[MPlayer-dev-eng] [PATCH] (mencoder) lavc option "notch_matrix"

[Steven M. Schultz]

On Sun, 19 Oct 2003, Michael Niedermayer wrote:

> > On their homepage
> >
> >   http://www.kvcd.org/portal/articles.php?lng=en&pg=20
> >
> > it says
> >
> >   The KVCD "Notch" Quantization Matrix is Copyright (c) 2002, 2003 of
> >   Softronex Corporation. Please contact us for commercial use.

	They have changed the page then from what it was a couple 
	months ago.   It used to say something to the effect that
	you could use the tables as long as the program was "labeled"
	as using them - i.e. a notice when running the encoder that
	"using the KVCD quantization matrices" or similar.
> 
> btw, does anyone know why this matrix is supposed to be better then the 
> default? did anyone do any tests? PSNR?, double blind subjective?

	Comparing the tables they give to the default MPEG-2 tables
	it looks like the high frequencies are quantized very differently.
	"Roll off" isn't the right term but that's what comes to mind.

	I've used the Kvcd tables as well as the TMPGEnc tables in 
	mpeg2enc (mpeg2 encoder from mjpegtools).   Not double blind
	testing but the resulting video was fine to me (not having golden
	eyeballs, using an older TV set, etc ;)).

	The difference in bitrate was a big surprise.   Using "-K kvcd"
	with mpeg2enc the average reduction in bitrate (as reported by
	mplex) was 20% (depends on the -q or quality setting - sometimes
	the savings was closer to 12%, other times 25%).   The TMPGEnc
	tables split the difference between the default and Kvcd tables -
	usually around 10-15% savings over the default tables.   Much
	depends of course on the source of the material being encoded.
	
	The claim that the Kvcd tables reduce the "splotches" during
	dark scenes is true - the dimly lit scenes do look better when
	using the Kvcd tables.

	Cheers,
	Steven Schultz