New HVSmetrics matrices

[ReferenceDivx]

I just made some new matrices with a program i wrote. They are called good, better, and best. Good makes files with about the same size as 263. Better and best produce video that is even better than best(however the file sizes will be larger). I would suggest best for 2cd encodes and better or good for 1cd encodes.

Please post any results you obtain.

[Rrrough]

Hi !

thanx for your matrices, will try them asap.
I'm still trying to figure out, what makes up a good matrix, I have a tiny bit of an idea.
could it be that there's a typo in your "better picture" inter matrix in row 7, column 2. shouldn't that be a 26 ?

cheers

[ReferenceDivx]

#include <iostream.h>
#include <math.h>

#define quant 18

double hvsfwm[64]={ 1,1,1,1,1,1,1,1,
1,1,.9599,1,.9571,1,.9599,
.8746,.9283,.8898,.8898,.9283,
.8746,.7684,.8404,.8192,.7617,
.8192,.8404,.7684,.6571,.7371,
.7371,.6669,.6669,.7371,.7371,
.6571,.6306,.6471,.5912,.5419,
.5912,.6471,.6306,.5558,.5196,
.4564,.4564,.5196,.5558,.4495,
.3930,.3598,.3930,.4495,.3393,
.2948,.2948,.3393,.2480,.2278,
.2480,.1828,.1828,.1391};

int order[64]={ 1,2,6,7,15,16,28,29,
3,5,8,14,17,27,30,43,
4,9,13,18,26,31,42,44,
10,12,19,25,32,41,45,54,
11,20,24,33,40,46,53,55,
21,23,34,39,47,52,56,61,
22,35,38,48,51,57,60,62,
36,37,49,50,58,59,63,64};


void main()
{
int n=0,b;

for(int i=0; i<8; i++)
{
for(int j=0; j<8; j++)
{
//cout << hvsfwm[(order[n])] << order[n] << " \n";
//cin >> b;

cout << int((quant / hvsfwm[(order[n])-1])+.5) << " ";
n++;
}
cout << "\n";

}

float ben1=15.95;

cout << int(ben1+.5);


}

[Rrrough]

errm, I was just looking at the matrix as is, and in all HVS matrices the sums of rows and colums are the same, just not there, so I was wondering why it's not the case in the "better matrix" at the aforementioned position.
So you switched to the "standard" intra matrix now, has it proven to be the best for you ? the intra matrix from your first posting had another one, which doesn't seem to have those HVS-characteristics, or am I wrong here ?

[ReferenceDivx]

hvsfwm[64] stands for human visual frequency weighting matrix. It is in the zig-zag order from 0 to 64. I found it is a good idea to use hvs for both intr and inter matrices. Both intra and inter matrices which i have posted use these values. However, they are scaled more. Inter matrix adjusts the picture a lot more than intra matrix. Best has about the same picture quality as mpeg, but produces files about 15% smaller. Good produces files about the same size as 263 files. However, i dont know how good the quality really is compared to 263 files.

We will need people to test these matrices.


I think my original hvs matrix was wrong. Even though it gave good picture quality, the size was a lot bigger than it should of been. I want to get good picture and good size ratio.

[rui]

Referencedivx, i posted an experience about using your "best" matrix in iago's lumi thread (sorry about this, but the post was mostly concerning iago's avs script, so i posted there).

[ReferenceDivx]

Ok i read your other post. I just got done encoding the matrix using hvsmetrics best matrix. I liked the overall picture quality of the movie. I only encoded the movie part, without the credits. I was aiming for a filesize of 728662.

Q:2:29241
Q:3:144044
Q:4:12342
Q:5:504
Q:6:56
Q:7:3

Ok got run.

[Koepi]

Currently retesting "Contact" with unfilter(5,5) added to the script, and using the "good-picture-hvs" matrix from ReferenceDivx (which indeed looks yummie, dunno what picture quality that produces, but the matrix itself looks nice and promising) - let's see how it works out.
Just wanted to mention that here, too, as the matrix itself needs some feedback ;)

From my running first pass I can tell that the filesize seems to be even below h.263, we'll know more about visuals in ~15 hours.... (hell, I need a faster compi, 700MHz duron is simply too slow to compete with our serial tester ;) )

Thanks for the matrices, they look really promising, I hope they keep what they look like :)

Regards,
Koepi

[OUTPinged_]

Are there any general rules when making a custom matrix?

