[XviD-devel] GMC material with 3 warppoints?

[skal]

	Gruel,

On Fri, 2003-04-04 at 17:38, Christoph Lampert wrote:
> On 4 Apr 2003, skal wrote:
> > (ps: i've been conducting some experiments with GMC. Conclusion:
> > it's very hard to have good results with GMC :( 
> > I can give a hand if you want, nevertheless... )
> 
> Sure, please do! 
> 
> What kind of GME did you use? 

	a home-brew mix resembling yours: during the global GMC pass,
	i first select the flat-laplacian'ed macroblocks
	as candidate for GME, provided they are textured
	enough. Then I statistically fit a 6-parameters
	model with these samples, discarding seemingly
	outliers (this is the harder part! these bad
	samples can *ruin* the overall confidence).
	Then I finish with a regular pass that actually
	compares the GMC-predicted SAD against normal
	SAD, and make a decision for each MB.
	The most difficult really is selecting good
	sample for fitting the model. All the more that
	a good ME pass may be greedy on finding motion
	vectors that eagerly lowers the bit-count at the
	expense of the smoothness of the motion field.
	But as far as GMC is concerned, a good motion
	vector for bitcount might be a bad one for evaluating
	the overall motion :(

	oh well... the only promising idea I could come
	with is weighting the samples with w=1/SAD weight.
	Hence, very-well matched motion vectors contribute
	more to GME. This is slightly in contradiction with
	the above remark, I admit :)

	bye!
			Skal