[XviD-devel] CheckCandidate

Radek Czyz radoslaw at syskin.cjb.net
Tue Feb 18 23:46:57 CET 2003


Heya,

> good timinig, I just sent my results on ModeDecision to the list a few
> seconds ago... Anyway, I don't really like the argument "It was better",
> even though it has a certain logic.

> I read your xvid-devel explaining it, but I would still prefer a somewhat
> "theoretical" reason for this, something like a model with parameters
> which we can then optimize.

I also prefer a scientific approach. However, probably the only way to
prove it would be to show that if SAD is high, it is no longer
accurate (or rather: even more inaccurate) as a prediction of bits
used (actually, I could prove it now using this bits thingy)
As it is inaccurate, but d_mv_bits() is not inaccurate, relaying
on d_mv_bits more in such case is statistically better...

Anyway, I got this idea and my initial test gave me 0.1 dB just like
that - so you see why I liked it.

> Is there any voluteer around to check what's wrong with medium quants at
> the moment? I guess from the practical point it would be better to have
> high errors at very high quantizers, and as good as possible decision in
> the area of quant 3 to 8 where most of the encoding takes place.

The inter/inter4v decision was difficult, I admit. Three factors play part
- how we add mv bits to SAD (lambda_vec8[] table, with expotential
scaling(quant) * (SAD + BIAS) ; where bias is constant) and the bias in the
decision itself (IMV16X16 , linear scaling(quant)).

First thing that comes to my head - if lambda_vec8[] was linear, equal
to current at q2 and q31, it would be higher in middle range. Exactly
what we need. It doesn't mean it would be better, though.

I'll check this, and other possibilities as well.

Radek



More information about the XviD-devel mailing list