[FFmpeg-devel] donation for snow

Jason Garrett-Glaser darkshikari
Fri Nov 7 05:53:58 CET 2008

On Thu, Nov 6, 2008 at 8:45 PM, Stefan de Konink <stefan at konink.de> wrote:
> On Thu, 6 Nov 2008, Jason Garrett-Glaser wrote:
>> That's odd.  My general experience with wavelets is that they give
>> much worse "visual quality per PSNR" than DCT-based does.  That is,
>> you can get many cases where a fully PSNR-optimal wavelet image is 3db
>> PSNR better than a fully PSNR-optimal DCT image (e.g. JPEG2K vs JPEG)
>> but the DCT-based one *still* looks better.
> the keywords are; psycho-visual model.

Of course; you're preaching to the choir here ;)  I wrote x264's psy model.

> But from the standpoint of Wavelets; would postprocessing with
> psycho-visual aspects in mind not help to improve the actual
> representation opposed to the optimal information carrier?

The only kind of postprocessing that I've found universally helps is
very slight fake grain and/or debanding, and only because it helps add
dither to compensate for banding.

The x264 psy model considerations all basically revolve around:

1.  Maximizing Preservation of detail.
2.  Places where detail preservation is not as important.

Neither of these can contribute to a postprocessing model because
postprocessing cannot really introduce new detail.  IMO, in practice,
most psy stuff has to be done encoder-side, or perhaps somehow in the
spec itself.  That is one of the major problems of modern video
encoding and formats; its still stuck in the stone age with regards to
psy stuff, and most encoders have either no psy optimizations at all
or extremely rudimentary ones that rely on random numbers pulled out
of thin air.

Interestingly enough, I've found most good video psy-opts have
significant parallels in audio encoding.

Dark Shikari

More information about the ffmpeg-devel mailing list