Can I train using PNG semantic masks as labels instead of COCO?

[fschvart]

I tried searching and couldn't an answer, is it possible to train Detectron2-based models with semantic mask annotations?
I have complex masks that I wasn't able to convert to COCO polygon annotations, any help in training with .PNG labels would be greatly appreciated

[SouLeo]

Have you found any further information on this? I'm in a similar situation.