Quantization matrices and final quality

[jonny]

I've got some testing encoding Gladiator and using different matrices.
I'd like to tell my experience, i'm courious to see if someone have done something similar.

I've done 2 encodes, 1 with the MPEG standard matrices, one pass vbr with Q=33
The second is with the original matrices of the dvd, one pass vbr with Q=70 (the final size of the 2 encodes differs of ~40MB, so the results are comparable)

Conclusion:
I really like the encode with the original matrices (and Q=70), the matrices is really "aggressive", so the Q jumps to an higher value, but the picture look more like the original, with nice grain and more detailed.

I wasn't expecting such difference and good results using an high Q, someone have experienced something like this before?

[tyee]

Jonny
How do you grab the matrices from the original DVD to use in CCE?

[Sulik]

The actual quantization for a coefficient C[i,j] is Q * Qmat[i,j],
so if you half all coefficients in the quantization matrix, you'll get the same result with double the Q parameter.

Actual differences in quality are more due to how much high frequency coefficients are quantized versus low frequencies, as well as intra vs non-intra, rather than the absolute values of Qmat and Q.

[jonny]

@tyee:

I demux a chapter with dvddecrypter, then i use restream to get the matrices (you can save both intra and nonintra in txt files).

After this you have some choices:

-copy manually the matrices to cce
-save an .ecl using a nonstandard matrices and replace the matrices inside the .ecl (you'll probably need to trim out "," and add a tab in front of each line... as you see in the .ecl file)
-merge the 2 matrices (saved with restream) in a single txt file, and load this file with my app (qcce), after this you can export the .ecl with EclCCE->Save ecl (a function that you find in the menu).


@Sulik:
thanks, a really interesting answer