[XviD-devel] _real_ adaptive quantization

[skal]

	Hi,

On Tue, 2002-10-01 at 12:10, Christoph Lampert wrote:
>
> > 
> > 	How to you decide the q vs q+1 quant? Do you call a round
> > 	of quant(q)/dequant(q)/quant(q+1)/dequant(q+1)/choose best ?
> 
> Currently it's brute force, as you describe, but not only for Q and Q+1,
> but also Q+2 (not more, because it wouldn't always be possible to return
> to Q in next step). It really calculated MSE of DCT-coefficients.
> And yes, it's veeeery slow ;-) 
> 
> I thought of using more advanced routine, like cycling through
> coefficients, calculating several distorsion in one step, but all this is
> "only" speedup and at the moment, I'm after the theory what can be gained
> at all. 

> >
> > 	  (a*256 + b)%q  = ( (a*256)%q + b%q ) % q. = (b+Offset(a))%q
> > 
> > 	(quant error is not exactly (a%q), but that's the idea)
> 
> Approximation is always good... 

	(it was just an approximation for the mail's clarity (!). Code 	snippet
contains the real expression...)
> 
> But I just got another idea: There are 4096 possible DCT values per coeff. 
> There are possible 31 quantizers. If I create a table of
> resulting distorsions, when quantizing any coeff with any quant it would
> "only" be 126976 entries, right? 

	Yup, it's just that the hashing reduces the table to a
	more reasonable 6Kb size...

> Now, the most frequently used values will become constant anyway, because
> the quantized coefficient becomes zero. If that happens for one quant, 
> its the same for all higher quants, too. So there is a huge potential 
> for saving space, if it's really needed. 

	Hem... measuring quantization distortion, one is interested by
	the *remainder* of the division Coeff / q, not the quotient
	(sent to bitstream coding later on, when Q is decided).



	Btw, my primary idea for this stuff was performing a global
	pass on the whole image to decide the VOL's quant value,
	raising it if the (lower) PSNR still falls within a user-
	specified acceptable range. But it may also apply to the
	q/q+1/q+2 macroblock-based tests you're conducting...


	bye,

		Skal