@SansGrip,

I tried 2 CD encodes with CQ and CQ_VBR. Flashing still remains a
problem, but Gibbs is much better :) I also, tried your method of
manual file prediction:

((CD size - Audio) / Total Frames) * test frames) = Target File Size

It's slightly more accurate, but I was wondering how you determined
for 1 CD the capacity is 813,019,155. For 2 CD's would it be correct
to multiply by 2. :?:

-black prince

@Kwag,

It looked bad on TV and PC. (flashing with 704x480 CQ). Also,
528x480 was a little better, but it's still noticalbe. Not until I used
352x240 did flashing disappear, but Gibbs seemed a little worst. :)
Seemed the higher the resolution flashing was more noticable :)

-black prince

Ok, you guys push me, so I'm going to look at this very closely :D
I'm going to go grab something to eat, and I'll be back in a couple of hours ( wife wants to go out :x ). I have something in my mind that might solve the problem.

Later :wink: ,

-kwag

With a calculator ;). Basically:

Capacity of CD in bytes = minutes * 60 * 75 * 2324

For an 80-minute CD, that would be:

80 * 60 * 75 * 2324 = 836,640,000

A good guess for VCD/SVCD filesystem overhead is 11mb, or 11,534,336. This brings us down to:

836,640,000 - 11,534,336 = 825,105,664

We then need to compensate for the system stream, which is calculated using the following formula:

Total size = (video size + audio size) * (2048 / 2018)

Therefore we do:

825,105,664 * (2018 / 2048) = 813,019,155

That's the total number of bytes we have available for our video and audio. Once we subtract the size of our audio from that, we have the maximum size for video.

160 * 60 * 75 * 2324 = 1,673,280,000
1,673,280,000 - 22 * 1024 * 1024 = 1,673,280,000 - 23,068,672 = 1,650,211,328
1,650,211,328 * (2018 / 2048) = 1,626,038,310

and

813,019,155 * 2 = 1,626,038,310

In other words, "yep" :mrgreen:.

@SansGrip,

SansGrip wrote:
Thanks :D

-black prince

@black prince

By the way, that's the formula I'm going to be using in the next release of KVCDP... It does seem to be more accurate, as you say, but it's a bit of a pain doing it manually.

@black prince, jorel, SansGrip, and All@,

Download this, and tell me if the "flashing" effect is gone or not.
8O http://www.kvcd.net/kpax-1-12-4-1-16.mpg 8O
Is the quality retained :?: ( To me, it seems the same )
Is the flashing gone :?: ( To me, yes, gone for good! )
But I need more eyes to look at it 8)
If it is, I'll explain what it means, and why ( I assume you already see what ) :wink:
Compare it to the sample I posted yesterday

-kwag

i see in powerdvd,seems......

amazing 8O
can't see any problem! 8) (perfect color cubs,skin faces..... :wink: )

see in zoomplayer too,a little dark and some blocks in dark! :?

Expect more changes tonight on the GOP. My goal is to find an optimal value to match the quality of the 1-12-1-1-24 at the same file size. This thread will be polluted by me with samples. 8O But when I put my mind on something, I "sock it to it" until I get what I want :D
So stay tunned, and compare the samples I'll start to post until we all agree that it matches the "Almost Gibbs free" kpax-newgop-newmat.mpg sample, but better stable background without flashes. This is the reference sample and size that I am focusing on right now.

-kwag

I see (in WMP) fairly strong quantization in low frequencies, yet an increase in Gibbs. No flashing though :).

Tunning, tunning..... 8)

-kwag

Making progress? :)

(I just realized... The more time you spend doing this, the less time you spend playing with GripFit 8O ;))

That’s why i LOVE this place! :lol: :flip: :lol:
Those flashes are (were) really annoying. :wink:

Take a look, and tell me SERIOUSLY what you think. If you see any :puke:, tell me. If it's :tup: , say it too :D

Original reference:
http://www.kvcd.net/kpax-newgop-newmat.mpg
New test model:
http://www.kvcd.net/kpax-1-24-4-1.24.mpg Ahh, chucks :roll: :mrgreen:

-kwag

Need some testers here :wink:

Encode a sample ( 15 second or so ) with GOP 1-12-1-1-24 and another sample of the same 15 second clip with GOP 1-24-4-1-24.
Play with CQ of second sample until size matches size of first sample. Then make a visual comparison.
Use KVCD BETA-1a Notch Matrix for both.

-kwag

@Kwag,

I created two 30 second samples. One with the GOP 1-12-1-1-24 and
the other using 1-24-4-1-24. I used file predictor to create the second
file by setting the target file size to the size of the first test file.
I agree with SansGrip, that the flashing is much improved. It appeared
to be there but much less noticable. I viewed the samples up close on
my PC and TV. Up close, meaning, within one foot. From 5+ feet,
It's not noticable. There was some Gibbs, but maybe your CQ settings
were not very high. I realize that flashing can not be completely
eliminated, but if it reduced to a point where it's not noticable that'ss
just as good. :D

-black prince

I think I learned my lesson on GOP. "If you have a lot of "Pees", you also need some to fill in with some "Bees" :tongue2:

Yes, the CQ I used is not the CQ to fit the movie on one CD. It's lower, so I could enhance the visible blocks and artifacts.

I can't see flashing anymore on that scene 8O. At least on that last sample I posted. I'll encode something else. As far as Gibbs, they're actually there, but they're "smoothed" because of the additional B frames, so they don't show as sharp as with 1-12-1-1-24.

-kwag

@Kwag,

Without realizing it, you are using psychovisual techniques. Reducing
picture effects the are very noticable and enhancing those the vision
focuses on more often. It's like listening to music and realizing the
audio range of the ear will never notice certain imperfections and
others it will pick very quickly. Knownig which visual effects are
ignored and which are very accute to sight is useful to creating
high quality video with greater compression. I think it's really interesting
and fun tweaking this process towards that goal. :D

-black prince

Yes, and we all will become Psychos soon :mrgreen: ( Just kidding :D ) You're right on target :D . The BETA-1a notch matrix has a strong effect on the low frequency domain of the material, causing a drop on visual DCT blocks and a blending effect on the blocks. Then on the high frequency end of the spectrum, the long GOP "blurrs" the artifacts for a more pleasing view. That sounds very much like some of the principles used on MP3s, where very close frequencies that the human ear can't make the difference, the lower level signal is discarded. That's how we get compression :) .I think we're very close to that goal now :wink:

-kwag