[XviD-devel] R-D-opt trellis lagrangian lambda (what a great subject!)

Christoph Lampert chl at math.uni-bonn.de
Thu May 15 12:03:49 CEST 2003


Hi,

a few days ago I had an minor idea for trellis and didn't find a
problem with it, so wanted to discuss with you: 

How about varying lagrangian parameter "lambda" not only with the
quantizer but also with position in the block (with frequency)?

As expressed in (typical) MPEG-quantization, higher frequencies are less
important for the human visual system, so saving bits there should be
enforced, and saving bits at well visible low frequencies should rather be
suppressed. With H263 quantization this isn't done, and fixed lambda
trellis quant doesn't care about it either. 
I don't find it logical to assume that the balance between rate and
distorsion should be constant over all frequencies. This could be 
expressed by a varying lambda. 
Overall PSNR would of course drop, just as for most "HVS" routines. 

But couldn't this work?

gruel


P.S. IIRC the value lambda=0.85*quant^2  was proposed by Sullivan and
Wiegand in 1998. But this was for Mode Decision/Adaptive Quantization, 
so on (macro)block level, not in DCT domain. 




More information about the XviD-devel mailing list