Custom matrix comparision... [Archive]

View Full Version : Custom matrix comparision...

Soulhunter

10th May 2004, 23:55

Source:

The Matrix - Lobby Shootout / 3:07 min. @ 25fps / 1024x768 pix.

Settings:

XviD v.1 RC4 (Koepis) / 2Pass @ VHQ1 / I-frame interval = 250

Results:

Matrix: H.263

@ Fixed Q 2 | PSNR = 48.60
@ 1000kbps | PSNR = 44.25
@ 2000kbps | PSNR = 46.54
@ 4000kbps | PSNR = 48.15

Matrix: MPEG

@ Fixed Q 2 | PSNR = 48.46
@ 1000kbps | PSNR = 43.96
@ 2000kbps | PSNR = 46.11
@ 4000kbps | PSNR = 48.06

Matrix: Alcailas

@ Fixed Q 2 | PSNR = 49.25
@ 1000kbps | PSNR = 42.67
@ 2000kbps | PSNR = 45.34
@ 4000kbps | PSNR = 47.20

Matrix: Andreas 78er

@ Fixed Q 2 | PSNR = 49.21
@ 1000kbps | PSNR = 43.62
@ 2000kbps | PSNR = 46.01
@ 4000kbps | PSNR = 48.02

Matrix: AVAMAT6

@ Fixed Q 2 | PSNR = 47.88
@ 1000kbps | PSNR = 44.24
@ 2000kbps | PSNR = 46.43
@ 4000kbps | PSNR = 48.08

Matrix: Bach1

@ Fixed Q 2 | PSNR = 48.27
@ 1000kbps | PSNR = 44.09
@ 2000kbps | PSNR = 46.35
@ 4000kbps | PSNR = 48.11

Matrix: Bulletproof's High Quality

@ Fixed Q 2 | PSNR = 48.20
@ 1000kbps | PSNR = 43.99
@ 2000kbps | PSNR = 46.01
@ 4000kbps | PSNR = 47.95

Matrix: Didée's SixOfNine Max=20

@ Fixed Q 2 | PSNR = 49.57
@ 1000kbps | PSNR = 43.22
@ 2000kbps | PSNR = 45.85
@ 4000kbps | PSNR = 47.76

Matrix: Didée's SixOfNine Max=24

@ Fixed Q 2 | PSNR = 49.54
@ 1000kbps | PSNR = 43.36
@ 2000kbps | PSNR = 45.73
@ 4000kbps | PSNR = 47.65

Matrix: evil v0.1

@ Fixed Q 2 | PSNR = 48.45
@ 1000kbps | PSNR = 43.80
@ 2000kbps | PSNR = 45.91
@ 4000kbps | PSNR = 47.79

Matrix: Fox Home Entertainment

@ Fixed Q 2 | PSNR = 50.22
@ 1000kbps | PSNR = 42.59
@ 2000kbps | PSNR = 45.64
@ 4000kbps | PSNR = 47.17

Matrix: HVS Best Picture

@ Fixed Q 2 | PSNR = 48.55
@ 1000kbps | PSNR = 43.84
@ 2000kbps | PSNR = 45.76
@ 4000kbps | PSNR = 47.79

Matrix: HVS Better Picture

@ Fixed Q 2 | PSNR = 48.32
@ 1000kbps | PSNR = 43.89
@ 2000kbps | PSNR = 45.93
@ 4000kbps | PSNR = 47.83

Matrix: HVS Good Picture

@ Fixed Q 2 | PSNR = 48.30
@ 1000kbps | PSNR = 43.87
@ 2000kbps | PSNR = 46.00
@ 4000kbps | PSNR = 47.83

Matrix: Jawor 1CD Matrix

@ Fixed Q 2 | PSNR = 48.10
@ 1000kbps | PSNR = 44.15
@ 2000kbps | PSNR = 46.42
@ 4000kbps | PSNR = 48.16

Matrix: Selur - high datarate

@ Fixed Q 2 | PSNR = 49.58
@ 1000kbps | PSNR = 42.62
@ 2000kbps | PSNR = 45.76
@ 4000kbps | PSNR = 47.68

Matrix: Selur's Matrix

@ Fixed Q 2 | PSNR = 49.73
@ 1000kbps | PSNR = 42.85
@ 2000kbps | PSNR = 45.87
@ 4000kbps | PSNR = 47.80

Matrix: Semi-insane

@ Fixed Q 2 | PSNR = 49.75
@ 1000kbps | PSNR = 41.84
@ 2000kbps | PSNR = 44.95
@ 4000kbps | PSNR = 46.34

Matrix: Soulhunters v3

@ Fixed Q 2 | PSNR = 48.73
@ 1000kbps | PSNR = 43.42
@ 2000kbps | PSNR = 45.47
@ 4000kbps | PSNR = 47.33

Matrix: Soulhunters v4

@ Fixed Q 2 | PSNR = 48.51
@ 1000kbps | PSNR = 43.51
@ 2000kbps | PSNR = 45.48
@ 4000kbps | PSNR = 47.46

Matrix: Soulhunters v5

@ Fixed Q 2 | PSNR = 48.99
@ 1000kbps | PSNR = 44.00
@ 2000kbps | PSNR = 46.24
@ 4000kbps | PSNR = 48.03

Note: I dont take any responsibility for this results !!!
Note: Post-processing was always disabled for the PSNR tests
Note: PSNR results are always expressed in average values
Note: The source was always upsized with Lanczos resize
Note: PSNR tests were done at bicubic fullscreen resizing
Note: The most matrices here are from LigH's zip pack
Note: All additional matrices are listed in my 2nd post

So, PSNR wise it seems like this...

Using CQM's is unless, if you stay under the H.263 Q2 limit !!!

But thats only the summary of a single test row... ;)

They could be very different with other sources n' settings !!!

To say it with the words of Albert Einstein...

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

My conclusion:

Its the best to make a new CQM for each source that matches the wanted filesize @ Q2 !!!

Comments someone...

Bye

Jawor

11th May 2004, 07:05

:thanks: for using my matrix in this comparison.

Guys, just in case you don't wanna type these 128 numbers into XviD's matrix editor, there's a link in my signature - you can download Jawor's 1CD Matrix in XviD-readable form there :D

kilg0r3

11th May 2004, 07:39

Thank you very much for your great effort Souhunter!

My conclusion, the standard matrices are tuned like hell _for PSNR_. And, we definitely could use some 'objective' measuring mechanisms, which model the human visual system. What is it that Sagitair is using?

Didée

11th May 2004, 08:26

>> Soulhunter, better known as the serial testing machine :) <<

Originally posted by kilg0r3
My conclusion, the standard matrices are tuned like hell _for PSNR_.
Seconded.

Remember, back when -h (was it him?) introduced the HVS good/better/best matrices, we all started using them because we liked the look they give. Then some PSNR tests were done, and it was figured that these HVS matrices can't stand a chance against H.263 or MPEG-standard. Well, our eyes must be bad.

For example, imagine a picture of a meadow with lots of grass, lots of flowers, etc. A scene with extreme complexity, that is.
Every matrix that does an "un-even" weigthing of the different frequencies will just die on PSNR here.
But I don't care about a little more distortion in all that grass ... simply because I cannot see it (untill I step by frame, or even do a frame diff.)

Summed up, it seems that I for one do prefer a crappy picture over a good picture:
I use always CM's, very seldom standard. I prefer qpel's look over non-qpel. I use B-frames all the time. Etc. pp. I simply use every feature that could help in giving me a lower PSNR ;)

Man, you cannot imagine how crappy my encodings are :D

- Didée

EDIT

Just discovered this, amongst others:

Matrix: Didée's SixOfNine Max=24
@ 4000kbps | PSNR = 47.65

Matrix: evil v0.1
@ 4000kbps | PSNR = 47.79

Come on, what's up here? Have I taken the blue pill, or was it the red one?

