[MEncoder-users] Experiences of video encoding

[Nicolas George]

Le septidi 7 germinal, an CCXV, Mathieu Monnier a écrit :
>  * Quantizer scales in XviD/lavc and x264 differ. Going from 1 to 51 by 
> step of 2 in x264 is fine, but doing so in mpeg4p2 isn't, because the 
> quantizer scale is too coarse around 1. I would do 1, 2, 3 and 4, then 
> by step of 2 till 16, then by steps of 4 or more. You'll do less tests 
> that way, and the RD points will be more smoothly spread along the RD curve.

You are right. I hesitated adding points for 1 and 3, but I let it go. If I
make a second phase with XviD or lavc, I will certainly include them. But
for now, I do not think the extra point would have changed my conclusion.

>  * Why do you denoise the video ?

I have met videos with a lot of unwanted noise, and denoising had a very
good impact on quality. There is no deeper reason. Maybe I should do some
tests without denoising.

>  * for XviD, I don't know if trellis & chroma motion are enabled by 
> default. If not, you should do so.

Trellis is on by default, and the documentation says that vhq is better
anyway. Chroma motion is on by default.

>  * XviD should be better, PSNR wise, for h263 quantization type.

That is what I find, except at very low quantizers. But the difference is
very small.

>  * How do you compare PSNR figures with different resolution ? In what 
> resolution do you compute the PSNR.

I reassure you, I did not make the mistake of evaluating quality on the
resulting downscaled video. If I had, I would have claimed lossless
compression at 27 kb/s by downscaling to 16×9 pixels.

The details of the evaluation is in the page. I re-scale the resulting video
to the original size, and compare at that size, and I scale both the source
and the encoded video to 1280×something (1280 being a common resolution for
screens and video projectors) and do the same.

>				      How do you conclude that native 
> resolution is good up to 25, then resolution should be lowered ?

I look on the plots where the curve for native and downscaled cross.
Quantizer 25 at native size gives almost exactly the same bitrate and
quality than quantizer 23 at "short" downscaled size.

>  * Your plots are mostly unreadable ( too many curves, not enough 
> colors ), and hard to visualize in Firefow, imho ( the original size is 
> too small, and zooming prevents you from seeing the big picture ).

The PNG versions are unreadable, I am aware of that, but I did not manage to
make things better. With SVG versions, you can toggle curves by clicking on
the legend, which gives much comfort.

>  * What does the curves PSNR HD/native size try to show ?

The goal is to see how much the way the quality has been evaluated may
change the results. These curves show that the answer is mostly: not much,
except for the upscaled encodings.

Regards,

-- 
  Nicolas George
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 185 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/mencoder-users/attachments/20070327/17240237/attachment.pgp>