The source code is up for download, too.
It seems reasonably accurate, but stop and think for a minute how much more accurate you could make it with two cams, one at each top corner of the monitor. Left to right accuracy could be pinpoint. Add another cam at the bottom center of the screen to ramp up the vertical accuracy and we might be able to throw away our mice...
I'm guessing the image recognition algos don't strictly need color, so all three cams could be black and white to reduce cost.