From the mere look of the evil matrix, I'm sure I'll never use it - I usually get an allergic shock from too much 255's ;)

Jawor

11th May 2004, 09:13

Originally posted by Didée
Summed up, it seems that I for one do prefer a crappy picture over a good picture:
I use always CM's, very seldom standard. I prefer qpel's look over non-qpel. I use B-frames all the time. Etc. pp. I simply use every feature that could help in giving me a lower PSNR ;)
That makes two of us :D

I've seen encodings that look quite OK even if the average quantizer is greater than 4. People, don't get confused by the magic of numbers :p

From the mere look of the evil matrix, I'm sure I'll never use it - I usually get an allergic shock from too much 255's ;)
I was shocked too when I saw first this matrix in LigH's editor, but in fact the matrix isn't that bad at all ;)

Soulhunter

11th May 2004, 18:13

Originally posted by Jawor
Guys, just in case you don't wanna type these 128 numbers into XviD's matrix editor, there's a link in my signature - you can download Jawor's 1CD Matrix in XviD-readable form there :D You could also use LigH's CQM Editor to import them from my 2nd post... ;)

Originally posted by kilg0r3
Thank you very much for your great effort Soulhunter!
No problem... :o

Originally posted by kilg0r3
What is it that Sagitair is using? IIRC, he used SSIM & VQM beside PSNR !!!

If someone gives me a working script to compare this...

Input is C:\Blah\Input.avs (4600 frames)

Output is C:\Blah\output.avs (4600 frames)

Originally posted by Didée
>> Soulhunter, better known as the serial testing machine :) << Hehe, maybe this gonna be my new custom title... :D

Btw, I will do a 2nd test-row within the next weeks !!!

Originally posted by Didée
From the mere look of the evil matrix, I'm sure I'll never use it - I usually get an allergic shock from too much 255's ;) Yeah, I was also scared after depending the results... :scared:

Nah, that only proofs that Einstein is right !!!

Bye

Selur

11th May 2004, 18:32

*gig* after seeing the evil0.1 matrix I'm wondering what psnr one would get if the whole inter matrix would be 255 and the rest would be normal h.263,..

btw. nice work soulhunter ;)

Cu Selur

CruNcher

11th May 2004, 19:48

hmm interesting the best Detail Preservation till today i allways achived with hvs better (also what the Extra_Details Profile XviD Build is useing) but seeing your Soulhunter Matrix v5 i think i have to give it a try nice work :)

SoonUDie

12th May 2004, 06:41

re: evil matrix
The evil matrix was "created" by me using DCTune on a specific frame of a specific scene of a specific movie. It suceeded in providing a higher-quality matrix than any of the others I was testing at the time, for that scene. I posted it because I wanted to see how it would stand up in other circumstances, and the results were rather bad. However, I did come up with a method that definately works better, and seems (to me) to work better than any other matrix... that method is described here (http://forum.doom9.org/showthread.php?s=&threadid=75416). Many other people have contributed, so it's turning out to be a rather intersting topic. I plan on running an extensive comparison soon (just started summer vacation here).

Selur

12th May 2004, 08:24

Hmmm,.. in principal I like the dctune idea, but iirc it can only create intra matrices. And the one thing I'm really uncertain about matrix creation is intra matrices.

So if anyone got the time it would be nice to see a comparison/some feedback/infos about how a inter matrix should look like. ;)

The main question for me in this field is:
Is it wise to use a inter matrix similiar to the intra matrix or is it 'better' to use a inter matrix with e.g. a fixed value?

Cu Selur

Sharktooth

12th May 2004, 11:32

For the lazy ppl here are the Soulhunter's and Bach1 matrices:
Soulhunter's V5 (http://ebola.gamersrevolt.it/Soulhunters%20V5.xcm)
Soulhunter's V4 (http://ebola.gamersrevolt.it/Soulhunters%20V4.xcm)
Soulhunter's V3 (http://ebola.gamersrevolt.it/Soulhunters%20V3.xcm)
Bach1 (http://ebola.gamersrevolt.it/Bach1.xcm)

Sharktooth

12th May 2004, 14:25

Ok, i've made my first custom matrix.
It's my first try so it could be "raw". If anyone is interested here it is:
8,9,10,12,14,16,20,24
9,10,11,13,15,17,22,26
10,11,12,14,17,19,24,28
11,12,14,16,19,23,28,34
12,14,16,19,22,27,32,40
14,16,18,20,24,31,40,48
16,18,20,24,28,36,46,58
18,20,22,26,30,38,52,64

16,16,17,17,18,19,21,24
16,17,17,18,19,20,24,30
17,17,17,19,20,22,26,32
17,18,19,20,22,26,32,40
18,19,20,22,26,32,40,48
19,20,22,26,32,40,50,58
20,22,24,30,38,48,60,72
22,24,27,32,40,50,70,92

As usual, for the lazy ones here's the direct link to the file: EQM V1 (http://ebola.gamersrevolt.it/eqm_v1.xcm)

Edit: Updated. See the post below.
Edit: Updated again... it should be "final" now.

Soulhunter

12th May 2004, 17:48

Originally posted by CruNcher
...but seeing your Soulhunter Matrix v5 i think i have to give it a tryThanks, feedback is always good... :)

Originally posted by Sharktooth
For the lazy ppl here are the Soulhunter's and Bach1 matrices.THX, very nice you host them !!!

Originally posted by Sharktooth
Ok, I've made my first custom matrix...
I will add it to my next tests... ;)

Bye

Soulhunter

12th May 2004, 20:16

Just a small update to get a idea how they cope with anime stuff... ;)

Source:

Animatrix (PAL/R2) - World Record / only 1000 frames @ 25fps / 720x576 pix.

Settings:

XviD v.1 RC4 (Koepis) / 2Pass @ VHQ1 / I-frame interval = 250 / Chroma motion

Results:

Matrix: H.263

@ Fixed Q 2 | PSNR = 47.38
@ 2000kbps | PSNR = 45.18

Matrix: MPEG

@ Fixed Q 2 | PSNR = 47.08
@ 2000kbps | PSNR = 45.18

Matrix: Acailas

@ Fixed Q 2 | PSNR = 47.76
@ 2000kbps | PSNR = 44.85

Matrix: Andreas 78er

@ Fixed Q 2 | PSNR = 47.25
@ 2000kbps | PSNR = 44.83

Matrix: AVAMAT6

@ Fixed Q 2 | PSNR = 45.36
@ 2000kbps | PSNR = 45.22

Matrix: Bach1

@ Fixed Q 2 | PSNR = 46.41
@ 2000kbps | PSNR = 45.18

Matrix: Bulletproof's High Quality

@ Fixed Q 2 | PSNR = 47.21
@ 2000kbps | PSNR = 44.95

Matrix: Didees SixOfNine Max=24

@ Fixed Q 2 | PSNR = 48.28
@ 2000kbps | PSNR = 44.97

Matrix: EQM v1

@ Fixed Q 2 | PSNR = 46.89
@ 2000kbps | PSNR = 44.88

Matrix: HVS Best Picture

@ Fixed Q 2 | PSNR = 46.74
@ 2000kbps | PSNR = 45.05

Matrix: HVS Better Picture

@ Fixed Q 2 | PSNR = 46.59
@ 2000kbps | PSNR = 45.03

Matrix: HVS Good Picture

@ Fixed Q 2 | PSNR = 46.48
@ 2000kbps | PSNR = 45.11

Matrix: Jawors 1CD Matrix

@ Fixed Q 2 | PSNR = 46.11
@ 2000kbps | PSNR = 45.11

Matrix: Semi-insane

@ Fixed Q 2 | PSNR = 48.99
@ 2000kbps | PSNR = 44.46

Matrix: Soulhunters v5

@ Fixed Q 2 | PSNR = 46.17
@ 2000kbps | PSNR = 45.08

Not very meaningful so far, but some "real" comparison will follow shortly !!!

Bye

Sharktooth

13th May 2004, 12:33

EQM is not intended for anime materials but... i also expect low a PSNR for movies too.
But i believe that a visual comparison says much more than PSNR.
Anyways thanks for your effort SoulHunter.
If you can test matrices at Q2 and 800, 1000, 1500, kbps it will be perfect for 1 and 2 CD and DVD encodings:)

EDIT: EQM V1 has been updated due to a nasty typo in the code i used to generate it... there are still some rounding issues. I'm working to fix them for V2.

Honza

13th May 2004, 19:26

Good work, thanks for great test, could you add filesizes with q2, if you still have files.

I tried do some tests and I thing, that Bulletproof's High Quality Matrix is good choice for fixed q2 with very good proportion of quality and size. I did´nt any PSNR measures I use human eyes only and I thing, that picture is sharp like MPEG and filesize is H263 similar.

