[XviD-devel] R-D "optimal"

[Christoph Lampert]

On Wed, 7 May 2003, Marco Al wrote:

> From: "Christoph Lampert"
> 
> > You mean perform inverse DCT and calculate error in the image domain?
> > I guess only performance penalties stop us from that.
> 
> I think he is saying the present error for an individual coefficient is
> taken as the difference between the unchanged quantized coefficient and the
> lowered/tresholded quantized coefficient ... whereas it should be between
> the original coefficient and the dequantized lowered/thresholded quantized
> coefficient.

Actually, Trellis doesn't use a "original quantized coefficient is
this... should we lower it?" approach, but simply has a set of possible
values and chooses the best from those. Error _is_ always calculated
between non-quantized and different quantized/dequantized values. 
 
> Anyway ... as you said before, for MSE the DCT remains orthogonal.

In maths/theory, yes. But since we talk about algoritms where a single bit
changes behaviour, rounding errors might be different in frequency
domain than in image domain. 
But image domain error can't be calculated for every DCT coefficient
independently, so it would be terribly slow.

gruel