[XviD-devel] Interview with Gej, mentioning XviD

[Christoph Lampert]

On Mon, 22 Mar 2004, Radek Czyz wrote:
> 
> 2. Have you guys heard of using non-linear distortion in R-D
> search/decision/whatever? Currently, we've got
> rd = bits + lambda*distortion
> which looks fine, but means that *any* level of distortion is fine as
> long as we save even more bits. However, a a different function, such as
> using distortion^2, would prevent distortions from going too far... Have
> you read something interesting about it?

Hi Radek, 

I know you know this, but maybe somebody else is interested, too: 

In theory, distorsion D of R+lambda*D is already quadratic (it's 
sum-of-squared quantization error). The rate R includes bits for 
vector as well as bits for coding of quantized coefficients. 

To speed this us, it is only approximated by  R + lambda*D with 
R = vector bits and D=SAD (and different lambda), and there, the 
SAD somehow compensates that the coefficient bits aren't taken into 
account. Usually, SAD doesn't tell anything about PSNR, because it's 
calculated before the reconstructed images (from quant/dequantized DCT 
coefficient) is subtracted from current. 

I don't there's a theoretical background for this working, it simply 
works, and since the number of vector bits is bounded from below by zero, 
too big distorion can't happen anyway. 

If you find a combination that works better, hey, that would be great. The 
problem might be how to "measure" is, because PSNR won't help there. 

gruel