I tried implementing this with face detection-based head tracking after that demo (or maybe before; I can't remember). I got it working but the effect was very underwhelming. It looks great in that video, but it kind of sucks in real life.

I think the problem is in real life you have an enormous number of other visual cues that tell you that you're not really seeing something 3D - focus, stereoscopy (not for me though sadly), the fact that you know you're looking at the screen, inevitable lag from cameras, etc.

I can't view the videos because of their stupid cookie screen, but I wouldn't be too excited about this. The camera lag especially probably impossible to solve.