KVCD template VS. Tmpgenc defaults?
 

Just out of curiousity, I decide to make a comparison between the two. I used the size 704 x 480, 500-2400 bitrate, and 224kbps 48khz audio as constant and varied the CQ from 50 in steps of 5 up to 80 (eg. 50, 55, 60, 65, etc.).

The source is a 15 minute dvd rip of the movie The Abyss. A background, there is a very difficult scene at the beginning where the title "The Abyss" gradually enlarges until the Y becomes an underwater scene of blue and block. This scene produces blocks even though most if not the rest of the movie are already blockless.

These are the resulting file sizes using the:

1. TmpGenc default GOP and matrix

CQ = 50 -> 155,457mb
= 55 -> 155,457mb
= 60 -> 155,457mb
= 65 -> 155,457mb
= 70 -> 155,457mb
= 75 -> 173,803mb
= 80 -> 199,633mb

Comment: Quality-wise, there was a little hint of blocks at the opening scene at CQ 50. 65 is good, 70 is excellent, no more eyeball difference for the higher CQs.


2. TempGenc default GOP and Andreas matrix

CQ = 50 -> 119,652mb
= 55 -> 126,581mb
= 60 -> 136,821mb
= 65 -> 155,457mb
= 70 -> 164,215mb
= 75 -> 195,457mb
= 80 -> 221,675mb

Comment: Filesizes are large. Acceptable quality starts at CQ 75.


3. Kwag GOP and TmpGenc default matrix

CQ = 50 -> 106,868mb
= 55 -> 110,452mb
= 60 -> 116,144mb
= 65 -> 125,410mb
= 70 -> 132,037mb
= 75 -> 155,452mb
= 80 -> 185,224mb

Comment: Very little filesize compared to the rest and quality-wise, the opening scene at CQ 50 is at par with Condition 1 CQ 65.

4. Kwag GOP and Andreas matrix

CQ = 50 -> 116,309mb
= 55 -> 122,144mb
= 60 -> 131,093mb
= 65 -> 147,543mb
= 70 -> 158,645mb
= 75 -> 189,640mb
= 80 -> 219,771mb

Comment: Quality-wise, there was blocks at the opening scene at CQ=50, and becomes worse as I increase the CQ. Blocks decreased at CQ=75 and lesser at 80.

Conclusion: I actualy graphed the resulting bitrates as against the CQ. Compression-wise, using the Kwag GOP and TempGenc default matrix is definitely tops and the quality is acceptable at CQ=50 right away and drastically improves as you increase the CQ. This is just using eyeball though. I'll do the same test on another dvd ripped movie and if the pattern holds true, then probably I'll be using the Kwag GOP/TempGenc matrix combination.

Great data Timberwolf! I'll have to make some tests myself at different resolutions. Apparently the "Andreas" Q matrix was optimizez for SVCD encoding. Or 352x480(576) resolutions. I think it was never tested at 704x480(576). Possibly a solution would be a different matrix for different resolutions. Volunteers :lol:, because that's a hell of a job!

kwag

Great post Timberwolf!

Thanks for all great info and work on that project..Definitely worth a bookmark for future reference!

Thx......... 8)

Thanks everyone! I hope it does any of you some good like it does me. Remember though that the result is just good for 704x480 size and for that particular movie. i have to do some more test to arrive to a better informed conclusion.

I did the test on two consecutive nights, just set the batch encode then let it run while I was sleeping. The Motion Search Precision was set to Normal to make the encodes faster.

Amazing info Timberwolf. Using the Kwag GOP, notice that you can bump the CQ from 55 to 65 _with nearly the same filesize_ just by switching back to the default matrix! That's huge! This means a whole range of movies that were just slightly too big before may now fit! And those movies that fit already can have their CQ pushed from the default of 50 up to nearly 65 in some cases. 8O

One oddity though... in your first set of tests, the file size is identical from CQ 50 up until CQ 70! Typo, I assume?

