How to separate predictions map to create id2label.json

[seand0101]

Then do I only need the mask for training? Every guide I found use id2label though, I understand that this is semantic segmentation so not only it is segmented but its also labelled as something