[XviD-devel] Questions on Quantization

[Luca Piccarreta]

Hi Skal,
don't be so frightened...
that 1 value on the 63th coefficient it's not visible... I don't have
floating
point iDCT here, but it should be something like a +1/16,-1/16 grid.
The objective should be "waking up" the iDCT as to avoid situations in which
small varying blocks are coded with some DCT coefficients, but the encoder
side
iDCT transforms to a null block...at the decoder side, another iDCT might
decode to a non null block, and after a while, very visible drifts appear.
(I can't really hope I was clear).
Actually, this situation is true only for very low QPs (1,2...)
If you come up with some numbers/results (I'm not sure about that 1/16), let
me know....
Ciao, Luca.



----- Original Message ----- 
From: "Skal" <skal at planet-d.net>
To: <xvid-devel at xvid.org>
Sent: Thursday, March 24, 2005 2:07 PM
Subject: Re: [XviD-devel] Questions on Quantization


>
> Hi Luca (nice to see you here;)
>
> On Thu, 2005-03-24 at 12:22, Luca Piccarreta wrote:
> > Hi Skal,
> > Are you sure about what you said about Intra DCT Mismatch Control?
> > I checked the MS ref sw (W4866) and it looks like mismatch control
> > for intra dct is there (see inverseQuantizeIntraDCTcoefMPEG).
> > Of course it's not there for H263 Quant Type
> > (see inverseQuantizeDCTcoefH263).
> > I'm also checking the latest 14496/2 I have available (N5546,March 2003)
> > and I can't see anything about different mismatch control for Intra DCT.
>
> Oops, you're right! Xvid is lacking mismatch control
> for MPEG4 intra-dequant. And it's been so since a
> loooong time! (as long as i can recall, actually)
>
> > Now, for my personal opinion, I think that the whole float DCT stuff is
> > crazy, mismatch control is crazy, long live integer-defined transform
and
> > bit exact conformance :)
>
> Fully agreed. Carving in stone some fixed 16bits (or 8bits,
> even) coeffs for the iDCT in MPEG2/4 would have saved a lot
> of troubles, instead of opening the luring floating-point
> Pandora box.
>
> > And of course the impact of not having mismatch control for Intra frames
> > is almost null.
>
> Well, i'm not sure (until i test it). Correcting this mistake
> in XviD would be easy, but i fear it might slightly break
> decoding of previously encoded material: one might notice a
> faint chessboard effect on intra blocks, due to the toggling
> of coeff #63. Any opinion?
>
>
> later!
> Skal
>
>
> _______________________________________________
> XviD-devel mailing list
> XviD-devel at xvid.org
> http://list.xvid.org/mailman/listinfo/xvid-devel