Yes, I tested it too and it's very nice! It finds images with differences in color, and even superposed text, etc, just as promised in the video. Amazing, really.
Maybe it calculates some kind of hash on every image, an then it looks for images that are at a distance lower than n to such hash? But I don't know any hashing system for images that could work this way... could anyone explain?
Edit: a fingerprint is mentioned in the video, so I guess there is some kind of hashing system after all.
I've had a program that did this for my on-disk pictures for a long time: DupDetector (http://www.photo-freeware.net/dupdetector.php). I'm sure they have some form of documentation on their algorithm somewhere...
Comparing a predefined list of pictures to each other is totally different from pre-generating a hash of some sort that can match any image.
An idea I had was some sort of contour tracing, that will track each contrast path in the image, and then create a code the defines it in terms of changes in direction of the contour - that will handle rotations.
You make a large number of those for each contour in the image. Then match new images against that list.
It really does work - and even finds rotated images.