[XviD-devel] [BUG] BFrame encoder doesn't force intra for 1st frame

Christoph Lampert xvid-devel@xvid.org
Wed, 7 Aug 2002 15:17:26 +0200 (CEST)


On Wed, 7 Aug 2002, Michael Militzer wrote:
> > Especially since B-frame ME is (and will stay, even after
> > optimization) much slower than P-frame ME.
> 
> I'm not sure if MV hint (as it's now) is really that good (I remember Daniel
> reported that encodings using MV hint are about 5 - 10% worse than without).

I'm rather sure that it's not _good_. But it's fast and if people want 
it... 

> I ask myself if it wouldn't give better results to do the first pass with a
> lower quality mode with integer ME only (that's quality=3, right?), which
> should be relatively fast and then do a normal second pass with high quality
> (quality 5 or 6). This could be nearly as fast as hinted ME and could give
> better results. (I suppose that sizes of quality 3 encoded frames are simply
> equally larger than their equivalent quality 5 encoded ones, so the size
> predictability (and quantizer calculations) for the second pass could still
> work pretty well if the formula is adopted a little bit...)

Yes, the possibilites for real "hinted" ME instead of "forcing old MVs to
a new situation" are much more promising. I guess not even a full search 
would be needed in second pass, but only a refinement step. Either normal
qpel/halfpel-refine, or diamond search with halfpel diamond size. 

Could someone do some tests for: Two-pass, hinted ME and _disabled_
Halfpel in both runs? 

Because I guess, halfpel positions will almost surely be wrong very often 
between first and second pass, but fullpel might be more "accurated",
due to the image being less smooth. 

gruel

-- 
Christoph H. Lampert chl@math,uni-bonn,de | Diese Signature wurde maschi-     
Beringstr. 6, Raum 14 Tel. (0228) 73-2948 | nell erstellt und bedarf
Sprechstunden: keine, aber meistens da    | keiner Unterschrift. AZ 27B-6