compressibilty tests with xvid matrices

[vinkes]

Hi everybody,

I made a test of all the quantization matrices collected by Nitrogen (thanx btw). For some time now I’ve been using the hvs-good matrix, and I had pretty good results. Once I saw the matrices from nitrogen, I became curious and I decided to test them all.

I tested the matrices on Band of Brother episode 5: crossroads. This episode features scenes with very high motion and some low motion scenes. I used koepi’s (thanx for all the great builds) 03052003 xvid-build with the following settings:

2pass – 1st pass
motion – 6
fourcc – xvid
vhq – 1
max I-frame – 300
min I-frame – 1
chroma motion
quarterpel
b-frames – 2/150/100/1 (with b-vop)
credits encoded at quant 18


neutral bicubic
640x352
fluxsmooth (5,7)

These were my results:

Matrix----------------------size (kb)------compressibility (%)

Bulletproof heavy compr.----------521.182---------100
Ultimate matrix-------------------581.366---------89,6
Hvs-good--------------------------738.720---------70,6
Ultra low bitrate-----------------743.046---------70,1
Hvs-better------------------------767.206---------67,9
H263------------------------------809.000---------64,4
Very low bitrate------------------823.656---------63,2
Low bitrate-----------------------844.194---------61,7
Bulletproof high quality----------855.418---------60,9
Hvs-best--------------------------856.128---------60,8
Andreas doppelte 99er-----------1.071.598---------48,6
Andreas einfache 99er-----------1.071.804---------48,6
Andreas 78er--------------------1.072.812---------48,6
Mpeg----------------------------1.085.676---------48,0

I my tests I didn’t include the cg-animation matrix. Perhaps later I’ll give some information about the visual quality of the different matrices, but I haven’t got much time at moment.

I hope the info is usefull to you guys.


Regards,

[JagPanzer]

vinkes, this is very useful. Tnx!

I hope you find some time and give us some information about the visual quality, and maybe even post some screen captures. ;)

bye

[N_F]

Your test is very useful!

I've been meaning to try the different matrices myself sometime, but time's been scarce so I haven't done it yet. I hadn't realised the difference would be so huge and I guess it's definitely something to check out.

Thanks!

[vinkes]

Alright, I did a short visual test:
- the mpeg and andreas's matrices definitely looked the best, although they do generate a lot of noise. Although I used fluxsmooth to get rid of the noise, these matrices seemed undo everything fluxsmooth did to minimize noise.
- bulletproof high quality and hvs best matrices also looked pretty damned good to me, almost the same amount of detail that the mpeg and andreas's matrices have. But these seem to have a lot less noise than these matrices. I would definitely recommend them if you had high compressibility.
- hvs good - better - ultra low bitrate - very low bitrate - low bitrate and h263: These matrices also looked very good to me. They look much like the h263. I'm most impressed by the hvs good matrix, I keeps a lot of detail and gives better compressibility than h263.
I don't know why the low bitrate matrices are called low bitrate matrices, because they don't seem to compress somewhat better than h263. But perhaps on other sources they do a much better job than with bob.
- ultra and bulletproof heavy compression matrix: they definitely lived up to their name, but the visual quality isn't very good. lots of ringing every time and frequent macroblocks in dark and very bright areas. I wouldn't recommend these matrices unless you were trying to fit SPR on to 1 cd-r :D .

I don't know, if i'll post some screenshots. As i said I'm very busy right now with school and all the b*llshit around it.

Regards

[Q-W-Y]

Great!

sticky?

[communist]

Not sticky coz this is an ancient thread using an ancient XviD build - if someone takes the time to redo this with newer (or wait for final 1.0 version) version of XviD then it certainly will be helpful :)

[sysKin]

The compressablity info is useless - this is not how you can compare matrices.

The visual info is *very* useful at the other hand, and I wouldn't say it's absolete - this kind of information does not become absolete.

[_rEuTeL_]

agreed, it's abvious that a "ultra low bitrate" matrix gives a good compressability, but I very much doubt that the detail level is comparable with let's say hvs_best

thats one long scentence!