[MPlayer-dev-eng] [LONG] PSNR comparison, new source and results

[Jan Knutar]

On Wednesday 16 June 2004 12:18, Romain Dolbeau wrote:
> Jan Knutar wrote:

> qpel not only raise the execution time by over 63% on
> average, it also decrease the PSNR in all case. trell
> isn't anywhere as expensive, and it does give you a
> nice improvement all the time. *dia=-2 is a tiny bit
> better and not much more expensive tha *dia=1.

I did a comparison of qpel at different bitrates, with very basic 
options. Source material is the trailer for "Kate & Leopold", from the 
PAL DVD. It's full screen 4:3, and I encode it unscaled.

I used mbd=2:v4mv, 2 pass encoding with different bitrates, with and 
without qpel.

The findings:

bitrate < ~750:	qpel decreases psnr
~ 750 < bitrate < 2000:	qpel increases psnr
2000 < bitrate:	qpel decreases psnr

Perhaps more interesting is the same expressed in bits per pixel instead 
of bitrate, like this:

bpp < ~ 0.075		qpel decreases psnr
~ 0.075 < bpp < .20	qpel increases psnr
.20 < qpel	qpel decreases psnr


What surprised me was that qpel at high bitrates actually decrease PSNR. 
It also explains why I've never seen qpel increase PSNR, because I most 
often encode at .25 - .30 bpp...


Unfortunately, my computer isn't fast enough to produce more 
comprehensive results. I want to do the same thing, but with trell 
enabled, and also using *cmp=2, and see if qpel behaves differently.