[MPlayer-dev-eng] 2-pass vqcomp [was: PSNR plot]

[Zoltan Hidvegi]

Jeff Clagg wrote:
> > PSNR may be high, even if 2% of the movie is almost unwatchable.  I
> > did not like this so much, that recently I only use constant QP lavc
> > encoding with well chosen lmin=lmax and v[bi]_q(factor|offset).
> 
> Not to derail this topic, but the default vqcomp is a bit low. 0.6 or
> 0.7 is usually better, and more closely matches constant QP. You should
> definitely play with this before you cosider ditching 2pass altogether.

That's an interesting topic in itself, so I've changed the subject.
Actually, I've gone up all the way to vqcomp=1, and I got better
global PSNR and more even quality.  But there was (and maybe still is)
a bug in lavc which sometimes causes the QP go up to 31 even when you
use vqcomp=1, this usually happens when some section is completely
black, or there are some very dark sections in the movie, and it is
even more likely to happen in 3-pass mode.  That's the main reason I
gave up on 2-pass, even though when it works, it seems that 3-pass
gives you less dispersed PSNR distribution than constant QP.

The other problem with encoding quality measurement is that high PSNR
does not always correlate with precievced quality.  You'll notice
blocky compression artifacts on a smooth region with high psnr much
more than a highly complex region with lower PSNR.  That's why thre
are complexity masking options, they may increase perceived quality
while reducing PSNR.

Zoli