[MPlayer-users] Re: encoding quality - bpp, resolution, quants and quality

[Matthias Wieser]

Am Dienstag, 9. September 2003 22:07 schrieb D Richard Felker III:

> > Those bpp values are probably thought for two-pass encoding. If you
> > only use "vhq" then you need more bpp than if you use
> > "vhq:precmp=2:cmp=2:subcmp=2:trell".
>
> If you want max quality/compression for live recording, you should
> record to a lossless format first (or perhaps vqscale=1) then do a
> 2pass encode later with denoising to make a copy at reasonable size
> for archival use.

That should give best quality. But it's more work for me. So why let the 
CPU do idle loops if it is possible to do realtime 
vhq:precmp=2:cmp=2:subcmp=2, crop,pp=fd,scale and denoise3d?

At least for TV it works without framedrops. ("-1 frames dropped" ??)

But with video-in frames get dropped, even without denoise3d (1). Strange, 
because every other option is the same.

> > The latter _was_ better. Ther are no recent comparisons between both
> > but Xvid is better than many other modern codecs, so Xvid has
> > probably become better than libavc.
>
> Doubtful. You could try a comparison (PSNR or better yet double-blind
> :), but it's generally accepted by mplayer developers that lavc is the
> encoder of choice...

PSNR is not really usefull to measure visual quality. JNDmetrix (2) should 
give better results. But it's expensive. With gimp I was not able to  
process the difference between original and compressed frames so, that 
highly visible block artefacts gets emphasized and hardly visible random 
noise gets softed.

(1) My Hi8 cam produces much less noise than cable tv, so denoise3d is not 
needed.
(2) 
http://www.sarnoff.com/products_services/video_vision/jndmetrix/index.asp

    Matthias