A question: You said that in test group 1, CQ 70 was excellent and you got no improvement going higher. In test group 3, you said CQ 50 was equal to CQ 65 in test group 1. What CQ in test group 3 was equal to CQ 70 of test group 1? Does test group 3 also have a cap where raising the CQ further gives no improvement?

Hmm... another question: Why do you use a bitrate range of 500-2400? The default for the template is 300-2300. Does this matter much?

Kwag... I'm willing to help test matricies if you have some in mind that need to be tested. I had always meant to ask you about the Andreas matrix, because I believe in German the name is 'Andreas simple 99er SVCD' (implying it was best for SVCD, I suppose), but I figured the matrix must have no relation to the resolution. Obviously does, eh? The 'Andreas 78er' matrix might be the compromise between the default and the 99er matrix?

Nope, not a typo. I actually redid those because I thought I made a mistake when I saw them equal. Why they turned out the same? i really don't know.

That's right! It seemed like there was no visible difference between CQ 70 to 80. That's from my eyeballing it though, you might have better eyes and could see some difference.



To my eyes, I thought CQ 50 in test group 3 was equal to CQ 60 in test group 1.

Truthfully, I find it hard to decide where the cap is because when you attain a superb quality the differences from then on are so subtle that I may not be aware of it. That's why I will just maximize the CQ as the 2 cdrs will permit you :)

Take note also that I'm commenting to the qualities of the encodes based on the most difficult scene in the movie, meaning if blocks will appear, it will probably appear at that scene first. When I say that the quality is good, it is only for that scene. For the rest of the movie, the quality may probably be excellent already.[/b]

No particular reason at all. lol

I was doing some experimenting with the bitrates, I didn't set out to do this particular Kwag/Tempgenc comparison at all. Then along the way I thought of this test. And since I already have some data gathered with those minimum-maximum settings, i decide to proceed with it so as not to waste that data.

Hey timberwolf

:oops: I´m not too much of an expert in this stuff yet, but i´m trying...

So, don´t mind if i ask but, how exactly did you setup, you loaded the tmpgenc default and use the 704X480 config or the opposite ?

Thanks

:wink:

Nor am I ....

yeah, that's about what I did.

Okay I set up a batch encode last night comparing further the Kwag GOP/Tempgenc default matrix versus the Kwag GOP/Andreas matrix.

The movie is The Matrix ripped from a dvd. The settings are 704x480, 23.976fps, CQ 66, 300-2300kbps, 48khz 224kbps audio.

Contrary to my findings with the The Abyss movie, the Kwag GOP/Andreas came out with lesser file size.

Kwag GOP/Andreas matrix = 1,362,614kb
Kwag GOP/TmpGenc default matrix = 1,392,614kb

I don't know the reason why, maybe it just depends on the movie on which method will produce a lesser filesize. Or the Kwag GOP/Andreas matrix may have had produced a lesser file size in "The Abyss" in the first place if I had encoded the entire movie and not just a 15 minute clip of it.

Anyways, there is a big difference in quality. The Kwag GOP/Andreas matrix have blocks all over the place. The Kwag GOP/TmpGenc default matrix is devoid of it, it is just perfect.

With the small difference in filesize and the quality a lot better, I'll be using the Kwag GOP/TmpGenc default matrix from now on for 704x480 encodes.

Thanks for your efforts Timberwolf! I'm going to make a batch encode myself, to double check your results. If I get the same result, I'll update the x2 templates with the standard Q.Matrix. Maybe the Andreas matrix is optimized for the lower resolutions. This is something I didn't test. And I believe all the tests that were done by the Andreas group, were done with 352x480(576) and 480x480 resolutions. None were done with 704x480(576). Obviously, there should be an "ideal" Q.Matrix for the 704x480 template. I haven't got involved in the Q.Matrix math or calculations yet. So that area is pretty dark for me right now. I was completely focused ( initially :wink: ) in the GOP structure, because I was more concerned with a descent space/quality goal. But I guess the next step is an analysis of quantization matrixes. For GOP, I consider that done. It's really optimized to the max. And the current GOP structure seems to scale well throughout all templates. Meaning it works the same for all resolutions. But Q.Matrix works very different, for what I can see in your results, with different resolutions.

