This paper was submitted in the AI channel on forecaster. I find it fascinating, as it has a potential to learn from video, and that opens up the possibility to learn from an endless source of good data. You can imagine learning from video from the world, and making a model of it.