Targets for multi-task semantic segmentation

I appeal to your experience, I am newbie in DL.

I like to test the semantic segmentation performance into a multitask, in which, a task specializes on a single label. In this scenario, each task has its own loss, so each task requires its own targets.

My main question is how to built the targets for each task ? I am working on CamVid dataset (on this, a target annotated image has 12 clases/labels ), Should I post-process each annotated image (if yes, how it would be the class ‘0’ zero, for example)?

All suggestions are welcome.