[MEncoder-users] Tuning hqdn3d

[Nicolas George]

Le quartidi 24 floréal, an CCXVII, RC a écrit :
> This is a pretty pointless question.  2-pass and CRF will both give you
> the latter.  The only difference being that CRF will only consider it's
> best-guess for quality, while 2-pass mode will chose the best possible
> quality for a given bitrate distributed over the length of the file,
> within specified min/max instantaneous bitrates.

The question is not pointless, but you fail to see my point, which is this:
for any file size, the best overall quality will will be achieved when the
local quality is constant.

> With CRF, you are dependant upon the equation correctly
> estimating visual quality.  With 2-pass, you are merely depending on it
> for an approximately appropriate distribution of bits.

And "appropriate" is precisely determined as "that will give the best visual
quality".

On other words: if the rate control algorithm has a few spare bits to put
either in scene A or in scene B, it needs to know which one already looks
better, to gives them to the other.

Yet in other words: when we do not care about the actual bitrate, a
multi-pass target-bitrate system tries to emulate a fixed quality system,
with more complex algorithms and an irrelevant arbitrary number.

>							 ie. Even if
> the ratecontrol equation guesses that an extremely low bitrate will work
> for a particular video, as long as it is inaccurate in the same
> direction over both comlpex and simple scenes, it would still work fine
> with 2-pass.

The hypothesis that his is inaccurate always in the same direction is very
strong: the more I think about it, the more I believe it is equivalent to
having a perfect subjective-quality comparison function.

> But more importantly, and less controvercially, some options just
> require look-ahead, which x264 in single-pass/CRF mode simply can't do.

This may be true, and in fact I wrote about it. Unfortunately, it does not
seem to be implemented. Just look at the contents of a 2-pass log file:

in:8 out:8 type:P q:10.00 tex:0 mv:0 misc:152 imb:0 pmb:0 smb:576 d:-;

There is really very little lookahead information. It would probably enough
to achieve near-optimal I-frames repartition, but not much else. And even
that, I do not think it is implemented.

To sum up:

- The best overall encoding should have a roughly constant subjective
  quality.

- Constant quality encoding tries to provide it directly, target bitrate
  encoding tries to achieve it with complex algorithms.

- Someone who do not care about fitting his video exactly on a CD or DVD
  wants to choose a quality level, not a target bitrate. Finding the bitrate
  to achieve the desired quality is hard.

Conclusion: constant quality is the best choice for someone who do not care
about the exact size of his videos.

Furthermore:

- Multi-pass encoding could be beneficial for global optimizations of the
  encoding, but current implementations of the codecs only use the global
  information for rate control.

Conclusion: multi-pass encoding is currently useless for someone who do not
care about the exact size of his videos.

Regards,

-- 
  Nicolas George
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 197 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/mencoder-users/attachments/20090513/09209829/attachment.pgp>