On Wed, Apr 19, 2023 at 11:06:09PM +0200, Anton Khirnov wrote:
> Quoting Michael Niedermayer (2023-04-19 22:53:02)
> > On Wed, Apr 19, 2023 at 09:48:32PM +0200, Anton Khirnov wrote:
> > > When an encoder exports sum-of-squared-differences information in
> > > encoded packets, print_report() will print PSNR information in the
> > > status line. However,
> > 
> > > * the code computing PSNR assumes 8bit 420 video and prints incorrect
> > >   values otherwise; there are no issues on trac about this
> > 
> > Are the values in the "otherwise" case maybe good enough so they
> > worked for people with noone noticing ?
> 
> While working on this with a 10bit sample I was suprised this code and
> vf_psnr showed significantly different values (IIRC not even the first
> digit was accurate) and it took me a while to realize the scaling made
> the assumptions it did.
> 
> > > * only a few encoders (namely aom, vpx, mpegvideo, snow) export this
> > >   information; other often-used encoders such as libx26[45] do not
> > >   export this, even though they could
> > > 
> > > This suggests that this feature is not useful and it is better to remove
> > > it rather than spend effort on fixing it.
> > > ---
> > > I needed to resolve this code's interaction with encoders as a part of
> > > my multithreading work and spent a few hours on it. Making it work
> > > correctly in all cases seems nontrivial and duplicates a lot of the
> > > logic from vf_psnr.
> > 
> > Can anything missing in vf_psnr be added into it and then vf_psnr used ?
> > I agree that duplicating PSNR code and logic is bad
> 
> Nothing is missing in vf_psnr AFAIK, the difference is that these values
> are produced directly by the encoder, so you don't need a
> decoding+filtering pass to obtain the numbers.

That sounds like a missing feature.
vf_psnr cannot use the encoder values, it always needs to recompute them.
The encoder could export these values in metadata and vf_psnr could
then check if the frame pairs already have their psnr computed and use
that.


> 
> The question more broadly is - what is this supposed to be useful for?
> PSNR is a highly flawed metric and AFAIU state of the art moved several
> generations away from it. And for a quick-and-dirty quality estimate,
> the bitrate and QP might be more informative for most users, and are
> supported by more encoders. So what is the point of printing this
> information?

The "problem" isnt psnr specific
having a filter which could provide another metric and be able to use
encoder supplied data when available or compute them from encoder input +
decoded images in a way that is automatic would be useful
It would mean a single syntax a user could use to get the metric he wants
while using the optimal implementation

thx

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Those who are best at talking, realize last or never when they are wrong.