@Kwag,

I don't know if tweaking GOP would really solve Gibbs. You sure are
making great progress with Flashing. Gibbs has been a problem
with old and new GOP's. I have an a crazy idea, what if,
SansGrip or someone re-encoded a high resolution KVCD mpg-1 using
UnDot to see if there was an improvement. Just a test clip would
satisify the question of improving Gibbs. UnDot requires avisynth
2.5 alpha. Right now there is no proactive filter that could do this,
but suppose the way is to re-encode

-black prince

Hi all

My 2 cents :

Movie : Panic Room 161075 frames, 6443 sec

old matrix : http://apoc.chez.tiscali.fr/panic-ro...ix-cq-79.8.mpg
new matrix : http://apoc.chez.tiscali.fr/panic-ro...trix-cq-79.mpg

CQ for 1 CD.

This film is pretty hard to compress : lot of DCT blocks (look at the wall)

--
apoc

@apoc,

I encoded "Panic Room" and found the DVD source was poor to
begin with Then of course, the KVCD backup looked even worst.
Dark scenes on the DVD had artifacts, blockiness, etc. I didn't
expect the backup to look better

-black prince

Probably not. I'm just trying to find the optimal point where there's no more flashing, while keeping the minimal Gibbs that we see on the 1-12-1-1-24. Other that that, there's not much we can do, as everything else is internal to the encoder.

-kwag

I've not played much with MPEG-2 short of noticing the big difference between I-frames. I did a few tests and determined that MPEG-1 is better for bitrates ~2500, but this is of course before I discovered KVCD .

Then freeze it. It's definitely better than the old one, and declaring it v2 will free up time (and mental energy ) for other things...

I don't think there would be for the simple reason that there's no Gibbs in the source, but there is in the final encode. Even if you could remove all traces of Gibbs with UnDot, when you re-encode that you'll just end up with Gibbs again.

@All,

I have a feeling that when I get the final GOP, I'm going to be doing a small tweak on the high frequency end of the matrix. I think this will be the ticket to kill the Gibbs, just like the low frequency notch killed some of the low frequency DCT blocks .

-kwag

@SansGrip,

SansGrip wrote:
Then Gibbs is a result of Tmpgenc motion search? That makes sense.
So the only way to really improve and/or remove Gibbs is to correct
Tmpgenc encoder. Hey Frankencoder, where are you!!!!!

-black prince

I don't think it's the motion search that creates Gibbs as it's usually seen in high-contrast areas like permanent subtitles etc. There probably is no way to a completely Gibbs-free encoding result, all we can do is try making it less noticable

Yep. You can see the Gibbs effect even on very high-quality JPEG files if you look closely enough. It's a consequence of the encoding method, and no amount of improvement to the encoder will get rid of it.

The best I've been able to do with the GOP without sacrificing compression and eliminating "almost all" flashing is 1-24-2-1-24.
One B, too much flashing.
Two Bs, flashing almost (if any!) gone.
Three B's flashing barely(almost static), but artifacts start to appear.
Four B's, zero flashing, artifacts start to increase.

I believe the "Happy medium" is 2 B's, with the GOP 1-XXXX-2-1-24 where xxxx is 24 or above, because last value is what dictates size of GOP. So I'll set it to the usual "5823"

Give it a try guys. There's not that much more we can do here, but I think we've done quite a lot .
I'm heading to the matrix now one more time to look at the high frequencies

Here's the latest test sample with 1-24-2-1-24: http://www.kvcd.net/1-24-2-1-24.mpg
Compare that to the old reference sample http://www.kvcd.net/kpax-newgop-newmat.mpg

-kwag

Tell me if I have this right for the current "state-of-the-art":

GOP: 1-5823-2-1-24
BETA-1a notch matrix
CQ

am I missing something?

Andy

You got it
Try it out

-kwag

Let's put the template up to for everyone to download

Only when everyone is happy and we all agree that the changes are positive . Need more testing . Feedback welcome. Please make small tests ( 15 seconds ) and compare samples with both GOPs. PLEASE, make sure that both samples are the same size before evaluation.

-kwag

Could you put it under a beta section or something, just so we can verify that everyone is using the same teplate?

Jim

Sure Jim ,
Here you go!: http://www.kvcd.net/KVCD-MASTER-BETA...-24-2-1-24.mcf

-kwag

@Kwag,

I tested 704x480, 528x480, 352x480, 352x240 using GOP 1-24-2-1-24,
Q-Matrix Notch Beta-1a, CQ 80 for all. I didn't create equal file size, since
I just wanted to see if Flashing had been effected by the new GOP. I did
however tried 528x480 CQ_VBR and I'll discuss that later. The results
were astonishing. Flashing disappeared in all the tests. Gibbs was
still noticable up close, but no worst than before with old GOP.

Now to 528x480 CQ_VBR. There was noticable Flashing. Gibbs was
slightly worst. I don't know if I want to compare CQ vs CQ_VBR
with the new GOP, since CQ would win by a mile.

-black prince

I agree. I just did a test with K-Pax at 352x240, and I'm up to CQ=100 for a one CD target (min=300,MAX=1,800) and I still can't meet the 11.68MB required . The sample size is 11.131MB . Look at K-Pax with the new matrix and GOP: http://www.kvcd.net/k-pax-352x240.mpg
I did this sample without blocks overscan, so it will look ok in WMP.

I think CQ with the latest changes is the best mode, even for 352x240
Anyone else doing more tests Feedback please Good or bad

-kwag