Soulhunter

13th May 2004, 20:11

Originally posted by Honza
Good work, thanks for great test, could you add filesizes with q2, if you still have files. Have already done something like this here (http://forum.doom9.org/showthread.php?s=&threadid=73666&perpage=20&pagenumber=3)...;)

Originally posted by Soulhunter

Source:

The Matrix - Lobby Shootout / 3:07 min. @ 25fps / 1024x576 pix. (Lanczos)

XviD v.1 RC4 (Koepis) / Quantizer2 / No B-VOP's / VHQ 1

Results:

Matrix: Fox Home Entertainment / Filesize: 365.252.608 Bytes

Matrix: SemiInsane / Filesize: 338.434.048 Bytes

Matrix: Selur - High datarate / Filesize: 261.918.720 Bytes

Matrix: Selur's Matrix / Filesize: 260.757.504 Bytes

Matrix: Didees SixOfNine Max=20 / Filesize: 236.367.872 Bytes

Matrix: Didees SixOfNine Max=24 / Filesize: 235.177.984 Bytes

Matrix: Acaila's / Filesize: 228.622.336 Bytes

Matrix: Andreas 78er / Filesize: 195.287.040 Bytes

Matrix: Soulhunters (v3) / Filesize: 194.543.616 Bytes

Matrix: Soulhunters (v4) / Filesize: 160.976.896 Bytes

Matrix: HVS Best Picture / Filesize: 159.289.344 Bytes

Matrix: Standard MPEG / Filesize: 150.237.184 Bytes

Matrix: HVS Better Picture / Filesize: 140.699.648 Bytes

Matrix: HVS Good Picture / Filesize: 136.632.320 Bytes

Matrix: Standard H.263 / Filesize: 130.297.856 Bytes

Matrix: Bach1 / Filesize: 128.780.288 Bytes (update)

Matrix: Jawor 1CD Matrix / Filesize: 121.341.952 Bytes (update)

Matrix: AVAMAT6 / Filesize: 106.104.832 Bytes

Matrix: Bulletproof's Heavy Compression / Filesize: 101.152.768 Bytes

Matrix: Ultimate Matrix / Filesize: 93.820.928 Bytes

Bye

Honza

14th May 2004, 17:22

Thanks

azerty@qwaerty

14th May 2004, 21:40

1) Use Overall PSNR or SSIM for a better convergence with Human Visual Measurement ...

2) For Best PSNR (quality ...?) in quant 2 ... lol

"My Best Quality Matrix" (Very High Bitrate and Mastering quality)

8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8

8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8
8,8,8,8,8,8,8,8

"My Very Good Quality Matrix"

08,09,10,11,12,13,14,15
09,10,11,12,13,14,15,16
10,11,12,13,14,15,16,17
11,12,13,14,15,16,17,18
12,13,14,15,16,17,18,19
13,14,15,16,17,18,19,20
14,15,16,17,16,19,20,21
15,16,17,18,19,20,21,22

16,18,20,22,24,26,28,30
18,20,22,24,26,28,30,32
20,22,24,26,28,30,32,34
22,24,26,28,30,32,34,36
24,26,28,30,32,34,36,38
26,28,30,32,34,36,38,40
28,30,32,34,36,38,40,42
30,32,34,36,38,40,42,44

Soulhunter

14th May 2004, 22:46

Originally posted by azerty@qwaerty
1) Use Overall PSNR or SSIM for a better convergence with Human Visual Measurement ...Yes, I will add SSIM test soon !!!

But all this stuff is very time consuming, so you have to wait a bit... ;)

Originally posted by azerty@qwaerty
2) For Best PSNR (quality ...?) in quant 2 ... lol
Good to get a idea whats the max. quality you can archive with each matrix...

1.) 120 min. usually DVD content on 1x DVD with H.263 = Impossible

2.) 120 min. usually DVD content on 1x DVD with the MPEG = Nearly impossible

3.) 120 min. usually DVD content on 1x DVD with 6of9 = Possible

Originally posted by azerty@qwaerty
My Best Quality Matrix (Very High Bitrate and Mastering quality)

Have already tested this matrix... ;)

Nice for capturing !!!

But for compressing movies its really bad... :(

The quantizer 2 size is much to high for storage like DVD-R's !!!

And using it in 2Pass mode will mostly give bad results depended to other matrices... :p

Btw, I know this wasn't meant seriously...

Originally posted by azerty@qwaerty
My Very Good Quality Matrix
So, you want that I add it to my tests... :D

Bye

Sharktooth

15th May 2004, 11:57

Ehr...Guess what? eqm v1 updated again.

Soulhunter

20th May 2004, 22:05

Sagittaire

22nd May 2004, 12:01

Make test with animate it is not a good idea: visually you will have much difficulty to see the difference between matrix then how SSIM test could do it. 90 for SSIM is a very high value: it's impossible to make the difference visually between encoding with ~90 for SSIM ...

a small test with HPII trailer in XviD quant2 ...

andreas_78er matrix
35 586 Ko
SSIM=83.64

H263
28 516 Ko
SSIM=82.53

Make the test with a real movie ...

Soulhunter

22nd May 2004, 20:36

Originally posted by Sagittaire
Make test with animate it is not a good idea...
But how to test matrices for anime then... :confused:

Originally posted by Sagittaire
Visually you will have much difficulty to see the difference between matrix then how SSIM test could do it. 90 for SSIM is a very high value: it's impossible to make the difference visually between encoding with ~90 for SSIM ...
For me the visually difference between Fox & H.263 is more than obvious... ;)

All detail that H.263 drops is still existent in the Fox encode !!!

Ok, H.263 is one of the best for this source @ 2000kbps (visually)... :rolleyes:

But better @ 2000kbps than the Fox Home Entertainment matrix @ Q2 ???

Originally posted by Sagittaire
Make the test with a real movie ...
Its in work... ;)

Bye

Soulhunter

23rd May 2004, 17:22

Ok, I wanna start my 3rd test next week !!!

But I dont know which source I should use...

- Gladiator PAL/R2 (complete chapter)

- Matrix Reloaded PAL/R2 (complete chapter)

Also I wanna add screenshots to compare the results !!!

But this would be a huge load of data... :(

So, how much screenshots per encode would be enough for a fair comparison ???

Suggestions someone...

Bye

malkion

24th May 2004, 01:51

So, how much screenshots per encode would be enough for a fair comparison ???

perhaps like Doom9's, we can have a no-mo scene comparison, and a hi-mo scene comparison. a wall shot might be good as well. the 2 movies u've selected are just as nice, either one should be fine, yah?

Soulhunter

24th May 2004, 17:31

Think the "Burly Brawl" chapter from Reloaded would do it... :rolleyes:

- At the beginning are mainly low-motion n' closeup scenes

- The rest (fight) is mainly hi-motion then

More suggestions maybe ???

Bye

Jawor

24th May 2004, 18:57

I think The Burly Brawl is a good chapter to test how do the matrices cope with high-motion, the close-ups of Neo's and Smith's faces can be also interesting, but The Matrix Reloaded has another good chapter with close-ups - the one where commander Lock and the captains meet the council (FACAR it comes right after The Burly Brawl). You can even see the wrinkles on councilman Hamman's face there!

Soulhunter

24th May 2004, 19:08

Good, I will check this... ;)

Btw, with "the beginning closeups" I meant the stuff with Neo and Oracle !!!

Bye

Sagittaire

24th May 2004, 21:17

don't forget to indicate size or bitrate in quant 2 ... ;-)

Soulhunter

25th May 2004, 17:52

Originally posted by Sagittaire
don't forget to indicate size or bitrate in quant 2 ... ;-) Nah, its gonna be a 2Pass test like the last one... :rolleyes:

Bye

Sharktooth

26th May 2004, 10:35

I am very curious to see the screenies :)
btw i expect high bitrate matrices to give the best results @2000kbs or quant 2 (less compression = better quality).
A 1000kbps test is needed to compare low/mid/high bitrate matrices scaling.
I know it will need soooooo much time......

EDIT: looking at the results... its weird how EQM and Fox HE (being extremely different matrices) produced comparable PSNR results.

CruNcher

26th May 2004, 10:59

Blocks are more bad than ringing, but ringing is more bad than detail degration

hmm i can't agree with that ok maybe in high bitrate ringing is not as bad as blocking, but i know from tests by myself with 3 hour content on very low bitrates trying to preserve as much details as possible @ arround 550 kbps that ringing can be more disturbing then blocking. It's also easy to explain why, because their are no real (effective) deringing pp methods @ the moment at least i didn't saw any that helped here only blurring helps (detail lose), but deblocker their alot and they work quiet good and effective as the one in XviD and yes together with deringing turned on it helps on this 550 kbps but with those 2 options on, the sample looks like a RV10 movie extreme washy but it's still impressive ;).

