Cool stuff; there are lots of other techniques that also do pretty well.
This really highlights how insensitive we are to color fidelity in images. The granularity of the marked images is very low quality, but, the results are very good. This is the same reason that for JPG and other image compression techniques, you can downsample the hell out of the color components with no apparent degradation to the image quality.
Humans are extremely sensitive to color fidelity. Look closely as those processed images.
You can tell that something is "off" about them, because while the algorithm is extremely good, it can't reproduce the true color gamut of reality, or even of a low quality jpeg that is based on a real photograph.
Humans are naturally sensitive to color, but not to hue fidelity, as hue synthesis shows: all hues are synthesized out of 3 basic wavelengths, color displays are based on this. Also, we tolerate B&W and sepia images well, we even love them. We are sensitive to hue saturation and contrast, and the discussed algorithm fulfills these two requirements. Furthermore, we have been desensitized by the heavy use of compression in color images and by uncalibrated displays.
Heh, I clicked to add this to del.icio.us, then noticed I'd already added and tagged it three years ago. It's amazing how much goes around and comes around again without us noticing :)
This sort of problem is really hard, but a little bit of cleverness gets you a long way. Theres some really cool information theoretic bounds that you can cook up on how good various algorithms can do for video using whatever assumptions you have on the video. (eg does the motion respect physics? do we have sufficient sampling that we don't some ambiguous choice? etc)
This really highlights how insensitive we are to color fidelity in images. The granularity of the marked images is very low quality, but, the results are very good. This is the same reason that for JPG and other image compression techniques, you can downsample the hell out of the color components with no apparent degradation to the image quality.