[XviD-devel] Interview with Gej, mentioning XviD

[Michael Militzer]

Hi Radek,

Quoting Radek Czyz <syskin at ihug.com.au>:

[...]

> > 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. 
> 
> Okay yes, but I was talking about VHQ mode decison and ME. I'm also 
> thinking how to repeat this step for trellis quantization, but I'm not 
> even trying to implement that one.
 
Apart from the fact that we're moving completely off-topic, I don't quite
understand what's so different about R + lambda*D^2 instead of using
R + lambda*D. Couldn't you achieve exactly the same by simply using a 
different  distortion metric (sum of absolute cubic error or something 
else)? That's the some as using SSE instead of SAD for D - nonetheless, 
the basic cost function remains linear.

Michael