[XviD-devel] Results of VQEG sequences

Christoph Lampert xvid-devel@xvid.org
Sun, 2 Feb 2003 14:31:03 +0100 (CET)


Hi,

I finally did the first run of VQEG sequences. For several reasons, I had
to switch from full 720x576 / 720x486  to CIF format 352x288 / 352x240 

You can find some results at 

http://andrei.myip.org:8000/users/chl/basic.html    (nothing special)

http://andrei.myip.org:8000/users/chl/chroma.html   (CHROMA_ME)

http://andrei.myip.org:8000/users/chl/gmc.html      (GMC enabled)

the rest should come later today or tomorrow, in total it'll be any
combination of CHROMA, GMC, EPZS, PMV_EXTENDED and USE_SQUARES (32
tables, 8 HTML pages) for the 20 VQEG sequences. 

My first impression is, that "extra" features like EXTENDED search and
SQUARES often don't help anything at all, or only in statistical margins.  
INTER4V mode as well. Heuristics for them has to be improved. Currently,
switching them on in default high quality modes often is a waste of time
and we should rather improve other things, like detecting the cases where
normal ME fails, instead of always spending more time in search. 

QPel performs very well in scenes with much but very slow motion
(src2,src18 = slow zoom, src10 = slow translation), maybe this could be a
hint when to enable it. 

GMC sometimes helps a lot in fullpel mode, but this effect vanishes with
halfpel, so I guess it's the GME that isn't good enough (as
expected). I'll redo these tests when GME is better. 


Anyway, I really really need some volunteer to go through all the data
and search for "patterns". Maybe do some plots, or statistical analysis. 

I have to hand in my Ph.D. thesis this week and do oral exams afterwards,
otherwise I would do it myself. 

gruel