Sharktooth

26th May 2004, 11:28

I usually try avisynth filters for playing back such lowbitrate materials.

Soulhunter

26th May 2004, 17:42

Originally posted by Sharktooth
I know it will need soooooo much time......
I wanna use two different bitrates for my test...

Estimated time... :rolleyes:

Ehm, be patient n' wait !!!

Bye

Sharktooth

26th May 2004, 17:46

if you need some help just gimme the source material and the .avs files...

Soulhunter

26th May 2004, 18:38

Originally posted by Sharktooth
if you need some help just gimme the source material and the .avs files...
Very nice offer, but I have only a 56k connection !!!

So, uploading the source would take more time than the encoding... :(

Or do you maybe have the PAL version of Matrix Reloaded ???

When yes, we have to check if its really the same... :)

Bye

Sharktooth

27th May 2004, 10:11

Well, i can rent it and just rip the chapter needed for testing.
Is it illegal?

kilg0r3

27th May 2004, 11:05

Originally posted by Sharktooth
Well, i can rent it and just rip the chapter needed for testing.
Is it illegal?
This depends on the laws in your contry and the EU. In IIRC Germany it would be illegal because it is forbidden to passby copy protection mechanisms in a way that is not presupposed by the manufacturer (viewing only). This applies if you own the DVD or not.

Sharktooth

27th May 2004, 18:43

Interesting. I will look for more informations on the italian laws.
However i should have no problems with movie trailers, should i?

Soulhunter

28th May 2004, 17:07

Originally posted by Sharktooth
Interesting. I will look for more informations on the italian laws.
However i should have no problems with movie trailers, should i? Yeah, the law... :rolleyes:

But the idea with the trailers is also good !!!

So, feel free to add your own comparision... :)

Bye

gnr

28th May 2004, 19:26

thank u very much Soulhunter for this nice comparision, found it very useful. i know a lot more comparision u like to make later. but so far what matrix would u like to recommand for 1cd rip for dvds. i mean the usual dvd movies with 80-110 min duration.

regards

Soulhunter

28th May 2004, 19:43

IMHO the H.263 n' the HVS ones are nice for low bitrates !!!

But try Jawor's 1CD Matrix, Sharktooth's EQM and my Soulhunters v5 as well... ;)

Bye

gnr

29th May 2004, 19:03

thanks for the suggestion, i will try these.

Soulhunter

1st June 2004, 19:06

So, once again... :D

Source:

Matrix Reloaded (PAL/R2) - Chapter No.3 - 3:13 min. @ 25fps

Source:

XviD v.1 Final / 2Pass @ VHQ1 / I-frame interval = 250

AviSynth:

mpeg2source("C:\Test3\Test3.d2v").LanczosResize(1024,576)

Results:

Matrix: H.263

@ Fixed Q 2 | Average PSNR = 49.96 | Average SSIM = 88.90 | Size = 65.53 MB
@ 2000kbps | Average PSNR = 49.23 | Average SSIM = 86.86
@ 4000kbps | Average PSNR = 49.96 | Average SSIM = 88.90 | Maxed out !!!

Matrix: MPEG

@ Fixed Q 2 | Average PSNR = 49.62 | Average SSIM = 88.35 | Size = 85.05 MB
@ 2000kbps | Average PSNR = 48.90 | Average SSIM = 85.65
@ 4000kbps | Average PSNR = 49.62 | Average SSIM = 88.35 | Maxed out !!!

Matrix: Acailas

@ Fixed Q 2 | Average PSNR = 50.59 | Average SSIM = 90.85 | Size = 142.60 MB
@ 2000kbps | Average PSNR = 48.42 | Average SSIM = 84.79
@ 4000kbps | Average PSNR = 49.82 | Average SSIM = 88.87

Matrix: Andreas 78er

@ Fixed Q 2 | Average PSNR = 50.48 | Average SSIM = 90.23 | Size = 123.23 MB
@ 2000kbps | Average PSNR = 48.74 | Average SSIM = 84.86
@ 4000kbps | Average PSNR = 49.95 | Average SSIM = 88.73

Matrix: Angel Best Low

@ Fixed Q 2 | Average PSNR = 49.71 | Average SSIM = 88.49 | Size = 84.30 MB
@ 2000kbps | Average PSNR = 48.91 | Average SSIM = 85.66
@ 4000kbps | Average PSNR = 49.71 | Average SSIM = 88.49 | Maxed out !!!

Matrix: AVAMAT6

@ Fixed Q 2 | Average PSNR = 49.07 | Average SSIM = 86.40 | Size = 63.05 MB
@ 2000kbps | Average PSNR = 48.77 | Average SSIM = 85.13
@ 4000kbps | Average PSNR = 49.07 | Average SSIM = 86.40 | Maxed out !!!

Matrix: Bach1

@ Fixed Q 2 | Average PSNR = 49.44 | Average SSIM = 87.61 | Size = 74.57 MB
@ 2000kbps | Average PSNR = 48.86 | Average SSIM = 85.44
@ 4000kbps | Average PSNR = 49.44 | Average SSIM = 87.61 | Maxed out !!!

Matrix: Bulletproof's Heavy Compression Matrix

@ Fixed Q 2 | Average PSNR = 48.83 | Average SSIM = 85.81 | Size = 59.13 MB
@ 2000kbps | Average PSNR = 48.62 | Average SSIM = 84.90
@ 4000kbps | Average PSNR = 48.83 | Average SSIM = 85.81 | Maxed out !!!

Matrix: Bulletproof's High Quality Matrix

@ Fixed Q 2 | Average PSNR = 49.33 | Average SSIM = 87.62 | Size = 71.40 MB
@ 2000kbps | Average PSNR = 48.91 | Average SSIM = 85.90
@ 4000kbps | Average PSNR = 49.33 | Average SSIM = 87.62 | Maxed out !!!

Matrix: Didée's SixOfNine Max=24

@ Fixed Q 2 | Average PSNR = 50.96 | Average SSIM = 91.56 | Size = 144.56 MB
@ 2000kbps | Average PSNR = 48.84 | Average SSIM = 85.52
@ 4000kbps | Average PSNR = 50.10 | Average SSIM = 89.34

Matrix: EQM v1

@ Fixed Q 2 | Average PSNR = 49.71 | Average SSIM = 88.82 | Size = 97.35 MB
@ 2000kbps | Average PSNR = 48.61 | Average SSIM = 84.97
@ 4000kbps | Average PSNR = 49.62 | Average SSIM = 88.50

Matrix: Fox Home Entertainment

@ Fixed Q 2 | Average PSNR = 52.58 | Average SSIM = 93.98 | Size = 253.78 MB
@ 2000kbps | Average PSNR = 48.54 | Average SSIM = 84.91
@ 4000kbps | Average PSNR = 50.06 | Average SSIM = 89.05

Matrix: HVS Best Picture

@ Fixed Q 2 | Average PSNR = 49.66 | Average SSIM = 88.61 | Size = 90.84 MB
@ 2000kbps | Average PSNR = 48.79 | Average SSIM = 85.49
@ 4000kbps | Average PSNR = 49.66 | Average SSIM = 88.61 | Maxed out !!!

Matrix: HVS Better Picture

@ Fixed Q 2 | Average PSNR = 49.39 | Average SSIM = 87.74 | Size = 71.40 MB
@ 2000kbps | Average PSNR = 48.89 | Average SSIM = 85.90
@ 4000kbps | Average PSNR = 49.39 | Average SSIM = 87.74 | Maxed out !!!

Matrix: HVS Good Picture

@ Fixed Q 2 | Average PSNR = 49.35 | Average SSIM = 87.61 | Size = 69.49 MB
@ 2000kbps | Average PSNR = 48.85 | Average SSIM = 85.77
@ 4000kbps | Average PSNR = 49.35 | Average SSIM = 87.61 | Maxed out !!!

Matrix: Jawors 1CD Matrix

@ Fixed Q 2 | Average PSNR = 49.28 | Average SSIM = 87.21 | Size = 67.66 MB
@ 2000kbps | Average PSNR = 48.89 | Average SSIM = 85.65
@ 4000kbps | Average PSNR = 49.28 | Average SSIM = 87.21 | Maxed out !!!

