Some test results - Effects of matrices on filesize

[DDogg]

The results of some tests on how different matrices effect filesize. Method used is Q28 OPV encode Min 300/Max 9000
creating a 1% sample using 'SelectRangeEvery(1200,12)' in a standard script using Mpeg2dec3DG.dll. Encoder used was
CCE 2.67.00.27 set for 16:9 with progressive settings. "D-ABR" is predicted bitrate for full encode derived from
the filesize of the 1% sample. I believe these results to be accurate within a range of +/- 2.5%.



Sample D-ABR
Source 23.976 Frames = 138,965 FileSize (Kbps) Deviation From CCE
(bytes) Standard Matrix

ISpy_Sample_6of9-24_Matrix.mpv 38,429,376 5304 + 47.05%
ISpy_Sample_Mpeg_Std_Matrix.mpv 30,605,780 4224 + 17.11%
ISpy_Sample_CCE_Std_Matrix.mpv 26,133,200 3607 0.00%
ISpy_Sample_Notch_Matrix.mpv 24,118,552 3329 - 7.71%
ISpy_Sample_PackedDvd_Matrix.mpv 23,446,300 3236 - 10.28%
ISpy_Sample_Bach1_Matrix.mpv 23,371,776 3226 - 10.57%
ISpy_Sample_Jawor_1CD_LowBR.mpv 22,394,472 3091 - 14.31%
Same Matrix - Size reduction of Filtered vs. Non-Filtered
ISpy_Sample_6of9-24_Matrix_filter.mpv 29,430,004 4062 + 12.62% - 23.42%
ISpy_Sample_Mpeg_Std_Matrix_filter.mpv 23,229,660 3206 - 11.11% - 24.10%
ISpy_Sample_CCE_Std_Matrix_filter.mpv 20,351,492 2809 - 22.12% - 22.12%
ISpy_Sample_Notch_Matrix_filter.mpv 19,381,892 2675 - 25.83% - 19.64%
ISpy_Sample_PackedDvd_Matrix_Filter.mpv 18,695,840 2581 - 28.46% - 20.26%
ISpy_Sample_Bach1_Matrix_filter.mpv 18,657,748 2575 - 28.61% - 20.17%
ISpy_Sample_Jawor_1CD_LowBR_Filter.mpv 17,957,524 2479 - 31.28% - 19.81%
Note1: The 6of9 matrix is intended for use with XviD, so results when used with CCE MPEG2 may not reflect
its proper usage. I provided it as a few people had expressed a wish to try it with MPEG2.
Note2: "Filter" refers to a simple 'Undot().Deen()' combination.
Note3: I found the Jawor_1CD matrix in the XviD forum. Thread (http://forum.doom9.org/showthread.php?s=&threadid=74463&perpage=30&pagenumber=1).

[Teegedeck]

Hi,

interestingly, the HVS-best-picture matrix (talked about in that very same Javor-1CD-matrix thread) seems to be the best choice for strong compression although it most probably wouldn't score high in your test.

For quite a long time, now, I haven't tried any new matrices for low bitrates - your test makes me want to test the Bach and Javor matrices at last. Thanks! :)

[DDogg]

T, I was thinking of you when I added the 6of9 :) I'll do the HVS-best-picture matrix, and maybe a few more and add the results because I'm still set up. Point me to what you would like to see.

/Add: Looks like I can't use the HVS-best-picture matrix because it has a value of 129 which CCE will not accept.

[Boulder]

DDogg,

did you do any visual comparisons? I'd be interested in hearing your opinion as matrices are not discussed that often:)

[DDogg]

Boulder, nothing really fancy. Just scanning through the samples and looking for boogers. Since OPV varies bitrate to allow a more constant quality, video quality is not quite the issue that it is with specified bitrate based encodes. Still, some of the more severe matrices can do some nasty things. Surprisingly, since it returns one of the lowest bitrates, Jawor seems pretty good. I know your eye is better than mine. If you get time could you compare Jawor to Bach to Standard?

[Boulder]

Originally posted by DDogg
If you get time could you compare Jawor to Bach to Standard?

Hrr..responsibility:D

Sure thing, I'm not exactly sure when I'll have the time but probably the next weekend. I do mostly low-bitrate MPEG-2 encodes (avg bitrate ~1500-2500kbps) so it'll be interesting to see how it comes out.

The matrix has to be modified a bit as three values are over 99 but that shouldn't really change anything. I don't know how MPEG-4's quants differ from MPEG-2's, I thought that a value of 99 would zero the corresponding frequency completely:confused:

[DDogg]

I thought that a value of 99 would zero the corresponding frequency completely Heck, don't get me to lying :) (Texan for, "I don't have a clue"). I just pasted the Jawor thing in at the last moment.

[Boulder]

Actually it seems that the coefficient can go up to 127, so no need to change a thing:)

[manolito]

@Boulder
Did you find some time to do a comparison between Bach1 and Jawor_1? I did 2 encodes with Jawor's matrix, ultra low bitrate, and compared to the Bach1 matrix my Q dropped by about 5%. The quality looked quite good, but I do not have a direct comparison to another encode done with Bach. In Jawor's thread some people say that Jawor's matrix emphasizes grain, but I could not confirm this on my standard 28" TV.

Cheers
manolito

[Boulder]

I didn't have the time to do any extensive testing, but I compared Jawor's matrix to KVCD notch. Compressibility was bigger with Jawor's but low-lit areas showed some DCT blocks. Sharpness was nearly the same. My opinion is that Jawor's could be quite good for TV caps (more noise in the source) whereas the notch matrix would suit DVD sources (usually clean sources) better.

[DDogg]

Well I am continuing to use Bach as it does less damage to the encode than Notch, IMO. Also using Bach results in a smaller file size and better Q within a given filesize. As always, I caveat the comments with the fact that my viewing method is an older 32 inch JVC.but I compared Jawor's matrix to KVCD notch. At what bitrate? What Q? Was the bitrate matched to the source? Both encodes were done at the same Q using OPV, correct?

[Boulder]

Hey, I have a 27" 4:3 TV set :p

For my analog capture test the average bitrate was ~1500kbps, resolution 704x576. Jawor's matrix gave a lower Q in CCE, I don't remember the exact values but it was a bit lower than 70 whereas the notch matrix gave a bit over 70. Jawor's matrix reduces the noise level quite nicely as even with Avisynth filters it's very hard to get a clean result without destroying the details completely. That's the reason why I favor it for encoding the captured clips. The noise in the source will keep the DCT blocks away at those low-lit areas of the image.

For the DVD test the avg bitrate was around 2200kbps, resolution 704x576 and Q somewhere between 20 and 30. To my eyes, the notch matrix gave a slightly sharper image. Like I said, the low-lit areas looked better with the notch matrix and that's the main reason why it is my choice for DVD sources.

Have you done any tests regarding this "dancing DCT blocks in low-lit areas" with Bach's matrix?

And yes, OPV was used in both cases with a little help from QCCE (where are the new versions?)

[yaz]

@ddogg (teegy)
i've tested 16 matrices with mencoder all hvs included. they performed pretty well (as expected :-). it would worth to try them w/cce. bach1, jawor, avamat made me a hard time at higher bitrates but performed well in the lower region (<1500kbps). note, bach1 seems to be a slight mod of notch but the difference is significant as regards performance.
another funny thing. i've tried xvid default matrix too (xvid mpeg?) just for fun & it was striking good in the range 1500-2000. the same psnr as with mpeg standard but compressibility was higher (~10-15%!) give it also a try !
the bests
y