3 ms·

Sounds like he's talking about an application of Principal Components Analysis (see, e.g. http://www.willamette.edu/~gorr/classes/cs449/Unsupervised/pca.html ht

• apu 16y ago
  PCA is optimal from an L2 perspective, but with images, that is often not the metric you want to use. Of course I was a little bit confused by the fact that he's talking about lossless compression but then mentioning jpeg, which is lossy. If you consider lossy compression algorithms, then for images what matters most is human perception. JPEG and the various MPEG standards for video already take advantage of this, by more heavily quantizing things that humans have trouble distinguishing changes in. Thus, while these compression schemes may increase the L2 reconstruction error, the results look "better" to human eyes than the results of PCA, where artifacts are quite noticeable. If you do want to use PCA, then the following method is quite a bit more efficient than straight PCA on images: http://www1.cs.columbia.edu/CAVE/publications/pdfs/Nishino_PAMI05.pdf http://www1.cs.columbia.edu/CAVE/publications/pdfs/Nishino_P... That's for a single image, but you could extend the idea out to multiple images. The problem with PCA is that it's a linear method, and lighting effects are not linear across a full image. However, in local settings (i.e., small patches of the image), the linearity assumption often holds, and you thus get much more bang-for-the-buck.