Matrix: Selur's high datarate

@ Fixed Q 2 | Average PSNR = 51.10 | Average SSIM = 91.80 | Size = 166.62 MB
@ 2000kbps | Average PSNR = 48.67 | Average SSIM = 84.96
@ 4000kbps | Average PSNR = 49.90 | Average SSIM = 88.92

Matrix: Selur's Matrix

@ Fixed Q 2 | Average PSNR = 51.30 | Average SSIM = 92.05 | Size = 167.81 MB
@ 2000kbps | Average PSNR = 48.77 | Average SSIM = 85.22
@ 4000kbps | Average PSNR = 50.04 | Average SSIM = 89.14

Matrix: Semi-insane

@ Fixed Q 2 | Average PSNR = 51.62 | Average SSIM = 92.95 | Size = 228.23 MB
@ 2000kbps | Average PSNR = 47.99 | Average SSIM = 84.69
@ 4000kbps | Average PSNR = 49.67 | Average SSIM = 88.77

Matrix: Soulhunters v3

@ Fixed Q 2 | Average PSNR = 49.83 | Average SSIM = 89.17 | Size = 105.86 MB
@ 2000kbps | Average PSNR = 48.59 | Average SSIM = 84.88
@ 4000kbps | Average PSNR = 49.62 | Average SSIM = 88.44

Matrix: Soulhunters v5

@ Fixed Q 2 | Average PSNR = 49.39 | Average SSIM = 87.53 | Size = 75.13 MB
@ 2000kbps | Average PSNR = 48.78 | Average SSIM = 85.25
@ 4000kbps | Average PSNR = 49.39 | Average SSIM = 87.53 | Maxed out !!!

Matrices sorted by Q2 size

Fox Home Entertainment = 253.78 MB

Semi-insane = 228.23 MB

Selur's Matrix = 167.81 MB

Selur's high datarate = 166.62 MB

Didée's SixOfNine Max 24 = 144.56 MB

Acailas = 142.60 MB

Andreas 78er = 123.23 MB

Soulhunters v3 = 105.86 MB

EQM v1 = 97.35 MB

HVS Best Picture = 90.84 MB

MPEG = 85.05 MB

Angel Best Low = 84.30 MB

Soulhunters v5 = 75.13 MB

Bach1 = 74.57 MB

Bulletproof's High Quality Matrix = 71.40 MB

HVS Better Picture = 71.40 MB

HVS Good Picture = 69.49 MB

Jawors 1CD Matrix = 67.66 MB

H.263 = 65.53 MB

AVAMAT6 = 63.05 MB

Bulletproof's Heavy Compression Matrix = 59.13 MB

Matrices sorted by PSNR results @ 2000kbps

H.263 = 49.23

Bulletproof's High Quality Matrix = 48.91

Angel Best Low = 48.91

MPEG = 48.90

HVS Better Picture = 48.89

Jawors 1CD Matrix = 48.89

Bach1 = 48.86

HVS Good Picture = 48.85

Didée's SixOfNine Max24 = 48.84

HVS Best Picture = 48.79

Soulhunters v5 = 48.78

Selur's Matrix = 48.77

AVAMAT6 = 48.77

Andreas 78er = 48.74

Selur's high datarate = 48.67

Bulletproof's Heavy Compression Matrix = 48.62

EQM v1 = 48.61

Soulhunters v3 = 48.59

Fox Home Entertainment = 48.54

Acailas = 48.42

Semi-insane = 47.99

Matrices sorted by PSNR results @ 4000kbps

Didée's SixOfNine Max24 = 50.10

Fox Home Entertainment = 50.06

Selur's Matrix = 50.04

H.263 = 49.96

Andreas 78er = 49.95

Selur's high datarate = 49.90

Acailas = 49.82

Angel Best Low = 49.71

Semi-insane = 49.67

HVS Best Picture = 49.66

Soulhunters v3 = 49.62

EQM v1 = 49.62

MPEG = 49.62

Bach1 = 49.44

Soulhunters v5 = 49.39

HVS Better Picture = 49.39

HVS Good Picture = 49.35

Bulletproof's High Quality Matrix = 49.33

Jawors 1CD Matrix = 49.28

AVAMAT6 = 49.07

Bulletproof's Heavy Compression Matrix = 48.83

Matrices sorted by SSIM results @ 2000kbps

H.263 = 86.86

HVS Better Picture = 85.90

Bulletproof's High Quality Matrix = 85.90

HVS Good Picture = 85.77

Angel Best Low = 85.66

Jawors 1CD Matrix = 85.65

MPEG = 85.65

Didée's SixOfNine Max24 = 85.52

HVS Best Picture = 85.49

Bach1 = 85.44

Soulhunters v5 = 85.25

Selur's Matrix = 85.22

AVAMAT6 = 85.13

EQM v1 = 84.97

Selur's high datarate = 84.96

Fox Home Entertainment = 84.91

Bulletproof's Heavy Compression Matrix = 84.90

Soulhunters v3 = 84.88

Andreas 78er = 84.86

Acailas = 84.79

Semi-insane = 84.69

Matrices sorted by SSIM results @ 4000kbps

Didée's SixOfNine Max24 = 89.34

Selur's Matrix = 89.14

Fox Home Entertainment = 89.05

Selur's high datarate = 88.92

H.263 = 88.90

Acailas = 88.87

Semi-insane = 88.77

Andreas 78er = 88.73

HVS Best Picture = 88.61

EQM v1 = 88.50

Angel Best Low = 88.49

Soulhunters v3 = 88.44

MPEG = 88.35

HVS Better Picture = 87.74

Bulletproof's High Quality Matrix = 87.62

HVS Good Picture = 87.61

Bach1 = 87.61

Soulhunters v5 = 87.53

Jawors 1CD Matrix = 87.21

AVAMAT6 = 86.40

Bulletproof's Heavy Compression Matrix = 85.81

Sample pictures of the 2000kbps encodes:

- Original (http://img7.imageshack.us/img7/6438/OrigB_1380.jpg)

- H.263 (http://img7.imageshack.us/img7/1039/H263_2000_1380.jpg)

- MPEG (http://img7.imageshack.us/img7/8643/MPEG_2000_1380.jpg)

- Acailas (http://img20.imageshack.us/img20/390/Acailas_2000_1380.jpg)

- Andreas 78er (http://img20.imageshack.us/img20/9459/Andreas78er_2000_1380.jpg)

- Angel BestLow (http://img20.imageshack.us/img20/5536/AngelBestLow_2000_1380.jpg)

- AVAMAT6 (http://img20.imageshack.us/img20/4001/AVAMAT6_2000_1380.jpg)

- Bach1 (http://img20.imageshack.us/img20/8339/Bach1_2000_1380.jpg)

- Bulletproof's Heavy Compression (http://img7.imageshack.us/img7/3710/BulletproofsHeavyCompressionMatrix_2000_1380.jpg)

- Bulletproof's High Quality (http://img7.imageshack.us/img7/4779/BulletproofsHighQualityMatrix_2000_1380.jpg)

- SixOfNine (http://img7.imageshack.us/img7/6456/DideesSixOfNine_2000_1380.jpg)

- EQM (http://img7.imageshack.us/img7/1893/eqm_2000_1380.jpg)

- Fox Home Entertainment (http://img7.imageshack.us/img7/6460/FoxHome_2000_1380.jpg)

- HVS Best (http://img7.imageshack.us/img7/1537/HVSBest_2000_1380.jpg)

- HVS Better (http://img7.imageshack.us/img7/5048/HVSBetter_2000_1380.jpg)

- HVS Good (http://img7.imageshack.us/img7/8491/HVSGood_2000_1380.jpg)

- Jawors 1CD (http://img7.imageshack.us/img7/3548/Jawor_1CD_2000_1380.jpg)

- Selur - high datarate (http://img1.imageshack.us/img1/2830/Selur-highdatarate_2000_1380.jpg)

- Selur's Matrix (http://img1.imageshack.us/img1/823/SelursMatrix_2000_1380.jpg)

- Semi-insane (http://img17.imageshack.us/img17/7038/Semi-insane_2000_1380.jpg)

- Soulhunters v3 (http://img17.imageshack.us/img17/5728/Soulhuntersv3_2000_1380.jpg)

- Soulhunters v5 (http://img17.imageshack.us/img17/2316/Soulhuntersv5_2000_1380.jpg)

More pictures of the 2000kbps encodes:

- Original (http://img7.imageshack.us/img7/7646/OrigB_2020.jpg)

- H.263 (http://img17.imageshack.us/img17/8741/H263_2000_2020.jpg)

- MPEG (http://img17.imageshack.us/img17/8713/MPEG_2000_2020.jpg)

