CQ vs. CQ_VBR ... Very Interesting... - Page 15

kwag · #**281** 12-31-2002, 02:50 PM

Quote:

Originally Posted by SansGrip

My tests with beta-1 showed quite a serious impact on CQ level. With the old matrix I got 69.9, and with beta-1 I got 62.8. This will probably equate to more macroblocks. I also saw more Gibbs

. Are my results really that out of whack with yours?

More tests have to be done. The screenshots above were done both with the same CQ value of 62.276. The file size for both samples is almost the same. I didn't use any noise filters. Just this:

Code:

LoadPlugin("C:\encoding\MPEG2DEC.dll")
LoadPlugin("C:\encoding\fluxsmooth.dll")
LoadPlugin("C:\encoding\legalclip.dll")
LoadPlugin("C:\encoding\sampler.dll")

Mpeg2Source("K:\K19\VIDEO_TS\k19.d2v")
LegalClip()
LanczosResize(496,336,8,0,704,480)
FluxSmooth()
AddBorders(16,72,16,72)
LegalClip()

#Sampler(length=24)

-kwag

SansGrip · #**282** 12-31-2002, 03:06 PM

Quote:

Originally Posted by kwag

More tests have to be done. The screenshots above were done both with the same CQ value of 62.276. The file size for both samples is almost the same.

For my samples (528x480, no noise) the file size difference was about 2mb

. Strange.

kwag · #**283** 12-31-2002, 03:35 PM

Quote:

Originally Posted by SansGrip

For my samples (528x480, no noise) the file size difference was about 2mb

. Strange.

Maybe the material does make a difference

I'm currently encoding a sample of "The sum of all fears" with the BETA-1. I'll also encode the sample with the original matrix. I'll let you know how much difference there is.

-kwag

SansGrip · #**284** 12-31-2002, 03:44 PM

Quote:

Originally Posted by kwag

Maybe the material does make a difference

That's what I'm thinking. Resident Evil is very high-motion.

kwag · #**285** 12-31-2002, 04:16 PM

Here's what I'm doing to zero in on CQ value faster. The way I do it in CQ_VBR doesn't work correctly with CQ, and takes many many tries to hit the target. That is New_CQ_Value=(Wanted_Size/Sample_Size) * Current_CQ_Value
So this is what I do for CQ:
Run a sample session. Apply the formula above to correct CQ value. Note the new CQ value and run a second sample. Apply the formula again, and note the new CQ value correction . Then use New_CQ_Value (Last_CQ_Value + Current_CQ_Value ) / 2 ) and repeat.
This way I make less tries and hit the target faster

-kwag

SansGrip · #**286** 12-31-2002, 04:46 PM

Quote:

Originally Posted by kwag

run a second sample

...with the new, corrected value?

Quote:

Apply the formula again, and note the new CQ value correction . Then use New_CQ_Value (Last_CQ_Value + Current_CQ_Value ) / 2 ) and repeat.

I'm not sure what you mean by "new", "last" and "current" here. Could you post an example?

kwag · #**287** 12-31-2002, 04:52 PM

Quote:

Originally Posted by SansGrip

Quote:

Originally Posted by kwag

run a second sample

...with the new, corrected value?

Quote:

Apply the formula again, and note the new CQ value correction . Then use New_CQ_Value (Last_CQ_Value + Current_CQ_Value ) / 2 ) and repeat.

I'm not sure what you mean by "new", "last" and "current" here. Could you post an example?

Just run a sample. Use the formula and get a CQ correction. Store that on left brain

.
Run another sample, use formula to get new CQ value. Store that on right brain

.
Now add both values and divide by 2. That's your new CQ value which is exactly between your last two CQ values. Next runs will be half of half of half, and you should get your target in 3 or 4 sample runs max.

-kwag

SansGrip · #**288** 12-31-2002, 05:16 PM

Quote:

Originally Posted by kwag

Just run a sample. Use the formula and get a CQ correction. Store that on left brain

.
Run another sample, use formula to get new CQ value. Store that on right brain

.
Now add both values and divide by 2.

Ahhhh... I was reading what you posted as "new CQ value multipled by ..." instead of "new CQ value equals ..."

.

kwag · #**289** 12-31-2002, 05:22 PM

Quote:

Originally Posted by SansGrip

Ahhhh... I was reading what you posted as "new CQ value multipled by ..." instead of "new CQ value equals ..."

.

*/+-

black prince · #**290** 12-31-2002, 11:24 PM

@SansGrip,

I tried Crop using VirtualDub -->Video->Filters->Add->null transform
and then clicking the "Cropping button" to bring up the cropping screen.
This was even better than Tmpgenc. Then I used FitCD to resize,
get the TV-overscan and correct aspect ratio. It all worked GREAT!!!