Any smart papers that describe what numbers should be picked and why?

[OUTPinged_]

Q: shouldnt matrix for reencoding of MPEG-encoded material be different than for encoding lossless source?

[MoonWalker]

I have made a test with the best hvs with whole Matrix for 2cd

Quantizer distribution :
Q:2:70967
Q:3:114631
Q:4:590

I must say that the quality wasn't what I expected.Some moving backround and a little noise..I have tested before TemporaSoften + MPEG and I had quite impressive results. The quality is good, but not to my eyes :)

Keep up the good work and be sure that I will test again these matrices..

MoonWalker

[Rrrough]

Hi,

not much time for testing right now, only a quick glance at quant 2.
I have a test setup with clips from 3 different movies (dark-bright, low-high motion).
60 snapshots are taken, framesize from debugview-output averaged, perceptional error from dctune and psnr from compare().
I calculated an "efficiency factor" which is
avg. framesize x avg. percepional error / 10,000
a lower value is better. hope this makes sense somehow.
the interesting part is imho, that best matrix is pretty much equal with builtin MPEG matrix, good and better come closer to h.263 from visual impression.
BUT good matrix is performing much better than h.263 as it gives much smaller filesize with almost same visual quality.

The results :

matrix |avg. psnr |avg. framesize |avg. perc. error |efficiency

MPEG builtin |45.73969836 |13782.39344 |4.573496639 |6.303373008

H.263 |45.87503443 |13430.06557 |4.789272066 |6.432023789

HVS good |45.23059063 |11946.34426 |4.834269098 |5.775184291

HVS better |45.30200656 |12334.01639 |4.824836508 |5.950961259

HVS best |45.67302131 |13980.55738 |4.471156115 |6.25092546

[manono]

Hi-

Instructive set of figures, but your MPEG builtin efficiency figure should be 6.303373008 (assuming the 2 previous numbers are correct).

[Koepi]

I did Contact with the hvs good matrix and got this quant distribution:

[992] Quantizer distribution for 2nd pass:
[992] Q:2:359
[992] Q:3:4038
[992] Q:4:94170
[992] Q:5:90789
[992] Q:6:12252
[992] Q:7:1428
[992] Q:8:374
[992] Q:9:208
[992] Q:10:92
[992] Q:11:64
[992] Q:12:64
[992] Q:13:24

...which is somewhat annoying, but let's do the visual test first.
EDIT: the visual results...

Well, the hvs_good_picture ,atrix seems to cope well with noisy sources. Unfortunately I had lumi masking switched on so some areas look bad, I'm testing with "Last Boyscout" now again without lumi masking and without post processing in the mpeg2dec.dll to see if it helps here too.
The sharpness of the picture stays intact, but the noise gets mostly filtered out. I made a very low bitrate encode and therefore the picture quality is really nice when thinking away the lumi masked blocks ;) (so we've to see what happens if the overall image gets a higher quantizer...)
Just want to share my first results, more to come.

Regards,
Koepi

[Rrrough]

Instructive set of figures, but your MPEG builtin efficiency figure should be 6.303373008 (assuming the 2 previous numbers are correct).

Yes, that's correct. I rechecked all figures, and the rest is correct. it was a copy'n'paste error.

MPEG builtin |45.73969836 |13782.39344 |4.573496639 |6.303373008

are the correct values. that makes HVS even more competitive. but of course the most important factor is visual impression.
corrected in prior post as well for better overview.
thanks for checking, manono
:)

[Koepi]

ok, quant distribution for Last Boyscout / hvs_good-matrix:

[992] Quantizer distribution for 2nd pass:
[992] Q:2:1631
[992] Q:3:93332
[992] Q:4:50647
[992] Q:5:371

direct mpeg2dec-pp.dll, cpu=0, lumoff=-2, lanczos resize to 640x272.

Checking visual quality now. Btw., file size desired: 627102kb - hit 627102kb! :)

EDIT: visual quality is fine. It looks better than h.263+lumi masking but achieving the same / better compression than that combination. Thanks for the great matrix! (redoing Contact now with this settings...)

