I just got back from a very cool demo by Brad Behm, who works in Dr. J.R. Parker's Computer Vision Lab here on campus. Very VERY cool stuff. They've got an app set up so that you can feed it an image, and it will search its database of several hundred images and return similar images. They run several comparisons simultaneously, checking edges, colours, etc...
It's not perfect yet, but they're batting over .500, which is much higher than the other computer-vision search groups are getting (apparently Brad's software averages about 57% accuracy, but this can go much higher to almost 100% depending on the images in the database and the source query image).
We talked about some of the issues and implications of this type of technology, and it could be an amazingly powerful utility. We're going to be talking with Dr. Parker about the possibility of hooking it up to CAREO (no promises, but we're interested - he's out of town at the moment, though) to provide another set of learning objects, with metadata generated automatically and on the fly by their application.