kwag

For the heck of it, I encoded the Warcraft 3 trailer from Blizzard (1024x488 DivX) to the KVCDx2 templates, and I found something similar to Timberwolf's initial findings... the clip is about 2:30 in length, and with the Andreas matrix it came out nearly 15meg, while the default matrix came out 12meg. Because the original source is DivX, it is very difficult to compare quality between the two (I don't know of a simple way to load up the original plus the two test samples side-by-side to compare). I guess the good news would be that I didn't notice any huge differences in quality, but shaved 3megs from the filesize... a very nice gain! I'm now going back and using the default matrix with higher CQ values to compare with.

Does anyone know of a nice method for loading up multiple mpegs plus the original .avi to be able to compare side by side?

Update: As with Timberwolf's first test, I could bump the CQ from 50 to 60, and got nearly the exact same filesize as when I use the Andreas matrix at CQ 50. Again, this is computer animation from a DivX source, so it may not be the best starting point. I still need to burn them all to CD and take a look on TV. PC viewing makes it hard to tell.

Please do check my results. I have become enamored with your 704x480 templates ever since I first tried them and is now exclusively concentrating my efforts on them.

Before I made this comparison test with the GOPs and matrices, my initial encodes were very good results for the 1-cd template but poor on the 2cd template-very blocky. I discounted my findings for the latter because I've only seen two aside from me complain in the board that they see lots of blocks with that template. So when I wanted to do a 2cd encode, I just used the 704x480-1cd template and just upped the CQ to anywhere up to 54. Just up to CQ 54 because beyond that blocks will appear and will increase as you continue upping the CQ. I am really puzzled why it was just the two or three of us who complained of blocks when using that 704x480-2cd template, since I don't think I made a mistake in using it.

My recent GOP and matrices comparison proved that the Kwag GOP and Andreas matrix combo gets blocky beyond CQ 54. I'm waiting for your own tests to confirm or disprove my findings. Anyways, the beautiful thing is I found your GOP settings work beautifully with the TempGenc default matrix. And it works the way it supposed to be, the quality increases as you increase the CQ.

The lack of people complaining about the 2CD template might be from the lack of people using the 2CD template ;) I know I haven't yet tried it, as most movies I've been playing with to this point either fit on 1-cd already, or were 'throw away' tests, so weren't wroth re-doing with the 2-cd template for 1-time viewing. You saved a lot of work for those of us that _will_ be using the 2CD template sooner or later :)

That's actually what i thought was the reason. Although, someone coming in to confirm the 2cd template being blocky beyond CQ 54 would be nice. :)

Timberwolf:

You're dead on RIGHT!
I just batch encoded the first 15 minutes of "Proof of Life" with the KVCD 352x480 template. Here are the results.
With "Andreas" matrix: 140,557KB
With "Default" matrix: 121,747KB 8O
At 352x480, I can see no visual quality difference between the encoded mpegs files. I'm encoding another 15 minute clip, "The Green Mile", to see if I get the same result. If I get the same result, I'm rolling back to the "Default" matrix on all templates.

kwag

Here we go. This is the result of the first 15 minutes of "The Green Mile"

Default matrix + KVCD GOP 1-18-3-1-48 : 118,559KB
Andreas matrix + KVCD GOP 1-18-3-1-48 : 141,728KB

There are some very very small visible artifacts around some edges in the default. They are also visible, but barely, in the Andreas test run.
I really can't see any other difference between the two. Even looking closely in the PC monitor. Both are pretty blockless.
But 23MB difference is HUGE. This multiplied by 4 is 80+ average MB savings per hour.
So I'm switching the matrixes back to default on all templates tonight.

kwag

I have been testing myself. I have been testing for 352x240 though. I compared the current 352x240 template with the original 352x480 template (changed to 352x240). I tinkered with CQ until I got comparable filesizes. The old template changed to 352x240 (w/ original matrix) produced noticably better quality.

question! How do you know if you're using the andreas matrix or the default matrix? When i load the template and click on the "quantize matrix" tab, the drop down menu is blank, there is no matrix selected. Should i select "default" from the drop down menu to use the default?