- Acailas (http://img17.imageshack.us/img17/8359/Acailas_2000_2020.jpg)

- Andreas 78er (http://img17.imageshack.us/img17/4383/Andreas78er_2000_2020.jpg)

- Angel BestLow (http://img17.imageshack.us/img17/8279/AngelBestLow_2000_2020.jpg)

- AVAMAT6 (http://img17.imageshack.us/img17/6566/AVAMAT6_2000_2020.jpg)

- Bach1 (http://img17.imageshack.us/img17/3908/Bach1_2000_2020.jpg)

- Bulletproof's Heavy Compression (http://img17.imageshack.us/img17/9965/BulletproofsHeavyCompressionMatrix_2000_2020.jpg)

- Bulletproof's High Quality (http://img7.imageshack.us/img7/3114/BulletproofsHighQualityMatrix_2000_2020.jpg)

- SixOfNine (http://img10.imageshack.us/img10/9572/DideesSixOfNine_2000_2020.jpg)

- EQM (http://img10.imageshack.us/img10/4720/eqm_2000_2020.jpg)

- Fox Home Entertainment (http://img10.imageshack.us/img10/9752/FoxHome_2000_2020.jpg)

- HVS Best (http://img10.imageshack.us/img10/8119/HVSBest_2000_2020.jpg)

- HVS Better (http://img10.imageshack.us/img10/4607/HVSBetter_2000_2020.jpg)

- HVS Good (http://img10.imageshack.us/img10/9489/HVSGood_2000_2020.jpg)

- Jawors 1CD (http://img10.imageshack.us/img10/1382/Jawor_1CD_2000_2020.jpg)

- Selur - high datarate (http://img10.imageshack.us/img10/2039/Selur-highdatarate_2000_2020.jpg)

- Selur's Matrix (http://img20.imageshack.us/img20/9931/SelursMatrix_2000_2020.jpg)

- Semi-insane (http://img20.imageshack.us/img20/898/Semi-insane_2000_2020.jpg)

- Soulhunters v3 (http://img20.imageshack.us/img20/760/Soulhuntersv3_2000_2020.jpg)

- Soulhunters v5 (http://img20.imageshack.us/img20/683/Soulhuntersv5_2000_2020.jpg)

Instead of the usually notes, just a big thanks to...

- LigH for his great CQM editor

- Virus, for his nice stats-log hack

- Sharktooth, for hosting my CQM's

- The XviD devs, for this fantastic codec

- All other people, just for reading this stuff

If you still have questions, feel free to ask me... ;)

Bye

CruNcher

1st June 2004, 20:31

550 and 750 kbps tests please :D
btw nice work but not really anything that suprises ;)

Soulhunter

1st June 2004, 20:43

Originally posted by CruNcher
550 and 750 kbps tests please :D
No, but maybe a 1000kbps test... :rolleyes:

Bye

Alxemi

1st June 2004, 21:03

Wou! Thanks for this incredible test!!
Yeah, 750/1000 kbits test could be usefull, it´s the bitrate most commonly used imo
Also, what do you think about posting .png screenshots? jpg is always loosy even in highest quality.
great job soulhunter!

lordadmira

2nd June 2004, 06:48

I don't know about u but clicking on those links trying to compare outputs is really annoying. So I wrote a quick DHTML app to make it easy to compare them. Just click on the matrix name. :D

http://w3.goodnews.net/~wagnerc/compare.htm

Did u know the "original" pic is different fromt he other ones? These are the 2000 kbps ones.

LA

cypher_soundz

2nd June 2004, 07:25

yeah i noticed this too, both original frames are out by a few frames.
Great job though, can i put in anothe vote for a lower bit rate?:cool:
Regards
cyph

sysKin

2nd June 2004, 07:36

Originally posted by lordadmira
I don't know about u but clicking on those links trying to compare outputs is really annoying. Huh? Open them all in separate tabs and then just tap ctrl-1, ctrl-2, ctrl-3 to switch to any picture.

ctrl-x is for mozilla/firefox, I don't know opera or konqueror or safari shortcuts, but I guess they are similar. (have I missed any worthwhile browser?).

Radek

virus

2nd June 2004, 07:54

Originally posted by sysKin
(have I missed any worthwhile browser?).No. :D

@all:
why going lower than 2000 kbps? Soulhunter is encoding at 1024x576! I guess you guys use 750-1000 kbps on a 640x272 resolution or something similar... ;)

kilg0r3

2nd June 2004, 08:03

Whew! Thank You!

What is this eqm matrix? Although it comes in very 'late' in the SSIM stats the screen shot looks promising. Couldn't find it via the search though.

lordadmira

2nd June 2004, 09:28

Originally posted by sysKin
Huh? Open them all in separate tabs and then just tap ctrl-1, ctrl-2, ctrl-3 to switch to any picture. All usm folks out heer in thu sticks use Netscape verzun fo'. :D

lordadmira

2nd June 2004, 09:39

Looking at the 2000 kbps screen caps, I don't see much difference between them. The bitrate's probly not low enough to really stress the matrices.

Sharktooth

2nd June 2004, 11:05

Originally posted by kilg0r3
Whew! Thank You!

What is this eqm matrix? Although it comes in very 'late' in the SSIM stats the screen shot looks promising. Couldn't find it via the search though.
It is on the first page of this thread.It should be a good matrix for high AND low bitrates.
Yes it looks promising and i'm still working on a new one, based on the same ideas.

Soulhunter

2nd June 2004, 17:41

Originally posted by lordadmira
Did u know the "original" pic is different from he other ones? These are the 2000 kbps ones.
Damn, you are right... :eek:

Thanks for the hint !!!

Should be fixed now... :o

Originally posted by lordadmira
I don't know about u but clicking on those links trying to compare outputs is really annoying. So I wrote a quick DHTML app to make it easy to compare them. Just click on the matrix name. :D

http://w3.goodnews.net/~wagnerc/compare.htm
Wow, really nice... :)

Could you maybe update the fixed "original" links ???

Originally posted by Alxemi
JPG is always lossy even in highest quality.
Thought about this... :rolleyes:

But to tell the truth, I cant see a real difference to a Q=100% JPEG !!!

Originally posted by cypher_soundz
Great job though, can i put in another vote for a lower bit rate?:cool: No problem, maybe the next week(s) !!!

Originally posted by lordadmira
Looking at the 2000 kbps screen caps, I don't see much difference between them. The bitrates probably not low enough to really stress the matrices. But I can see... :D

Nah, this bitrate was a compromise to give both, high n' low bitrate matrices a fair chance !!!

But a low bitrate test is the next thing on my shedule... ;)

Originally posted by Sharktooth
It is on the first page of this thread.It should be a good matrix for high AND low bitrates.
Yeah, also when the PSNR/SSIM results tell something different...

Your EQM gave me a very good visual impression @ this source !!!

Originally posted by Sharktooth
Yes it looks promising and I'm still working on a new one, based on the same ideas. Hehe, Im still working on Soulhunters V6... :p

Bye

Sagittaire

2nd June 2004, 19:13

and your visual impression ???

Make a blind test ... ;-)

Teegedeck

2nd June 2004, 19:20

I strongly support that! :rolleyes:

...but seriously, a blindtest with so many samples would probably strain even the best pair of eyes. A short impression of yours would be nice though, for those who haven't had the time, yet, to compare the snapshots themselves (like me).

Soulhunter

2nd June 2004, 19:30

Originally posted by Sagittaire
And your visual impression ???
Hmm, I already considered to post my visual impressions here...

But IMHO its better to let the ppl decide for their self !!!

Think, some people could simply adopt my impression instead comparing the pics... :rolleyes:

Dont wanna put my words in other ppl's mouth !!!

Originally posted by Sagittaire
Make a blind test ... ;-)
Would be very interesting, but also very time consuming... :(

I will think about this !!!

Bye

Teegedeck

2nd June 2004, 19:46

OK, first impressions on your close-up snapshot; I've left out quite many of the samples but those I've had a look at: Not surprisingly, low-bitrate matrices are distinguishable (Jawor's, Bach1, HVS-best) - but I think MPEG standard looks remarkably good on this shot. Andreas 78er also smoothes too much for my liking. I wouldn't dare choosing between the other candidates here (Acaila's, Didée's 6o9, Selur's hdr, Didée's semi-insane, FoxHome), quite un-ABXable for me... ;) (At least for as long as my GF blocks our 'better' display.)

