The performance difference between a model's ability to understand cropped regions versus full images.