Regards,
Koepi

[rui]

Well, this weekend i tried the "best" matrix with the full movie "The Replacements", with lumi in both passes, also using iago's script to try to eliminate the "black" problem, for 1 xcd, and a res. of 608x352 (i guess, not sure, was made in home),797MB with ogg audio, so my final avi has a little more than 700MB, and...the final avi looked ver bad :(

I am not blaming "best" matrix here! Or iago's script!

I believe that this movie just can't be done in 1 cd (in spite of having a little short of 2 hours), his compressiblility is horrible, and by using iago's script and "best" matrix didn't helped the situation either, even by using CPU=4 (but got ride of the some black in the jail scene, in spite of using lumi ;).
So next i will try using the "good" matrix, has Koepi's stated he got good results.
By the way, i like to use CPU=4 in mpegdec, because later when watching the movie i don't enable post processing. I saw a post by MaTTer saying that he looses detail in the process, but this could be a good option for those with lesser cpu's, has they don't have to enable post processing when watching (matter of taste, but they could try this).

[Koepi]

I'll report about visuals in some hours - CPU=0 even for Contact (usually not meant to be encoded for 1CD... very much movement, noise, camera-on-TV sets,... all bitrate killers you can think of), using lanczos resize - and I tried unfilter(-5,-5) now to get rid of some of the over-sharpness of the images,...
Compression is about 2.7:1, which usually looks bad, but I hope it'll turn out to be good.

The last Contact encoding suffered from CPU=4 - as did Last Boyscout. It's quite unusable for me now. This explains the unsatisfying result with Monster's Inc., too. I'll redo without lumi masking and with CPU=0 after Contact is finished ;)

I like this, some development with some quality degradation in the start, and big improvements in the end... _that's_ evolution :)

Keep the ideas and impressions/results coming guys!

Best regards,
Koepi

[Didée]

@ Koepi, a little OT:

Me too was very annoyed by the EE of Contact. I tried unfilter(-80,-80) (!) followed by some nifty warpsharp (hehe) and Lanczos. Encoding (3 diff. 2nd-pass variants) will finish this night, I´ll report tomorrow. The previews looked very promising.
Shall I prepare Screenshots? (´cause of warpsharp ;) )

[MaTTeR]

Originally posted by rui
I saw a post by MaTTer saying that he looses detail in the process, but this could be a good option for those with lesser cpu's, has they don't have to enable post processing when watching (matter of taste, but they could try this).

Yes unfortunately I'm a detail freak so the slightest loss to my eyes is noticed, I don't even use post processing during playback to be honest. However, Marc mentioned a tip using PP during the encode that I need to try out, it might eliminate the blur effect I was seeing(crossing fingers).

[ReferenceDivx]

I've been thinking it might be useful to find a hvs optimized matrix for each keyframe(or every frame) on the first pass. This could then be used on the next pass, to provide better compression and perceived image quality. I've been reading a lot of papers on the internet about the concepts behind optimizing dct matrices. Dctune uses these algorithms to tune image compression for each image. However, i havn't been able to track down the source code(plus it has a patent). Anyways I think we are on the right track.

I don't think one matrix or group of matrices are going to give all the answers. I think we need to have an algorithm that that everyone is comfortable using. It also needs to provide the option of better image or greater compression. I think all these things can be achieved.

Anyways, time to make some homemade doughnuts. Umhhh

[Mango Madness]

I just finished an encode of Van Wilder at original resolution, latest Koepi build, Mod HQ, best matrix, shootings for 2 700meg files with 224kbps mp3 audio (can't play ogm till i have a faster comp). The Best Matrix was very good and I'd say it reduced percepible noise by around 10% as just a guess. The Van Wilder DVD is very VERY crappy indeed (25% NTSC, 75% FILM), so i'm glad to see the matrix help as much as possible.

[Rrrough]

@Mango Madness
errm, the result isn't due to the matrix then, when you were using ModHQ, as it's modulating between h.263 for low quants and builtin MPEG matrix in higher quants. best matrix wasn't used at all then, you have to choose MPEG-Custom as quantization type.

cheers

[rui]

Well, i made another encode of the movie "The Replacements", this time using in iago's script the CPU=0 option in mpegdec.dll, and using the "good" matrix by Referencedivx.
The resulting avi is again a little bad.
Not so much blocks as using the the "best" matrix, but, by freezing frames in Vdub and comparing, one can see that the "best" matrix retains more detail, even if i used the "best" in conjunction with CPU=4, and now CPU=0 with "good" matrix.

