[MPlayer-users] Re: [POOL] Moderate some users (was: Re: [-] Re: divx 6)

[Matthias Wieser]

Am Samstag, 6. Mai 2006 20:19 schrieb Vladimir Mosgalin:
> Hi Matthias Wieser!
>
>  On 2006.05.06 at 17:10:49 +0200, Matthias Wieser wrote next:
> > > No, PSNR proves nothing, especially so close. Differences in the type
> > > of motion comp (which gmc is) will make differences on this scale
> > > without reflecting anything about quality. You'll need to use a
> > > better metric or doubleblind viewing test to get any useful
> > > information.
> >
> > I need nothing. You will need to prove your claims. I'f you can't it's
> > your problem. I took a random video (with much panning) and surprise,
> > surprise, GMC improved psnr nearly 0.1 dB. If you wan't to use a better
> > metric or
>
> PSNR PROVES NOTHING.
>
> Now read this again as many times as you need to remember it.

Think about it's definition.

PSNR CAN PROVE MANY THINGS

"Now read this again as many times as you need to remember it."

Regarding image quality it's at least an indication.

> 0.1 dB improvement - I can only laugh at that. Even 0.5 dB difference in
> PSNR doesn't mean that one of the encoded version looks better.

You are free to invent a better metric. I have suggested to use SSIM.


> I've seen so many cases when psnr and my eye tell me opposite, so now I
> don't even try to compare psnr.

Nobody forces you to compare psnr. Just do your own gmc tests.

Regards,
  Matthias