[XviD-devel] Global Motion Compensation

Christoph Lampert xvid-devel@xvid.org
Thu, 24 Oct 2002 13:27:57 +0200 (CEST)


On Thu, 24 Oct 2002, Michael Niedermayer wrote:
> encoding is not more difficult than decoding IMHO, as the motion estimation is 
> allready done, we just need to analyse it to find good warp points ...

For translational: yes.  For the others: No, ordinary motion vectors are 
by far not exact enough to find subpictional zoom/pan etc. parameters. 

Most likely that's the reason why DivX5 does translations only.

> > Maybe more tweaks are possible for fast decoding, too.
> > (pre calculated tables or something...)
> hmm, i doubt it a bit

Me, too... but who knows ;-) Is GMC decoding MMX/XMMed? Maybe it's 
3dnow/SSE/SSE2able? 

> > The only argument would be that you don't gain anything from GMC, which
> > may be true for most videos, but sometimes it seems to help during
> > zoom/pan, so I'll at least have a look.
> hmm, its allways a question of complexity vs. compression gain, never just 
> compression alone, or why dont we all do full ME search? 

Okay, partly agreed. A slow feature should only be enabled if there is 
a sufficient gain, but what "sufficient" means depends on the application. 

Full search ME is an extreme example of slowdown at hardly any
better ME results (than PMVfast/EPZS) in all sequences I know. 
If GMC gives 1dB gain in 5% of sequences and nothing in 95%, it's somehow
a different story... those things can be tested and the feature only
actived when needed. 

gruel