[XviD-devel] New Trellis Quant

[Christoph Lampert]

On Mon, 12 May 2003, Radek Czyz wrote:

> Hi,
> 
> > For me too, and also at fixed quantizer PSNR dropped heavily, so it can't
> > be too long vectors or something. Actually, I can't explain this at all.
> 
> That's all true. I don't know why, but BITS search hurts quality in
> most cases... Actually at quant 10, even the decision itself hurts
> PSNR and this is even weirder.
> My tests show that BITS search is reasonably good at quants 2, 3, but
> not higher.

Maybe one shouldn't just compare BITS, but use the usual lagrangian R-D

value = lambda*bits + distorsion

where lambda = 0.85 * quant*quant (for MSE distorsion measure)
or some similar value (e.g. 1 to 1.15 was better in my tests). 

But my real question is: Why was encoding _faster_ when BITS was activated
for halfpelrefine. I can imagine BITS for whole search being faster, if
the search terminates much too early. But refinement is always done. How
can a slower refinement step speed up the encoding? 

gruel