Very nice.

One question: as the top-level comment in this thread noted, this algorithm may be specialized to certain kinds of images.

Can you release info about the kind of images/datsets this approach will pathologically fail with? That would be really really awesome.