[XviD-devel] 2pass results

[Edouard Gomez]

Edouard Gomez (ed.gomez at free.fr) wrote:
> some results

I added the bquant formula compensation (i transform the bquant in
2pass2 so when the encoder applies its (ref+quant)*ratio/2 + offset, it
respects the one computed in our 2pass2 code)

The results are strange, really weird.

It helps a lot in respecting the frame size prediction, but it hurts
very bad the overall PSNR of some sequencies (not all, and probably only
non natural sequences like credits)

I outputed two graphs showing the ratio "expected frame size" (called
scaled size in 2pass2) and real frame size (called actual frame size in
2pass2)

2pass w/  inverse bquant formula: http://ed.gomez.free.fr/vrac/bi-comp.png
2pass w/o inverse bquant formula: http://ed.gomez.free.fr/vrac/bi.png

As you can see, we are much closer to the 1 ratio in the copensated
case, but we have big loss of PSNR in certain scenes. See below.

Some numeric results now:

Complete star wars stats but i encoded only the first 5000 frames:
Old code:          y(dB):44,076629 u(dB):47,707196 v(dB):48,760355
Old code + comp:   y(dB):43,685551 u(dB):47,230051 v(dB):48,471633
uni + comp:        y(dB):43,821356 u(dB):47,469009 v(dB):48,601345
uni:               y(dB):44,400525 u(dB):48,009571 v(dB):48,931644
bi + comp:         y(dB):43,716695 u(dB):47,392992 v(dB):48,540499
bi:                y(dB):44,400631 u(dB):48,015106 v(dB):48,928525

Matrix1 Chapter29 5000 frames:
Uni      : y(dB):41,259619 u(dB):49,825857 v(dB):51,505201
Bi       : y(dB):41,287525 u(dB):49,857670 v(dB):51,545100
Uni+Comp : y(dB):41,114162 u(dB):49,698736 v(dB):51,379006
Bi+Comp  : y(dB):41,134136 u(dB):49,696732 v(dB):51,392731
Old+Comp : y(dB):40,919226 u(dB):49,574534 v(dB):51,288877

"Old" means i used old code from cvs (quite buggy)
"uni" means i only use backward refernce quant to compute the bquant
"bi" means i use the backward reference quant + a prediction of the
     forward reference.
"Comp" means i compensate the formula applied by the encoder loop, so
       the bquant computed in 2pass2 is really used by the encoder

I must go to school more results and a patch tonight so everybody can
test and give its opinion about my changes.

NB: frame size respect is quite different from constant quality respect,
that's why perhaps the 1dB loss is not that important... (but 1dB in
some cases seems to be quite huge)

-- 
Edouard Gomez-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://edu.bnhof.de/pipermail/xvid-devel/attachments/20030528/28ab83ff/attachment.bin