Re[2]: [XviD-devel] Global Motion Compensation

[Christoph Lampert]

Hi Radek,


On Fri, 25 Oct 2002, Radek 'sysKin' Czyz wrote:
> 
> Hmm OK. So if there is only one point (translation), I have a halfpel
> resolution? Is the GMC referece picture made by interpolation (so it
> would just be a copy of REf, REfH, RefV, RefHV in halfpel mode - and
> we have to re-compute halfpel reference in qpel mode?)) or there is
> some special filter (well I guess not).

There is another important task: mode decision! Decide whether to use GMC 
for a block or INTER/INTER4V. GMC might be better even if INTER gives
slightly lower SAD, because you don't have to encode extra vectors (GMC
will also be used to calculate a "median" predictor for this MB, so if
motion vectors are identical between GMC and INTER, the influence on
neighbouring blocks is very similar. 

gruel