Edit: On the 'fog scene', not surprisingly the smoothing matrices do a good job. This includes the 78er and all low bitrate matrices with the exception of HVS-best. FoxHome, Acaila's hdr and the semi-insane matrix exhibit too many blocks for my liking.

But again, when watched in motion the video might look a whole lot different...

Soulhunter

2nd June 2004, 20:14

Hmm, I think I have to note something you cant see @ still pictures !!!

Fox Home Entertainment, Semi-insane and Acailas...

This 3 matrices produced really bad "floating" in some scenes !!!

Bye

Alxemi

2nd June 2004, 20:36

But to tell the truth, I cant see a real difference to a Q=100% JPEG !!!
just make some zooms and compare an uncompresed bmp with a Q100%jpg and then do the same with png. png is identical at pixel-level and jpg isn´t, at least with my image processing software :P

Soulhunter

2nd June 2004, 21:13

Originally posted by Alxemi
just make some zooms and compare an uncompresed bmp with a Q100%jpg...
Nah, meant usually watching here !!!

So, normal size and at least 50cm away from the monitor... :p

Bye

lordadmira

2nd June 2004, 21:35

Updated.

Sharktooth

3rd June 2004, 12:26

Originally posted by Soulhunter
Your EQM gave me a very good visual impression @ this source !!!
EQM V1 approach is to provide a sharp image (as the MPEG standard matrix if not better) using low coefficients for the "visible" frequencies, good compression (as the MPEG standard matrix if not better) using high coefficients for removing less visible frequencies and unwanted noise reduction (that's not film grain) using a unique coefficients scaling and distribuition.
Visual quality should be really good but since the matrix is removing details (where the human eye can't see them) and removing noise, all the tests (PSNR, SSIM, etc.) will end up in "not so high" results.
But as i always say: try it for yourself and look at it with your eyes.

Lefungus

3rd June 2004, 16:26

Originally posted by Sharktooth
since the matrix is removing details (where the human eye can't see them) and removing noise

Do you have anything to backup this claim ?
This custom matrix trend is getting annoying when nobody can prove anything about their "subjective superiority".

And those comparisons are also biased, as:

- people don't watch the source in motion. Video != Still Picture
- people knows which matrices are associated to each picture/video sample. Full placebo effect, like: "it's my matrix so it's good" or "H263 is bad, let's watch this sample, YEAH ! i knew it"

When somebody tells me a matrix is HVS optimized, i want facts, not just the usual top-left->low quantizers, bottom-right->high quantizers with a different random scaling. Because this is not enough.

no offense, but i'll take you as an example:

Originally posted by Sharktooth
using high coefficients for removing less visible frequencies and unwanted noise reduction (that's not film grain) using a unique coefficients scaling and distribuition.

Could you provide some facts on how you determined which scaling has to be applied ? How do you know how to remove "unvisible" frequencies ? How are you sure it's unwanted noise you remove and not film grain ?

I expect most answers to be, i tried random numbers, or i picked a scaling, and i tried, and i tried, and i tried until results were 'ok' perceptually. And voilà, a new HVS optimized Matrix!

Now let's grow a bit, and let's stop this nonsense.
The only solution i see is to use double blind tests like they do for audio codecs. Until then, all claims are unverified, misleading and false most of the time.

Soulhunter

3rd June 2004, 18:09

Originally posted by lordadmira
Updated.
Thanks !!!

Originally posted by Lefungus
The only solution i see is to use double blind tests like they do for audio codecs.Again, I will think about this !!!

Bye

CruNcher

3rd June 2004, 18:16

@ Lefungus
:goodpost:

Sharktooth

3rd June 2004, 18:48

OMG... calm down BOY.
I use a proprietary algorithm to base coefficients scaling on a spatial frequency analysis of a variety of images.
Since it is in an early stage of development i wont give any detail.
Hope it is enaugh to back up.

Manao

3rd June 2004, 19:05

Sharktooth : did you read that thread (http://forum.doom9.org/showthread.php?s=&threadid=75416). If you didn't, you may find it interesting. There, I'm raising the issue that it is completely irrelevant to tune an inter DCT matrices from computations made on real frames, because such matrices are not applied on real frames but on the result of the motion compensation.

Anyway, I totally agree with Lefungus on the need of a double blind test.

Soulhunter

3rd June 2004, 19:21

Originally posted by Manao
Anyway, I totally agree with Lefungus on the need of a double blind test.
And another vote to make a double blind test... :D

Ok, before I start thinking, I wanna hear some suggestions ???

What would be the best way to do such a test...

Bye

Teegedeck

3rd June 2004, 20:16

Now I've checked on the 'good' display and could verify that Didée's SixOfNine-HVS, EQM, Soulhunters v3 and Bulletproof's HQ are practically transparent when compared to the original.

BTW, on Didée's SixOfNine '24'; I've been a bit confused what this thing was - until Didée enlightened me in another thread that it actually is his SixOfNine-HVS!

CruNcher

3rd June 2004, 20:31

@Teegedeck

Sorry but i watched on my TFT and for me Hvs-Better is achiving better results then EQM for example wich doesn't preserve the whole noise it blurs in between. And you also can see that in detailfull areas that details get blurred

Teegedeck

3rd June 2004, 21:07

Which sample are you talking about? I meant the close-up of agent xyz, sorry for forgetting to mention it.

CruNcher

3rd June 2004, 21:16

yes i mean that jpg shoots
but maybe also the jpg compression is doing it bad here soulhunter should definatly used png for this :(

Teegedeck

3rd June 2004, 21:29

I think PNGs would have been overkill here. At 90% quality there is not artifacting or anything mentionable IMHO. Don't know what exactly Soulhunter's used, but I guess very low compression.

As for HVS-better, I haven't even bothered looking at it because I've always found it too blurry. But as soon as I have some time (weekend...) I'll do.

CruNcher

3rd June 2004, 21:40

@Teegedeck
if you have good eyes and look really carefully you'll see that the Noise pattern is nearer with HVS-better on the Original than it is with EQM. You can see that especialy well on the white stripe (left side,background,door) the stripe is much clearer with Hvs-Better as it is with EQM, the color merges into each other with EQM and doesn't stay as a dot pattern. Btw im talking about the 2000kbps pics and i could also show you areas in the Face which loseing details because of this effect with EQM were Hvs-better is preserving it imho i think their is near to nothing that can be improved as refdivx did a great job with his code :) but sure you right if you say they are transparent, because nobody will ever see so much detail from some feet away but it goes about the mathematical preserve of details here, else we wouldn't need this matrix discussions anymore :p

Sharktooth

4th June 2004, 09:36

Originally posted by Manao
Sharktooth : did you read that thread (http://forum.doom9.org/showthread.php?s=&threadid=75416). If you didn't, you may find it interesting. There, I'm raising the issue that it is completely irrelevant to tune an inter DCT matrices from computations made on real frames, because such matrices are not applied on real frames but on the result of the motion compensation.

Anyway, I totally agree with Lefungus on the need of a double blind test.
Yes i've read it and i think i'm doing things in the right way... :)
Blind test is a must.

Sharktooth

4th June 2004, 09:53

Originally posted by CruNcher
@Teegedeck
if you have good eyes and look really carefully you'll see that the Noise pattern is nearer with HVS-better on the Original than it is with EQM. You can see that especialy well on the white stripe (left side,background,door) the stripe is much clearer with Hvs-Better as it is with EQM, the color merges into each other with EQM and doesn't stay as a dot pattern. Btw im talking about the 2000kbps pics and i could also show you areas in the Face which loseing details because of this effect with EQM were Hvs-better is preserving it imho i think their is near to nothing that can be improved as refdivx did a great job with his code :) but sure you right if you say they are transparent, because nobody will ever see so much detail from some feet away but it goes about the mathematical preserve of details here, else we wouldn't need this matrix discussions anymore :p
Those are the "invisible" details i talked about on my previous posts.
During playback they're "invisible" :)

lordadmira

4th June 2004, 12:28

Originally posted by Soulhunter
And another vote to make a double blind test... What would be the best way to do such a test... I'll look into making a DHTML app.

Soulhunter

4th June 2004, 17:53

Originally posted by Teegedeck
Don't know what exactly Soulhunter's used, but I guess very low compression.The lowest one... Q = 100 % !!!

