[XviD-devel] adaptive quant improvement

[Marco Al]

Christoph Lampert wrote:

> You can't do adaptive quant after MC, because MC is done with the
> dequantized coefficients, so you have to know the quantizer before MC.

Sounds reasonable, then again ... you cant make a good decision about DQUANT
without knowing the residual, so you have to know the motion field before
determining the quantizers.

They are both true ... you cannot make the decision about the rate optimal MV
and the rate optimal quantizers in a single pass, fundamentally impossible.
Iterative solutions can work of course, but to make a choice what to put first
with only a single pass you need to know the potential rate reduction for both
mechanisms and the hit they will take by not having all the information they
need.

Personally I think assuming a slightly different quantizer than the real one
will hurt ME less than having no DFD to work with will hurt "good" adaptive
quantization methods.

Marco