For this particular problem, how does the proposed solution compare to just classifying based on image size?