Originally posted by lordadmira
I'll look into making a DHTML app.
Very nice, thanks... :)

Bye

*.mp4 guy

8th June 2004, 11:18

i made a new matrix, its very similar to hvs but with some changes that reduce noise at lower bitrates, in my opinion that is.

well i hope you will consider including it:D

HVS with less noise:

8,15,15,16,17,18,21,24
15,15,16,16,17,19,22,25
15,16,17,18,20,22,25,29
16,16,18,21,24,27,31,36
17,17,20,24,30,34,40,46
18,19,22,27,34,43,53,64
21,22,25,31,40,52,68,85
24,25,29,36,46,63,85,110

8,14,15,17,20,22,25,30
14,14,17,19,21,24,28,33
15,17,20,22,25,28,33,38
17,18,22,25,27,34,39,47
19,21,24,27,36,41,50,57
22,23,25,33,41,52,62,74
25,25,29,39,49,61,77,100
26,29,38,46,57,74,100,126

I think it is better but i need some confirmation :confused:

aketon

8th June 2004, 12:06

Is the HVS-GOOD matrix better than H.263??? I am making 1 CD encodes!

Manao

8th June 2004, 12:13

aketon : we don't have your eyes, so we can't judge for you.

It's however widely agreed that HVS_Good gives more pleasing results than H263 for one CD encodes, and that HVS_Best is even better.

But, for example, those matrices reduce the PSNR, so if you had got cybernetic eyes, you should use H263.

You should also have made a search on the subject before asking.

aketon

8th June 2004, 13:34

Thanks for answering! I did search about this, but I didn't found nowhere something saying "HVS-Good is better than H.263" (or something similar)! The only things I found is comparisons about PSNR and some other info! That is why I came here and asked, as I couldn't find something to help me!

Soulhunter

8th June 2004, 18:15

Originally posted by *.mp4 guy
Well i hope you will consider including it
If you think its ready to join, I will add it to my test... :)

But unfortunately Im a bit busy atm !!!

So, if you think there is still something to tweak @ your matrix, just do it... ;)

I estimate to start the next comparison in 1-2 week(s) !!!

Bye

*.mp4 guy

8th June 2004, 18:35

Thanks for letting me enter it in the test, i will continue to update the matrix in my above post until i find the one i am most satisfied with;)

... i think my computer will be doing some overtime.

Soulhunter

8th June 2004, 19:00

No problem !!!

If I start earlier than estimated, I simply send you a pm... ;)

Bye

Sharktooth

25th June 2004, 17:12

And here we go again: EQM V2 matrix (http://ebola.gamersrevolt.it/eqm_v2.xcm)
It's the result of the fine tuning of the ideas behind the EQM V1 matrix (http://ebola.gamersrevolt.it/eqm_v1.xcm).
It compresses a bit less than V1 but visual quality should be much better (not that V1 looks bad...).
A reminder: Both matrixes are standalones compatible.

Sagittaire

26th June 2004, 23:42

In conclusion HVS-Better is a good choice for Metric, Visual and Compressibility compromise ... ???

Alxemi

27th June 2004, 01:22

It compresses a bit less than V1 but visual quality should be much better (not that V1 looks bad...).
Humm maybe it´s what i was looking for my 2cds jobs..
In 1cd (600/1000) i always move between HVS-better, jawor1cd... but i always feel confused when choosing a matrix for standar 2cd rip (900/1400) What do you guys like to use with that bitrates? (normal movie)??

malkion

27th June 2004, 02:59

CruNcher

27th June 2004, 10:07

@Sagittaire
The best comment i ever saw from you congratulations i totaly agree ;)

guada

27th June 2004, 11:47

Hello cruNcher :)

Very very interesting your answer.. yes, Sarge :p
To soon ;)

Alxemi

27th June 2004, 12:28

Alxemi, maybe you'll like a HVS-best picture matrix for your 2 cd encodes. Sometimes I just stick with MPEG-standard for 2 cd's. Basically any inter matrix starting with coefficient 16 or higher will do if you like the matrix that is...
It´s was i usually use for 2cd... but every test shows that all the HVS are very similar, and hvs best is not always the best :)
And the difference of bitrate usually between 1cd and 2cd is big.
MPEG-Standar always gave me very bad results, mosquito etc..

guada

27th June 2004, 13:52

hello Alexemi,

I'm so surprised, but it's your own opinion.
however, could you tell the name of movie that you would have encoded( don't need of your script ).

Thank you very much. :)

Alxemi

27th June 2004, 14:14

Hum.. i remember some movies that y rip some years ago with MPEG standar, were 2cd rip, one was excalibur and the other... hell i don´t remember, but i remember that my bitrate was more than 1K and the MPEG matrix created mosquito noise that was not in the source.
Anyway i was very newbie (just like today xD) and maybe messed up the configuration. It was early XviD too.

Soulhunter

27th June 2004, 17:39

Originally posted by Alxemi
It´s was i usually use for 2cd... but every test shows that all the HVS are very similar, and hvs best is not always the best :)
And the difference of bitrate usually between 1cd and 2cd is big.
MPEG-Standar always gave me very bad results, mosquito etc.. Maybe you could do a quick compare between EQM2, HVS-Best and my v5 & v6 matrix ???

Bye

guada

27th June 2004, 17:41

"Anyway i was very newbie (just like today xD) and maybe messed up the configuration. It was early XviD too": Alxemi

Each of us is a few newbie in his existence.
For more news on methods of encoding, look Crusty's FAQ at this address:

http://www.vslcatena.nl/~ronald/docs/xvidfaq.html

Soulhunter

28th June 2004, 19:57

Think, its time to discuss the way I should post the results !!!

As this is meant to be a blind test, I considered this method...

- I give each screenshot a random number

- Write the appropriate matrix to this number down

- Post only the screenshots

- PPL mark them (points from 1=bad / 10=best)

- After some time I post the (number x = matrix y) list

Bye

stephanV

28th June 2004, 19:59

do we get to know what the original looks like?

Soulhunter

28th June 2004, 20:23

Originally posted by stephanV
do we get to know what the original looks like?
Before or after Ive posted the "result" list ???

Bye

Alxemi

28th June 2004, 20:39

Maybe you could do a quick compare between EQM2, HVS-Best and my v5 & v6 matrix ???
I will, as quick as i could finish my actual akira backup (which i think will take me some more time). Will post the results :)

hey your new results-method is very interesting ^_^ i cannot wait!

stephanV

28th June 2004, 20:48

Originally posted by Soulhunter
Before or after Ive posted the "result" list ???

Bye

that would depend on how you conduct the test. im just asking about the method. results might vary if we know how the frame should look like on beforehand, because testing without reference is different.

if we know what the frame looks like, we could use that as reference point (score = 10 on your scale). if we dont, then we will judge the visual quality of each frame independantly. in that case, the original might actually score lower then some of the encoded frames.

both testing methods are valid i think, only the results need different interpretation.

Alxemi

28th June 2004, 21:13

if we know what the frame looks like, we could use that as reference point (score = 10 on your scale). if we dont, then we will judge the visual quality of each frame independantly. in that case, the original might actually score lower then some of the encoded frames.

Maybe the second method is better in a filter testing, when you try to improve the original quality, but in a matrix-test i think is better to have the original in hand to compare.

Soulhunter

28th June 2004, 21:36

OK, I will post the reference frames also !!!

Bye

kitanai

1st July 2004, 19:36

Hi there,

I keep reading about the different Custom matrices. But: where can I find the FOX HOME ENTERTAINMENT matrix?

Could someone please post a link!? :D

guada

1st July 2004, 20:20

All you want:

- CQME 1.0,
- FFDSHOW,
- CCE ( but..)

Bye

Jawor

2nd July 2004, 11:29

It's bundled with all my XviD builds (http://jawormat.republika.pl) - unzip the file into a temporary directory and you'll find this matrix in a subdirectory named Custom Matrices.

kitanai

2nd July 2004, 11:45

Got it, thank you!

Sagittaire

2nd July 2004, 18:33

Soulhunter

2nd July 2004, 19:51

Originally posted by Sagittaire
Fox q2 is better than H263 q2 but perhabs H263 q1 is better than Fox q2...As this are the results of the 1st/2Pass, it should be...

Fox Home Entertainment @ Q1 is still better than H263 @ Q1 !!!

Originally posted by Sagittaire
...use 1-31 for quant interval for 2 pass with very high quality encoding and no size limitation
Read the stuff above... ;)

Bye

Soulhunter

4th July 2004, 20:06

So, here (http://forum.doom9.org/showthread.php?s=&threadid=79173) is the link to the blind-test... ;)

Hope ya like it !!!

Bye