[XviD-devel] Adaptive quantisation
Dirk Knop
xvid-devel@xvid.org
Sun, 04 Aug 2002 19:46:09 +0200
Hi,
Edouard Gomez wrote:
>I don't know if doom9 uses to review open source projects. But i've
>never seen reviews using experimental code. Koepi's builds are cvs
>snapshots + the koepi's decision to include square me or ... i would
>have never used that to test a codec. I'm not criticizing koepi, I'm criticizing doom9's choice.
>
>When someone has to review the codec, it should use a stable release
>(or the current stable snapshot, as our releases are not called
>releases). I would have laugh if someone had tested mozilla 2 years
>ago with a nightly build : "well mozilla is shit, it crashes all the
>time, renders an html page in more than 1 minute..." or tested my
>speedtouch usb driver cvs when i'm working on new features...
>
>Perhaps the solution is koepi proposes a stable - use for review
>binary... don't know.
>
>
I made a "plain cvs" build too which we tested on SPR. It looked even
worse. (Somehow EPZS is working even if it isn't coded "correctly".
There were more visible blocks with PMVfast in these extreme
situations.). That's the only thing I modified vs. the CVS. The first
build (doom9 used the CVS+EPZS builds for a second test, respect for
using other binaries too and check the results) was with hardcoded "cc
kf" treatment, which he then told "not to recommend" (I totally agree
there, it was just a proof-of-concept or better "let's see if it works"
version). But the next builds he used (it was from CVS 01.08.) was
"vanilla" with EPZS and looked better (the screenshots are from that
encodings btw.). Well, we have to see if the old luma masking code is
working better there. The curve treatment sure did a good job, but if
the new algo is quantizing everything below/above some thresholds away
the results are explainable.
Well, just wanted to mention that Doom9 tried to be fair, and he was
kind of nice in his review.
And please, with all respect, the CVS is in a stable state for "vanilla"
builds.
If anyone try to do a PSNR check and encode something with the "new"
luma masking code, we could compare that to the "plain" version without
the usage of adaptive quantization to see if the impact is that bad.
Don't get too picky on me(I once tried b-frames and took the build down
when you aske me to, gruel!), if I made a failure I'd apologize, but I
can't see where this was wrong.
Regards,
Koepi