[XviD-devel] Adaptive Quantization

Christoph Lampert xvid-devel@xvid.org
Mon, 30 Sep 2002 10:18:27 +0200 (CEST)


On Mon, 30 Sep 2002, Marco Al wrote:

> Christoph Lampert wrote:
> 
> > That's why I was wondering what other effect there might be.
> > Maybe there is a bug in adaptive quant which wasn't noticed before...
> 
> You are doing transform/quantization twice? I was assuming you did a slight
> rewrite and were doing the adaptive quantization inside mbtransquant. If not why
> not just break in mbtransquant when it reaches the MB you modified the dquant
> for and step through it to see where it starts to differ from what you did in
> the adaptive quantization code?

Yes, that's what my next steps would be for debugging. I just thought I'd
ask about side effects I might have overlooked (I found another one: 
With higher "current" for the an MB (e.g. the one _after_ the one with
Q+1) ME gives higher weights to motion vector bits, so motion vectors
tend to get shorter. That's because my changes in quant come _after_ ME,
because they act on motion compensated data. 
But the effect is very small and not the problem, either... no different
if I disable MV penalties.

So I guess I'll do some debugging later...

gruel


> There might be a bug, the whole dquant LUT obfusciation business promotes it
> (seems easier to just work with an unencoded mb.dquant and just encode it
> correctly in mbcoding.)
> 
> Marco
> 
> _______________________________________________
> XviD-devel mailing list
> XviD-devel@xvid.org
> http://list.xvid.org/mailman/listinfo/xvid-devel
>