The below results were achieved from a movie with 1h53 minutes, at a res. of 608x336, with a final avi size of 704 MB. (not including sound, it's a XCD).


Quantizers Analisis
---------------------

Quantizers Used For Movie :
------------------------------
Quant 2 Used : 74 Times, Percentage Used : 0.05%
Quant 3 Used : 43 Times, Percentage Used : 0.03%
Quant 4 Used : 13320 Times, Percentage Used : 8.18%
Quant 5 Used : 70126 Times, Percentage Used : 43.07%
Quant 6 Used : 59883 Times, Percentage Used : 36.78%
Quant 7 Used : 16505 Times, Percentage Used : 10.14%
Quant 8 Used : 2583 Times, Percentage Used : 1.59%
Quant 9 Used : 195 Times, Percentage Used : 0.12%
Quant 10 Used : 66 Times, Percentage Used : 0.04%
Quant 11 Used : 7 Times, Percentage Used : 0.00%
Quant 12 Used : 5 Times, Percentage Used : 0.00%
Quant 13 Used : 4 Times, Percentage Used : 0.00%
Quant 14 Used : 4 Times, Percentage Used : 0.00%
Quant 15 Used : 2 Times, Percentage Used : 0.00%

Average Quantizer Used for Movie : 5.542

Quantizers Used For Credits :
--------------------------------
Quant 20 Used : 7385 Times.

Size Analysis
----------------

1-Pass Size : 2393374362 Bytes or 2337279 KBytes or 2282 MBytes
Scaled Size : 735022237 Bytes or 717795 KBytes or 700 MBytes
Actual Size : 735006152 Bytes or 717779 KBytes or 700 MBytes
Actual Size of Avi File without Sound : 739183616 Bytes or 721859 KBytes or 704 MBytes

Usefull Statistics
------------------
Compressibility : -38.65%
Relative Quality of XviD avi : 36.09%
Absolute Quality of XviD avi : 89.37%

Pretty tuff piece if video, isn't it? ;)

[Bulletproof]

I tried a few tests and noticed that the "best" matrix makes dots in the picture which don't belong, like the MPEG matrix does. After a few seconds into a scene I saw tiny colored blocks appearing in the frame, until the next keyframe came.

I then tried the "better" matrix, and the blocks are no longer there.

[ultimatebilly]

@ReferenceDivx:
Could you maybe repost your attachments? Due to server-problems all attachments seem to have been deleted...
Does anybody know if these problems have been solved?
Or could anybody just post the matrix-values into this thread, so I can copy and paste them into a txt file?
Thanks in advance!!

@Bulletproof:
The way I understand quantizer matrices, it has to be this way, unless you provide a matrix like the best-matrix with rather high bitrates, because it doesn't restricts the dct-values that much, causing the run-length-encoding don't being as effective.
Therefore the encoder has to quantize single blocks much more heavier than suggested by the matrix, which results in the effects you saw...
So my suggestion is to use the best-matrix only with 2cd+ encodes, and to use the good-matrix for 1cds...
The good-matrix restricts mainly the higher frequency-values much more, which results in a bit more blurriness, but needs less bitrate (because more values are brought to be 0)...

Uups... I just saw that this thread is rather old... anyways :D

[manono]

Hi-

After someone approves the attachment, you should be able to get them.

[ultimatebilly]

Can't wait :D
Thank you very much!

[duartix]

Well, I've searching the board as crazy and I can't find these matrices. Where are they?

[Koepi]

There they are.

;)

Regards
Koepi

[Nibor]

Yeeeehaw!

I think many people (including me) were searching for attachements which couldn't be found :D
But now the attachement is finally here (again) =)
Many thanks!!!

PS: XviD is just great! Tnx to all developers!!!

[NiTroGen]

You can also found a small collection of 15 matrices here (http://homepages.pathfinder.gr/nitrogen/files/XviD%20Quant%20Matrices.rar). It includes the matrices created by ReferenceDivx, Andreas, Bulletproof...

