[XviD-devel] H263 quant

[skal]

	Hi Pete,

On Fri, 2002-09-13 at 01:39, peter ross wrote:
> >	Hi,
> >
> >	there's something I find strange:
> >	H263 dequantization is the same
> >	for intra or non-intra coeffs:
> >	( C' = C*2q + (q&1?q:q-1) )
> >	But there's a different bias
> >	used during *quantization*.
> >	For intra, it's 0, and for
> >	inter, it's -q/2. I've checked
> >	MoMuSys, and it's the same.
> >	Does someone has an explanation/pointer?
> >	Something to do with the AC-values
> >	distribution function?
> >
> 
> both h263 and mpeg *quantization* algorithms were implemented
> directly from the VM18 document. from memory (i dont have VM18
> with me atm), the reason for this is to reduce rounding error caused
> by // division. i guess we could try to impove the quantization algo,
> such that the coeffs are more acurately decoded.

	hmm... that's what I first thought, but with a strictly linear
	quantizer, and assuming uniform distribution, the errors during
	division would be systematic, and cancel each others...
	*But* we're not assuming uniform distribution of AC, do we? :)
	An quantizers are not that linear either, due to the dead-zone
	around zero. Oh well.. this bias looks to me like a very 
	fine-tuned thingy, if any. Nothing really obvious at first
	sight, unless I've missed something...

> i guess we could try to impove the quantization algo,
> such that the coeffs are more acurately decoded.

	exactly :) I'm conducting tests for an improved
	TransQuantIntra/inter() module. Scratch notes
	(ie: some are unchecked, or simply wrong)
	are here:

http://skal.planet-d.net/coding/quantize.html

	I far from having sorted out this biases
	weirdness, for now...

	bye,

			Skal