Why do you limit your preprocessing with only youtube-real subset of FF++?
Namely, you indicate -d Original parameter while running crop_dlib_ff.py and crop_retina_ff.py. And you throw NotImplementedError if a subset is not indicated. So how to preprocess the whole FF++ at a single run?