It would be interesting to pair this with something like MaskRCNN or something which does semantic masking, and then be able to choose the compression levels of different masks.