This is really cool but I wish the code were available. It seems that increasingly it is becoming the norm not to release the training code. The paper seems thorough, though. I can think of some great applications of this and I hope that frameworks like ARKit incorporate it.
I can't wait to find out in 5 years that this approach ends up being used for something nefarious like automatic threat detection and elimination to allow some Boston Dynamics robot to autonomously shoot people.
This wouldn't have been my first thought 5 years ago but that's how things are going.
