[XviD-devel] adaptive quant improvement

[Christoph Lampert]

On Tue, 28 Jan 2003, Marco "elcabesa" Belli wrote:

> i'm worknig on a new adaptive quant to give more bit to mblock where
> there are more details or edge. i am testing it and it seems to work..
> but i doesn't know how to make enance his performance. 

Can you explain a little what you did? Others might be interested in 
adaptive quant as well. 

And since I just read "details or edge". I think it's commonly agree
that a good strategy is to give more bits to edges (to prevent mosquito
noise), but less bits to "highly structured areas", because errors are
much less visible there. Maybe also more bits to flat areas, because
errors are very well visible there, and also, there aren't many bits to 
safe there anyway. 

In case you didn't check, Dirk Farin did some research on this for his
MPEG-2 encoder. I don't know an english language URL for this, but it's
surely in the sources

http://rachmaninoff.informatik.uni-mannheim.de/sampeg/ 

I guess what he uses is called "luminance masking" and "activity masking". 


P.S. Even with EXTRASTATS, don't rely on PSNR to judge your 
routines. Adaptive quant normally _lowers_ PSNR while it can still 
improve visual quality. 


gruel