[Two_hands]

Right now that's pretty much what I do. I use YuNet to get faces, crop them out and run detection. It's probably a factor of a not enough data/poor model choice.