Have you tried making the output a palette with potentially multiple color choices? I'm trying to think of a way of representing "it could be any of these colors, choose one." The problem is training it will be tough without being able to semantically segment objects that are the same but colored differently a priori.
I'd imagine the output would also have to be ordered, but this is easy -- just use the EM spectrum as an imposed ordering for the palette, and say, cap it at size 3 max with potential null hues for things that really only have one (or less than max) color(s).
I'd imagine the output would also have to be ordered, but this is easy -- just use the EM spectrum as an imposed ordering for the palette, and say, cap it at size 3 max with potential null hues for things that really only have one (or less than max) color(s).
Any ideas?