This HOG computation is traditionally performed by repeatedly stepping a window of, say, 64 pixels wide by 128 pixels high across a test image frame and computing the HOG descriptors. As the HOG calculation contains no intrinsic sense of scale and objects can occur at multiple scales within an image, the HOG calculation is stepped and repeated across each level of a scale pyramid. The scaling factor between each level in the scale pyramid is commonly between 1.05 and 1.2 and the image is repeatedly scaled down until the scaled source frame can no longer accommodate a complete HOG window. If the SVM classifier predicts the detection of an object at any scale, the corresponding bounding box is returned. The following diagram shows a typical HOG object (pedestrian) detection workflow:

This technique is more accurate but more computationally complex than Viola-Jones object detection.