Because HOG descriptor was developed for human
detection the authors propose resizing all images to 64 x 128
pixels and then computing the descriptor. For other types of
objects different image sizes are required (e.g. for cars: 104 x
56 pixels; for bicycles: 120 x 80) [16].