Given a set of image processing requirements, how would you optimize the memory footprint in a Qualcomm vision pipeline?
