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

Christoph Lampert xvid-devel@xvid.org
Fri, 25 Oct 2002 15:36:18 +0200 (CEST)


On Fri, 25 Oct 2002, Radek 'sysKin' Czyz wrote:
> First, it checks all possible vectors, but in a big squares of 32x32
> halfpixels. I look for a square in which the biggest number of vectors
> is. (this can be done faster, because there is a huge probability that
> more than 50% vector is -15..16, -15..16).
> After I know this square, I divide this square and find the 'most
> popular' vector which would fit in this squre (by dividing again and
> again). My final square is 3x3 pixels big, end even if there is only
> one vector in it - it would still be close to 'original' (32x32), only
> more precise.

Interesting... I'll do some tests.

gruel


P.S. one URL:

http://bs.hhi.de/~smolic/publications.html

The paper "Real Time Global Motion Estimation for an MPEG-4 Video Encoder",
is very interesting, also the PhD thesis (in german) 
"Globale Bewegungsbeschreibung und Video Mosaiking unter Verwendung
parametrischer 2-D Modelle, Schaetzverfahren und Anwendungen"

it explains lots of methods (e.g. GMC from MPEG2 motion vectors or from
optical flow). 

Smolic and I discussed a few things via email. He's interested in XVID,
but unfortunately can't contribute himself because his GMC stuff is
developed at Heinrich-Hertz-Institute as commercial application.