[MPlayer-users] Is PSNR a definitive way of judging the quality of encoded files?

[Jason Tackaberry]

On Sat, 2005-09-24 at 13:28 +0100, VJ wrote:
> I found that with mbcmp=4,cmp=2,subcmp=2 the PSNR was the highest
> among various values of cmp and subcmp. Also the different values of
> mbcmp I tested for were all the ones mentioned in man page of
> mplayer(0-10). I used 2 pass encoding.

PSNR is one metric, and it's by no means definitive.  I've seen people
debate endlessly over which encoder values should be used for what, and
that foo and bar yields 0.65 higher PSNR for this, but don't use baz for
that because it will lower PSNR by 0.3, and so on.

PSNR might be a good starting point but it's no substitute for your eye.
The problem when measuring the "quality" of lossy audio and video codecs
is that they're going to introduce artifacts.  X may be closer (by some
algorithmic measure like PSNR) to the original than Y, but it may
introduce artifacts that are very perceptible, or worse, annoying.  I'm
extremely sensitive to and annoyed by blockiness in my videos.  The eye
will notice blockiness much more than it will notice random noise, but
yet noise will lower the PSNR more (especially if it's randomly
introduced, rather than using something like cmp=NSSE).

On playback, I will pass my video through post processing filters that
inevitably lower PSNR because they make the video further different from
the original.  (The filter chain I tend to use is -vf
fspp,noise=8ah:5ah,eq2=1.2:1.1.)  Introducing noise is a common enough
practice for post processing, because it masks the kind of digital
artifacts that the eye is perceptive to, whereas high frequency noise is
much more easily ignored by the brain.  Given that we purposely filter
the video whose result is lowering the PSNR, it should tell you that
PSNR is not a definitive measure of quality.

There are general practices over what options to use for what kinds of
video, and which options use should use for everything if quality is
important.  When you're experimenting, PSNR is one metric you can use
for judging for yourself.  For example, if your eye has a hard time
discerning a difference between two samples, you may fall back on
whichever one has the best PSNR.  But personally, I trust my eye a lot
more than PSNR.

Cheers,
Jason.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 229 bytes
Desc: This is a digitally signed message part
URL: <http://lists.mplayerhq.hu/pipermail/mplayer-users/attachments/20050924/627c93ad/attachment.pgp>