Re[2]: [XviD-devel] qpel - problems I found

[Marc FD]

> True. Allow me to post quality modes from my VfW:
>
> int pmvfast_presets[8] = {
>         0, 0,
>         PMV_HALFPELREFINE16,
>         PMV_HALFPELREFINE16 | PMV_ADVANCEDDIAMOND16,
>
>         PMV_HALFPELREFINE16 | PMV_ADVANCEDDIAMOND16 | PMV_HALFPELREFINE8,
>
>         PMV_HALFPELREFINE16 | PMV_HALFPELREFINE8 | PMV_USESQUARES16,
>
>         PMV_HALFPELREFINE16 | PMV_HALFPELREFINE8 | PMV_USESQUARES16 |
PMV_EXTSEARCH16,
>
>         PMV_QUARTERPELREFINE16 | PMV_QUARTERPELREFINE8 | PMV_USESQUARES16
|
>         PMV_HALFPELREFINE16 | PMV_EXTSEARCH16 |
>         PMV_HALFPELREFINE8
> };

another funny thing : PMV_USESQUARES16 is not always better than
PMV_ADVANCEDDIAMOND16. i've made many tests and sometimes advdi > sqr,
sometimes advdi ~ sqr, and sometimes advdi < sqr. (in hpel and qpel mode)

> Now this is funny :) maybe qpel gives so small SADs that we have less
> INTRA blocks?

strange anyway. i don't think qpel is _that_ accurate.

> > - The crispness (noise?) i see with qpel, is it normal ?
>
> Crispness - yes. Noise - I think so.

it gives a sort of "film effect". like EE artifacts :(

> > i noticed that qpel makes files larger and lowers the PSNR, but i assume
> > it's
> > because it's not finished.
>
> Does new code still give you bigger filesize?

i downloaded it 4 minutes after your commit.
and yes,it gived me bigger filesize (~10%)

> As for PSNR - I think it's the noise. However, at fixed quant 3 PSNR

the decrease is ~0.1 dB at q2,
and was about ~5-6 dB at q10 with the old code (i will retest) !!!

> decrease for xvid is as big as 0.75dB, while the same in divx5
> decreases psnr only by ~0.1dB.

divx5 ? encoded with ? / decoded with ?

i only use the decoder (divx5 basic) and the result is _ugly_ with qpel
it typically looks like wrong compensation.

> This is strange, anyway.

sure. maybe a motion compensation problem ?