[JagPanzer]

NiTroGen tnx! No more searching for me. :cool:

[ultimatebilly]

Yeah!
A matrix collections is what we needed!
Thanks!

[duartix]

First of all big thanks NiTroGen.

What's this thing with attachments lately?:confused:
Can anyone spear me the irony and explain what's happening?
Pretty please?

[Swede]

Serverproblems. We will announce when everything is up again.

[Rash]

The HVS matrices made VirtualDubMod to crash. :(
I'm using Koepi's build XviD-03052003-1.exe. Is it just me?

(Yes, I'm sure it is the matrices. Going back to h.263 and Modulated HQ made VDubMod behave normally now).

[Nibor]

@Rash

I tried the hvs-good matrix with VirtualDubMod 1.4.13 using the same XviD build... 2 pass...
Its strange, the first pass started but at the end, when it was nearly done, VDubMod crashed! When I opened the stats file in StatsReader it seemed bo be ok, correct number of frames and normal stats...
Then I did the second pass using the stats file from the first pass, and VDubMod worked, but at the end it crashed again! But the output video was ok, no weird things...
But that's weird! If the matrix (no not THAT matrix :D) was corrupt, the output wouldn't be normal, but if using an other custom matrix, VDubMod doesn't crash, so the problem must be something with this matrix in combination with VDubMod, or am I totally wrong?
Hmm, I think I'll try a newer version of VirtualDubMod...

[Koepi]

I'm using customn matrices (especially hvs-good-picture) quite frequently and they never produce crashes.

Koepi

[rcjc]

has anyone had problems with using the custom quant matricies with 10/4 stable?

[corigan]

I'm with koepi on this one, done a few encodes with koepi's 5-03 codec and hvs_good matrices and had zero problems in the new vdubmod 1.5.1.1a

Corigan

[iago]

Crashes might be Avisynth 2.5.1 and/or VDubMod related, as being discussed in the avisynth forum. I don't think the problem is matrix-related ;).

Recently I've also been using the hvs-good-picture matrix quite often (imho it's ideal for 1CD rips) and so far (fortunately ;)) I haven't got any crashes either.

[Rash]

Later I'll try using another matrix.

The second time I tried, it made my computer hang. (WindowsXP SP1)

Just a note. I was using the hvs-better and not hvs-good matrix.

[manono]

Hi-

I use all 3 of them exclusively. Which I use depends whether or not I'm going for one or two CDs , and the results of the compress test. I used to see Nibor's problem sometimes, where it would crash on the last frame, but the AVI turned out fine. But I haven't seen that problem for awhile. And I've never seen it crashing in mid-encode, or at the beginning, or freezing the computer, or any type of instability. This is using all XviD versions from 10-4 stable forward to the most recent ones. So I don't think the problems are matrix related.

I guess I'm stating the obvious, but to be able to use them, you have to choose MPEG-Custom in the Global Tab-Quantization Type.

[Rash]

Yes, I was using MPEG-Custom. So when I switched back to h.263 and Modulated HQ I had no more craches (I haven't even reinstalled VirtualDubMod).

Anyway, are those matrices recommended to use on both passes?

[iago]

Originally posted by Rash
Anyway, are those matrices recommended to use on both passes? Absolutely, that's what you should do! Do not change the matrix you used for the first pass.

Btw, afaik, using Modulated or Modulated HQ is not recommended anymore.

In short, use either h263, or mpeg, or an mpeg-custom matrix for both passes and do not change your matrix between passes ;).

[Nic]

Wise words as always iago :)

-Nic

[Rash]

That's very good to know. Even though I've just made a perfect rip with Modulated HQ. ;)

[ultimatebilly]

Even though I've just made a perfect rip with Modulated HQ
The problem is only that Xvid-files which use more than one matrix aren't MPEG-4 compliant, and will therefore maybe (or probably) not be played by MPEG-4 capable standalones...
But for Computer-playback, you can of course use modulated or modulated HQ...

[iago]

But for Computer-playback, you can of course use modulated or modulated HQ...Maybe, but with one bitter drawback I guess -> as long as you don't use b-frames!

http://forum.doom9.org/showthread.php?threadid=53136&highlight=bframes+modulated

[ultimatebilly]

Oh! I didn't know that!:scared: This explains why my AI-encode I did a while back was totally fucked up... (The max I-frames setting wasn't respected)... I had in one place more than one minute without a new I-frame :D )...
Thank you very much for pointing me to this!!

