EDIT: here's a particularly blatant example, around 0:17: http://i.imgur.com/8mSv8oC.jpg

The user is swiping from right to left, causing their hand to be highly motion-blurred. But the rendered UI elements smoothly fade in behind the hand.

There are foreground segmentation algorithms that can do this kind of thing, assuming you have some information about which regions of the image are foreground/background (say, from a depth sensor). But those algorithms typically take on the order of seconds or minutes to process a single frame. For this video to be real, the total latency -- sensing, matte generation, and rendering -- would have to be in the single-digit milliseconds. It's barely on the edge of plausibility, but I'm very very skeptical.