Measuring multipass improvements?

[stormy1777]

Is there a way to figure out if another pass actually helps the quality of the final video?

Say I run 2 passes get a 500MB avi, now run 3 passes also 500MB, now 4 and 8 passes.. all 500MB.. how can I compare other than my eyes which version is better, by how much and where the added benefit is very low..

I assume 3 passes (of course minimum 2 passes) is the optimum, but say I wanted to measure the improvment of another pass..

In GK I don't see any output in the encoder window that says what it gained from that additional pass, I'm sure a higher number is a better, but what's the limit and how do I measure it?

Thanks!

[Soulhunter]

Click the links 3-5 in my sign... ;)


Bye

[stormy1777]

ok.. thanks... lots of nice tests you post.. but we might need an 'executive summary' i.e. what do all these numbers mean, in simple words.. to be safe I go with 3 pass.. ;-)

let your PC sleep a little bit.

[Soulhunter]

Originally posted by stormy1777
but we might need an 'executive summary'...
Read link 18... ;)
Originally posted by Soulhunter
To say it with the words of Albert Einstein...

" Not everything that can be counted counts, and not everything that counts can be counted... "

So, comparing with your eyes is always the best !!!

My eyes tell me...

- IMHO 2Passes are enough (only tested in 2-3 CD scenarios)

- XviD looks anyway better than DivX (my personal preference)


My opinion:

If you cant see a difference between two methods, use the faster one... ;)


Bye

[bond]

i also once did a visual comparison, which you wuill find if you use search, which showed that 3passes make sense qualitywise in divx5 (not speedwise tough ;) )
more than 3passes definitely made no sense in the sample i tested

still my opinion is that 2passes should be enough for a good codec to find the maximum quality possible

[DigitAl56K]

Actually, PSNR is a poor way to measure the benefit of Multipass.

Remember Multipass aims to bring consistency to the quality of the video. In 1-Pass your video quality may be great in some areas, but poor in others. Think of multipass as a means to average that out in some sense, such that the quality is steady throughout the file.

When we talk about quality, the closest gauge we really have is what frame quantizer the encoder is using for each frame. Lower frame quantizers indicate the picture has been preserved more acurately (higher quality), but also lead to higher bitspend. The inverse is true for higher frame quantizers.

Create a file on your hard drive:
C:\FrameLevelControl.txt

.. then run an encoding. Look down to the end of this file, and you will see a list of how many frames have been encoded at each available quantizer. In a 1-Pass encoding these will probably be fairly distributed. After a multipass they should be far less distributed.

Lower average quantizer = Better quality.
Lesser quantizer distribution (all the frames grouped around the same quantizer) = Better quality consistency. -- This one is what to look for in Multipass.

Once you have a very low distribution, you should stop running additional passes.

HTH,
-Al

[stormy1777]

wow!, nice...

Quantiser distribution:
Q=1 0
Q=2 521
Q=3 4139
Q=4 0
Q=5 0
Q=6 0


that special debug file is not mentioned anywhere else on doom9.. now I see it's all over the divx forum... ok.. now is there a way to make it spit this Quant distribution after each pass? it seems like it dumps it at the very end.. so I have to do a full 2pass encode, look at the Q-dist.. then another full 3pass compare the Q-dist.. and maybe even 4pass just to be 100% sure.. would be nice to just run the 4pass and just dump this nice chart after each phase.. but I guess it's not possible since it still knows it's got more passes so it doesn't make any decisions yet...

[Soulhunter]

How does the number of passes affect the quantizer distribution ???

Would be another nice test... :rolleyes:


But no, first I have to play around with the XviD v.1 final !!!

Someone else maybe... :D


Bye

[DigitAl56K]

Originally posted by Soulhunter
How does the number of passes affect the quantizer distribution ?


More passes should reduce the distribution, leading to a more consistent quality.

Before anyone spends hours of their life testing the next inevitable question: No, it is not always possible to do a zillion and one passes and get 100% of all frames using the same quantizer, because of things like the VBV (equal quantizer for different frames will use a different numbers of bits).