1 | initial version |
HOG has, according to the Dallal and Triggs paper that was followed for the HOG interface, a window size of 64x128pixels. As far as I know this is the model size of the OpenCV implementation, which means that smaller objects cannot be detected. Reason for this is that the image pyramid is generated by downsampling the original image and thus leading to multi scale detections (from model size to larger persons). If you want to be able to detect smaller persons you need to manually upscale your image. However keep in mind that upscaling to much will introduce scaling artefacts that will definitely influence the models detection performance.
I have applications where upscaling it to double the original size worked, but than the smallest objects did not contain enough usefull pixel information anymore to be detected by the trained model.
Keep in mind that detections found on the upscaled version need to be manually resized the the original window size by the user, since the image pyramid doesn't contain upscaling in OpenCV.