it seems like you could use facial recognition and face tracking to identify more data samples from video. if you recognize 1 sample go back and forth in time and mark the face closest in space as another sample. at 30 fps you would have 1799 new samples if you just find a single match in one frame for instances.