Thanks for your help.

@Kwag,

Tried the file prediction formula for CQ and this time I achieved 0%
difference between test file and predicted file size in 4 passes.

This used to take 8 to 12 passes. Bless you Kwag

I'm going
to wait until the Q-Matrix development is stable, so I used the
standard Q-Matrix for now. Seems you and SansGrip are still testing
and changing it.

-black prince

SansGrip · #**291** 01-01-2003, 07:00 PM

Quote:

Originally Posted by black prince

I tried Crop using VirtualDub -->Video->Filters->Add->null transform and then clicking the "Cropping button" to bring up the cropping screen. This was even better than Tmpgenc. Then I used FitCD to resize, get the TV-overscan and correct aspect ratio. It all worked GREAT!!!

You're now doing it exactly how I do it

. It's pretty easy, but I still wish FitCD could determine the borders automatically

.

muaddib · #**292** 01-01-2003, 08:11 PM

Quote:

Originally Posted by SansGrip

but I still wish FitCD could determine the borders automatically

.

Yeah! If just FitCD had a crop dialog like Vdub... That would be nice.

I'll search if I can find a delphi plugin that can read a dv2 file (or just extract some frames snapshots).

SansGrip · #**293** 01-01-2003, 08:33 PM

Quote:

Originally Posted by muaddib

Yeah! If just FitCD had a crop dialog like Vdub... That would be nice.

If Delphi can call external DLLs it wouldn't be hard to write a DLL to open an Avisynth script (or AVI) and display a frame for cropping. Reading from a d2v would be harder.

muaddib · #**294** 01-01-2003, 08:44 PM

Quote:

Originally Posted by SansGrip

If Delphi can call external DLLs it wouldn't be hard to write a DLL to open an Avisynth script (or AVI) and display a frame for cropping. Reading from a d2v would be harder.

I guess delphi do can call external dlls! I just have to confirm that.
But it would be better if we can also do it from a d2v, cause that's what we usually use as the source file in FitCD.

SansGrip · #**295** 01-01-2003, 08:46 PM

Quote:

Originally Posted by muaddib

But it would be better if we can also do it from a d2v, cause that's what we usually use as the source file in FitCD.

I believe mpeg2dec can be called as a regular DLL, not just an Avisynth filter. If it can, reading d2v shouldn't be too difficult.

SansGrip · #**296** 01-01-2003, 09:40 PM

Quote:

Originally Posted by muaddib

I guess delphi do can call external dlls! I just have to confirm that.

If it can, try this with it. Everything you need should be in the zip; it's just a question of translating the C header into something Delphi can work with.

In C, I would call it like this:

Code:

CROPINFO ci;
DisplayCropDialog("the_file.avi", 1, &ci);

It doesn't matter what filename you put at the moment: it doesn't actually open any files. It just displays a message box (to let you know it worked) and then fills in the CROPINFO structure with pretend values:

ci.left = ci.right = 8
ci.top = ci.bottom = 16
ci.width = 704
ci.height = 448

So to test the DLL work out how to call it from the FitCD source then, on return, check the values in the structure match the above.

If you can get FitCD calling this DLL when a button is pressed, I'll make it actually do something

.

jorel · #**297** 01-01-2003, 10:41 PM

any solution or resolution for matrix,cq or cq vbr......???

something?

SansGrip · #**298** 01-01-2003, 11:01 PM

Quote:

Originally Posted by jorel

any solution or resolution for matrix,cq or cq vbr......???

Here's what I use, right now:

Old matrix (I find the new ones increase the file size too much

)
New GOP
For LBR, CQ_VBR without noise
For 352x240 and 352x480, CQ_VBR with noise
For 528x480 and above, CQ without noise

jorel · #**299** 01-01-2003, 11:11 PM

Quote:

Originally Posted by SansGrip

Quote:

Originally Posted by jorel

any solution or resolution for matrix,cq or cq vbr......???

Here's what I use, right now:

Old matrix (I find the new ones increase the file size too much

)
New GOP
For LBR, CQ_VBR without noise
For 352x240 and 352x480, CQ_VBR with noise
For 528x480 and above, CQ without noise

"Old matrix (I find the new ones increase the file size too much

)"
"For LBR, CQ_VBR without noise"
yes i get the same result in tests.

thank you!

kwag · #**300** 01-01-2003, 11:19 PM

Quote:

Originally Posted by SansGrip

Old matrix (I find the new ones increase the file size too much

)
New GOP

Hi SansGrip,

Even though the new matrix ( BETA-1 ) increases the file size, did you compare the quality produced to the old matrix, by adjusting the CQ to match file sizes on both encodes

At least on the tests I've made, the quality produced was slightly better on low light levels with the BETA-1 matrix. Did you notice that, or were your results different

-kwag