evaluate MaskRCNN on semantic segmentation task

[andreaceruti]

I would like to transform the JSON file produced by evaluation, that I use to evaluate both instance segmentation and object detection tasks into another format that I could use for evaluating also on semantic segmentation task.
I mean I have the ground truth and the detection masks (polygons/RLEs) for each instance, and so I would like to receive hints on how I could transform these segmentation into one binary image for each class and then evaluate the binary image of the detection against the ground truth.