We cannot do the convolution directly on thousands of regions on the raw image to do the object detection. It takes too much running...