I would like to start a general thread about vector quantization, which seems a promising direction especially for multimedia compression.
I know it is already used in Opus audio codec and is considered for AV1 video coded - any other compressors using it?
A natural approach is PVQ (Pyramidal/Perceptual Vector Quantization - nice introduction: https://people.xiph.org/~jm/daala/pvq_demo/ ) - any other promising approaches?
I have recently taken a closer look at PVQ and was able to reduce its mean-square error (MSE) usually by more than 10% thanks to replacing its projection with a more tunable one: https://www.dropbox.com/s/jgghkeov1swu53x/pvq.pdf
Any other interesting improvements?
And generally - do you think vector quantization is indeed a promising direction?
For multimedia only or also for other types of data? Which ones?
update: discussion with Jean-Marc Valin: https://groups.google.com/a/webmproj...el/4P9il8Pbb58