[powerslave]

I think maybe the virtualdub crashes with custom matrices occurs when using gknot. At least mines does, and ive read some posts that seem to indicate using vdubmod on its own, there's no problems with these matrices. Are the ones who have no problem using gknot or just vdubmod?

[powerslave]

Yes, i just confirmed that when using virtualdubmod as a standalone, it will take the custom matrix (in my case hvs-good) and encode the avi but still fails when doing it through gknot. I tried only 1 pass for time purposes, still need to mess around and learn a little bit more about using vdubmod because the file size ended up higher then i had set up through gknot. Anyway, for those of you who had trouble with gknot, try setting up the avisynth script in gknot, then instead of adding it to the queue via gknot, close gknot and open the avisynth with vdubmod and set up the codec parameters, etc. Look at the xvid guide here on doom9, that explains a bit how to do it manually. Then you can use nandub to plex the video and audio together. The quality i got was descent for a first try, but i was finally able to get the filesize down without using qpel and thus no more qpel streaks and blotches. Still alot of noise, i guess i have to now concentrate on avisynth filters to get the encodes looking the way i want. Anyways, that was my breakthrough for the day haha.

[d'Oursse]

No problem when using VdM alone. Problems with GK. I think that GK does not write the custom matrix components in the entries of the registry.

[len0x]

Now that I have this problem also I'm very keen on solving it.
Actually I have an idea:
GK uses registry to set settings for the codec (and they are OK - I verified that). VDubMod propagates those settings via SetCompData mechanism (I think) that's why it works there.

Can XviD developers comment on that? (i.e. probably it's worth cheking how matrix is being picked up from the registry when SetCompData is not set).

P.S. I'm talking about crash straight in the beginning of encode...

[len0x]

After I got no help from xvid developers (although it's fair to say that some of them tried :) ) I hacked this problem around.

I now use SetCompData for xvid settings in VDubMod script and it works starting from GK 0.28.6.1

[alucard83]

Niceness! I encountered this prob last night when I loaded a custom matrix and I immediately got prompted with 3 errors in a row. Finally gave up, it's good to see I wasn't alone on this:D

[DarkDudae]

Originally posted by ReferenceDivx
I've been thinking it might be useful to find a hvs optimized matrix for each keyframe(or every frame) on the first pass. This could then be used on the next pass, to provide better compression and perceived image quality.

We also are looking for info about that to add that function in AVSOptimizer. Btw, great work with your matrix :)

[Tuning]

Eventhough very late to post this, Thanks Referencedivx.

So again most obvious question arises!. How to make custom matrices ?
Or How the numbers in matrices related to quality.

And as XviD 1.0 has released, what will you preffer as XviD 1.0 settings. (I think XviD 1.0 has now modulated quant in name of adaptive quantization...and many other changes).
So what settings i need to discard/include while using hvs_matrices ?

Thanks XviD enthusiasts in advance!

[Koepi]

TUNING! Please use the search button more extensively and _read up a lot first_ - adaptive quantisation is what we called luminance masking before (it does the same and is scientifically more acurate).

Quantizer type modulation has never been MPEG4 compliant. That's why you don't see it in XviD-1.0. It was a thing to play with, a proof-of-concept.

There are plenty threads in this forum explaining custom quantisation matrices, please refine your search.

Regards
Koepi

[Tuning]

Originally posted by Koepi
TUNING! Please use the search button more extensively and _read up a lot first_ - adaptive quantisation is what we called luminance masking before (it does the same and is scientifically more acurate).
Quantizer type modulation has never been MPEG4 compliant. That's why you don't see it in XviD-1.0. It was a thing to play with, a proof-of-concept.
There are plenty threads in this forum explaining custom quantisation matrices, please refine your search.

Regards
Koepi

Thanks koepi again. I just found the required thread by :search:

Thanks.

[MfA]

Modulated quant might make a comeback ... some of the developers would really like qpel to be adaptive (and for that you need to change VOP parameters too).

Although personally I think changing the quant type every frame doesnt make a whole lotta sense.