see what's going on.

Best,

[june94]

Thank you!
I've already debugged this part but unfortunately could not find answer for my question.
I also found that PALETTE is only used for output/plot/show result, but not for mapping "uncommon" masks to labels/classes 0-N range (for example, in LoadAnnotations or loss calculations).
Could you please show the part where this mapping happend?