[XviD-devel] Global Motion Compensation

[Christoph Lampert]

On Fri, 25 Oct 2002, Marco Al wrote:

> Christoph Lampert wrote:
> 
> >> Least median of squares seems to be the most popular method.
> >
> > Smolic starts with "structured" blocks only and uses a Maximum-Likelihood
> > approach after that. He claims results are very good (and fast).
> 
> That might be nicely combined with a MB classifier for adaptive quantization. He
> uses a faster (but slightly less-) robust estimator than LMedS.
> 
> Since you talked with him ... why does he in the "Real Time ..." paper only use
> the robust estimator for finding an estimate for the translation and not the
> other parameters of the affine transform? He does do so in his thesis for
> finding global motion from the MPEG-2 motion vectors, but he does not refine
> them afterwards for some reason ... he must have considered it, estimating all
> parameters before gradient descent search, but Id still like to know the reasons
> why it was not a good idea :) 

He just answered that it _would_ indeed be a good idea. Since we can
arrange things to have results of ME available before GMC (maybe not final
ones, but close to it), the procedure could be: 

1) ordinary (local) ME 
2) calculate start values of GMC coefficents from (some) ME MVs. 
3) calculate _real_ GMC coefficients (nonlinear optimization)
4) apply mode decision (inter[4v]/intra/gmc) for every MB and set correct
   prediction accordingly. 

This might be slightly modified, of course, e.g. mode decision
inter/intra at step 2 already etc. 

gruel