3 ms·
Keep in mind that when you're fitting a compression algorithm so close to the test cases, what you're really doing is pushing data from the test cases into the
by patio11 18y ago
Keep in mind that when you're fitting a compression algorithm so close to the test cases, what you're really doing is pushing data from the test cases into the compression algorithm. For instance, prior knowledge that blurs are important gets pushed into the algorithm.
This is why when testing compression algorithms against each other you usually weigh the total of the code and the compressed versions of some common corpora which is large relative to what you expect code sizes to be.
If you don't do this, I mean, you can do lossless compression of the Mona Lisa into one bit while remaining able to display any other JPG. Check bits of input. If first bit is 1, then decompress into the Mona Lisa. If first bit is 0, the rest of the stream is a JPG, interpret as normal.
Still, I'm of the opinion that it remains an outrageously good tech demo for GAs. (Visual, easily described task, results anyone can appreciate, etc.)
- reitzensteinm 18y agoI have been watching the development of this for a while, and I don't think that the algorithm has been tailored towards the test cases at all. The file format is extremely general (vertices forming polygons, color, alpha and blur per vertex). It's been fed anything and everything from around the net, but it takes so many days of computation to make the images look good that only a few examples were used in the article itself. If anything, Mona Lisa and Lena were specifically chosen to be hard for a system like this to compress (fine brush detail does not mix well n sided polygons). Blur for example compensates for the shortcomings of polygon rendering to make the whole system more general purpose.
- tlb 18y agoIt probably does poorly on periodic repeating textures. The DCT in JPEG handles periodic brightness (like a striped shirt) fairly well because it turns into a few high-frequency components (which are still roughly periodic despite the zig-zag transformation.) This compression scheme has no support for periodicity so it can only represent a striped shirt with a polygon for each stripe. Neither of his test images have stripes.