Not trying to be the proverbial "wet blanket", but the concept is not really new. I believe we have even heard this concept expressed a few times here, and some of the issues brought up were:
Distance required, especially for larger screens. I.e. how well does it work when dealing with an incomplete trapezoid?
The need to frame the image, in order to provide markers. Mame provides this ability, and some low complexity images may not need it. But the possible issues are that light will need to be controlled around the screen, some calibration will still be necessary to isolate the image from the background and tell the application about the game being played (resolution, aspect ratio, etc) and the frame could throw off the display for those looking for 1:1 display accuracy. Some of this could be automated by linking into a database, or Mame itself, but given the number of possible display options, I'd be surprised if calibration wasn't an eventual necessity.
The other concern is that the demos seem to show a pretty strict 90 degree orientation of the gun. The effects of turning the gun require some more complicated math, and if it's less than perfect, so will be your positioning.
Finally, the way this is done really isn't so different from using a "light bar" as a reference. It does provide more markers, but the possible downside is, again, the need to see the entire screen (if it's incapable of handling partial references.)
It's definitely an interesting proof-of-concept, but I still see a number of hurdles which need to be overcome. The extent to which these will, or can be, is the risk in buying into the idea.