[Dark Shikari]

Originally Posted by TGEN

You can compute discrete first- and second-order differences of every frame to come up with optimal tesselations, depending on your maximum resolution. On big SIMD hardware like GPUs, that's easily parallelisable, but a lot of work elsewhere. I don't see fixed-function decoding DSPs achieve that sort of parallelism within the next five years either, so it's more of a future thing. But at that point, the bandwidth and storage available will make it irrelevant (like buying a quadcore to run your word processor these days).

Additional note, you could compute the entropy of each frame, and based on that assign a number of prototypes to it, then run stuff like RLVQ on it to come up with midpoints for your n-gons. You would expend ridiculous amounts of computation time per frame though when encoding, but decoding should be pretty straightforward.

Have fun doing DCTs of non-square shapes.