Maybe that's actually feasible, in the sense that you don't have to train a complete new model but can spit out a personalized model based on a calibration sequence of ten images.