Possibly poorer quality in later builds

[AndyP]

Hi

I have just done a few encodes of The Matrix (R2) using the last few versions of XviD (18082002 onwards) and notice some abnormal 'blocking' over moving parts. I compared this to a XviD encode I did a while back (using 30062002 I think) and this was much better.

Current

Target size 747238kb
Search precision 6
MPEG/Modulated quants I range 2-3, P range 2-8
Credits fixed quant 12
Two pass with lumimasking and MVH
Alt-CC as default (500/90 strength 30 etc.)
AVS file (part)
--
crop(4,80,712,416)
BicubicResize(640,256,0,0.5)
--

Old

Target size - about the same
Search precision 6
H263/Modulated quants rage 2-31 (I think - i can't remember)
Credits fixed quant 12
Two pass with lumimasking and MVH
Alt-CC high 250 low 50 strength 50
I can't remember the crops but the size is 672,272 - would be either bicubic or simpleresize

I know that this is not an entirely fair test and will happily reinstall 30062002 and redo with same settings as done before, but just wondered if anybody has any suggestions off the top of their head. My guess was as the new one was at a lower resolution, if anything, it should look slightly better.

If I have done anything stupid then I apologise now!

Kind Regards,
Andy

I have hopefully inserted two images:
New Face.jpg - using the new builds
Old Face.jpg - using the old builds
Zoom in 3x to the left of the mouth extending past the nose to the eye. I think that the old builds look less blocky. It is more prominent when switching between the two in VDub. I guess something is lost in the JPEG compression.

file:///e:\temp\New Face.jpg
file:///e:\temp\Old Face.jpg

[AndyP]

Can't get the images to work so have attached them.

[AndyP]

And the second (old) one.

[Koepi]

ok, you DID some no-nos for quality:

1st, use h263 on first pass when using modulated on 2nd pass and the average quantiser will be >3 (even if it's going to be 3.01!). This can suffice to gain your old quality again.

Don't use MVhints, they degrade quality significantly - this is just an option for SPEED, not for quality (some people want just speed *shrug*)

Koepi

PS: i think it's not necessary to look at those pictures as everybody _knows_ what blockyness is, so we can save some bandwidth. I'll delete them if you don't care.

EDIT: I-range should be 2-4 for matrix, or even 2-5.
EDIT2: and keep your hands away from iframe boost, set it to 0.
maybe the consecutive iframe treatment doesn't suite your needs, you can set it to 0 to disable it, or just use 10% or something similar.

[-h]

This could be TOOSMALL_LIMIT. Did that change make it into your binaries Koepi?

-h

[Koepi]

i set it to 1, i thought of that, too :)

"Sunday, 18.08.2002
16:00h MET
A modified binary without filename change is up, I set the block-skip threshold to 1 now again."

_but_ i have the other conditions checked, too whcih sykIn suggested (not just a threshold check but

...

now damnit.
it's 3. maybe the newer CVS checkout?

I'll put up a new binary very soon.

Still the first pass quantizer should be h263 for modulated quants...

Thanks for making me check that again, -h, I absolutely trusted that it's 1 since I already changed that :-/

Regards,
Koepi

[Koepi]

XviD-29082002-2:
- Activated EPZS(^2) motion estimation algorithm.
- Fixed macro block skip-threshold. Sorry I didn't notice before
that this was changed in CVS - my "private" builds had them
disabled so i blindly trusted they were fixed in CVS, too.
- As bonbon I added a "new modulated HQ (experimental)" mode for
modulated quantizers, Isibaar suggested that: h263 quant for <4,
mpeg quant for higher quantizers. Should eleminate some noise
from the "huge" frames, and mpeg quant seems to perform better
with high quantizers vs. blockyness than h263 quant.


Man, I'm sorry pals!

[unplugged]

Originally posted by Koepi
Don't use MVhints, they degrade quality significantly
Heh (http://forum.doom9.org/showthread.php?s=&threadid=24890&perpage=20&highlight=mvhints AND more AND blocks&pagenumber=1) he he... :D, time ago I found and post about this negative behaviour, sadly almost nobody gave me the answers to my findings.

[Koepi]

Well, sometimes you have to experience it yourself on different footage to realize it.

[AndyP]

Wow! Lots of replys :) Thanks guys.

Right, will download new version and try again.

I used MPEG for first pass as (Acaila, I think) suggested you use the one that is most commonly used in the 2nd pass and for this I found that 80%ish of 2nd pass frames were MPEG. I will use H263 as I did before.

I won't use MVH, I noticed you did say that recently but I did this encode a few days ago, although I did use it in the old one with no probs. :confused: I did do it without MVH as well (forgot to mention, sorry) and it was no different.

Iframe boost was 0 for the new builds. I used the defaults.

Feel free to delete the pics. (It's just I know you don't like it when people don't give all the facts ;) ) Oh and I will use the iframe range you recommend.

Again, Many thanks for all the replys

Kind Regards
Andy

[AndyP]

Hi

Just had a look at the new encode and much better. Still not quite as good as 30062002 build but only really noticable if you look for it and not really noticable when played back at full screen with the Nic's filter set to deblock and dering.

Many Thanks

Andy

FYI Set-up as before but used H263/Modulated (old), I quants 2-4, P quants 2-8, no MVH, AltCC (and iframe settings) as default

[Koepi]

Strange, I get _better_ results.

Did you use lumi masking? Then you'd have weird artefacts with the old encoder... the newer does a proper job.

[AndyP]

Yeah, I used lumimasking on both old and new....

Andy