[XviD-devel] problem

[Michael Militzer]

Hi,

Quoting skal <skal at planet-d.net>:

> On Fri, 2003-04-04 at 18:05, elcabesa wrote:
> > i'm testing lot of stuff of xvid an i'm very happy about result=)
> > the only oine problem/bug i found is that enabling qpel in fixed quant 
> > encoding make psnr go down
> > you cannot notice it in a cbr encoding couse bitrate saving  make xvid
> using 
> > an  lower quantizer   so psnr raise =)
> > 
> > i don't know if it's a know effect of qpel, or if it's due to interpolation
> 
> > reuested by standard
> 
> 	So far I've seen, it's an effect of the interpolation required
> 	(6taps FIR+mirroring) by the norm: even at hpel position, the
> 	psnr lowers significantly (~.3db or worse) when using qpel.

yes, that's true. I noticed this as well. So in case you don't go for real qpel
precision but use only hpel position motion vectors, the lowpass filtered hpel
positions seem to _always_ lower PSNR (and increase filesize) compared to normal
halfpel mode with bilinear interpolation. But I'm not sure if this is really the
case because the FIR filter+mirroring is simple bad or because the lowpass
filter approximation (without mirroring) we use during ME is responsible for
this. - Well, I should simply try...

> 	( i mean: if you take the mv found by hpel search and simply
> 	promote them qpel mv without any further refinement search,
> 	the psnr lowers). Good candidate scenes for qpel are still
> 	to be found (but Michael (M) has some hints ;)

yes, there are scenes were qpel performs really well. This means that for a
fixed quant encoding, qpel gives smaller filesize _and_ higher PSNR than hpel.

bye,
Michael