Re[6]: [XviD-devel] motion estimation for B/P/I decision

[Marc FD]

> > i saw it. sort of sad of MB decision ?
> 
> Just a simple motion estimation (the simlest I could think of) and a
> sum of all SADs found is used for the decision... I also tried sum of
> squared SADs but it worked worse.

that's what i understood mainly ;)

> > i want to say it CLEAR. i've done LOTS of test/coding/compare on
> > SKIP decision the last past month, and i'm almost sure of a thing now :
> > you can't avoid smearing with SAD/SSD or anything else
> > (Hadamrad based,ect..) because it's impossible to know if it's due
> > to noise or to very slow gradient move.
> 
> OK I understand now. What do you suggest? We can't just forget about
> skip-mode imho...

yes i know :(. 
but XviD is THE better MPEG-4 codec ever (tm), so i hope we'll 
found a solution. after a very precise ME, i saw that very smooth&slow 
movement was detected (maybe thanks to Hpel)
but i eager we get a lot of noise movement too.

> However, did you notice that B-frames help with the smearing effect?
> It seems that very small movement is just bigger, and as a result
> skip-mode is not used. However, it's used for intermidiate  B-frames
> and this results in a nice interpolation between one P-frame and
> another. 

an idea to avoid any smearing : P-200B-P like gops in long still scenes.
but i'm not sure it's even MPEG-4 compliant !
 
> I found that out on some anime, where smooth moving walls sure have
> this artifact.

moving clouds are the ultimate smearing source. i have a 200 frames 
long scene with mainly a cloud moving few pels/frames. it almost all 
my tests failed on it. it's really a torture clip. 

on anime it's very easy to see because it's more "clean" than a movie.

MarcFD  marc.fd@libertysurf.fr