Re: [XviD-devel] Adaptive quantisation

[Dirk Knop <dknop@xxxxxxx>]

Hi,

Michael Militzer wrote:

> before. Now I have no idea what went wrong. Blockyness can be caused by a
> lot of things, simplest is that the quantizers were too high (could be
> caused by lumi masking - btw: was lumi masking activated for the SPR encode
> at all?), worst would be a serious problem within core. btw: I have SPR
Luma masking was activated, and I asked Doom9 to tell me where the 
quantizers range around. he took a short look and they were around 8-12 
(and sometimes higher), which is definatly the cause for blockyness in 
my experience (even with XviD which in general handles higher quantizers 
more gracefully than any other codec I "played" around with).

> myself and used a (longer) sequence of it for my qpel encoding tests: I had
> no blockyness problems (but I had a problem where you can see a "quality
> jump" whenever a keyframe is decoded - that's something doom9 mentioned
> already in his last comparison and I didn't believed it at first, but I have
> to admit it's really annoying) - So there's lots of work to do now...
>  
Well, this problem of "annoying degrading quality after iframe" should 
be solved with the new "scene based" curve compression which foxer was 
so kind to implement. Now the quality should be stable if no consecutive 
keyframes occur (which is a generally hard to handle thing, I think the 
treatment we inserted there helps it a lot, at least it does when I head 
for 1CD with "action" movies).

Michael, could you please test the first scenes of SPR if they still 
look that nasty when heading for 2CDs bitrate with the old luma code? 
I'm really eager to see that... The bitrates I get with the matrix now 
in the 1st pass are amazing compared to those with the former code, but 
I'm not sure if this is really good as it could have visual impact (I 
mean this as in: very easy to see).

It would be really nice if this only derives from the "wrong" luma code 
(btw., luma masking was switched on in all XviD tests of Doom9 for this 
comparison).

Best regards,
Dirk