Of course, it doesn't actually look invisible unless you're standing in exactly the right place. That is why the stationary shots at the beginning of the video look so much more "invisible" than the moving shots later in the video. If you pause on those later shots, you'll see that there are big discontinuities around the boundary of the car.
Thought experiment: imagine the car as a sheet of glass. Think of yourself standing at position A and looking at a point X on the glass. You see a position A' behind the glass. Now imagine moving to a position B and looking at the same point X on the glass. You see a different position B'. The LEDs don't know if you are at position A or B, so they can't know whether to show the light from A' or B'.
Do you really have to be such an extreme kill joy? I think anyone that watched the video would see the exact same thing we both saw, yes, it isn't perfect, yes if you aren't looked straight on you won't see it, but goddamnit if it isn't an awesome effect that is going to get people talking and will generate buzz for Mercedes and overall is an absolutely awesome way to advertise that your car looks "invisible" to the world it is driven in.
They use a deceptive camera angle to make it look like it works much better than it does in reality. As thoughtful people, what are we supposed to conclude about hydrogen fuel cells?
It's still a great ad, and I enjoyed it. Some of the secondary press coverage is terrible though (PC World: "Mercedes makes invisible car, tricky to find where you parked it"; says that if the invisibility isn't perfect, it's because the LED sheet is too low resolution.)
The same trick was used in the recent Mission Impossible movie. The device, which was controlled by an iPad (ha!), used eye tracking to adjust the image to look correct from the guards pov. And of course it failed when another person entered the room. I thought it was quite well done (though the real-time adjustment aspect seems very unlikely) but I wonder if what was happening was lost on most of the audience.
I was already familiar with the concept, I remember reading about invisibility cloak prototypes[1] in the early 2000s.
This could be worked around by making the LEDs more directional and taking input from an array of directional cameras on the other side (or presumably, a single Lytro camera per LED). I don't believe there's a fundamental difficulty in amelioration the problem of parallax.
Mission Impossible: Ghost Protocol actually dealt with this tech in a semi-realistic way, and acknowledged the "only works from one viewing angle" problem. The invisible car in "Die Another day" ignored it.
Thought experiment: imagine the car as a sheet of glass. Think of yourself standing at position A and looking at a point X on the glass. You see a position A' behind the glass. Now imagine moving to a position B and looking at the same point X on the glass. You see a different position B'. The LEDs don't know if you are at position A or B, so they can't know whether to show the light from A' or B'.