KWAG already updated tamplates on home page .
Download from there and when u load it to TMPG u'll see "Default" appear in matrixes box.
Anyway, All templates are updated since 30-6-2002 night .
bman

i never liked andreas matrix, it always appeared unsmooth and noisy. glad to hear that all the templates are back to using the default. hey KWAG, try changing the B/P spoilage back to 0/20, ive noticed significant filesize reduction with it back to its default as well.

question: anybody know if the default matrix in tmpgenc is the same as the natural image matrix within the panasonic encoder? ive always liked the results that pwi could produce, besides the color crap-out. i wonder if there is a way to extraxt the internal settings from panasonic and compare and test them.

I am using the KVCDx2 2-CD templates to encode NTSC programs captured with AVI_IO MJPEG=19. The CQ=70 and the quality is amazing! No blockiness at all. Could the problem be related to ripped sources rather than AVI?

Hey All,

I've also been running similar tests using the KDVD templates.

Here's a file listing of my results for a 2 min & 26 second trailer from the Star Wars Ep one DVD, using the Kwag GOP and a standard matrix;

GOP Used: 1-18-3-1-18

36,184,068 Full-D1-CQ.mpg
29,374,468 New-Full-D1-CQ-70.mpg
26,722,308 New-Full-D1-CQ-65.mpg

I found that at a CQ of 65 it looked very good but there were some artifacts, but bringing the CQ up to 70 improved the visual quality up to what the "Andreas" matrix provided.

At a CQ of 70, this is still significantly smaller than what "Andreas" provides. :D


22,337,540 Half-D1-CQ.mpg
21,948,420 New-Half-D1-CQ-75.mpg
18,079,748 New-Half-D1-CQ-65.mpg

Half D1 is a different story, it seems that only at a CQ of 75 do I get a result similar to the Andreas version but the file size is almost the same. Personally, I think that the Andreas matrix encoded version still looks a little better in this test anyway, I guess it'll need some more playing with. :?

Hope this info is usefull ....

Later

Why the non-Kwag GOP? Since it is known that the 1-18-3-1-48 provides better compression, this may have skewed your results some (at least if others do the same test).

DAMN DAMN DAMN #$%@$# :twisted: It was supposed to be 0/20 and not 0/0 8O
Just too many details to keep track. Thanks again deltaboy :lol:
I'm updating all templates in a few minutes.......

.......Verified. Done.

kwag

Yes, but I'm not testing the KVCD templates, but the KDVD templates. 1-18-3-1-18 is the same GOP as in the KDVD templates with the Andreas matrix.

This is not the DVD standard GOP that TMPGenc has. DVD compatibility demands following strict guidlines.

TMPGenc's standard DVD GOP for NTSC is 1-5-2-1-18.

Don't get me wrong, I'm just following what was in Kwags template. I have no idea what the difference between the two GOPs is or what they mean. 8O

Hey guys, help me out on this one...

I´ve been using the Kwag´s GOP on the Default TMPGENC matrix, but i haven´t made the results you guys made... :cry:

I try to encode ALI. It´s not a difficult movie, low action, no underwater...
The image is good ( CQ 60 ) but still not good in filesize... ( 15 min = 149,678 )

It´s about 9,8 per min, so it will not fit even a 90 min movie :?

Any suggestion ?

Thanks ! :wink:

By the way : The templates in the download page are already with the GOP ( Kwag ) and TMPGENC ( matrix ) Changed ?

Sorry, my mistake. I don't read the acronyms closely enough!

Yes. All templates are updated with default matrix and KVCD GOP.
About your movie, is it a wide screen or full screen movie?
Also did you do IVTC?

kwag

Kwag

I have been converting your templates to 352x240 VCD and 480x480 SVCD. These formats are the only two that work both in the Apex 1500 and the Apex 1200.

Using test samples, I converted older and newer templates to 480x480 SVCD. Older templates convert better (size/quality). I also tinkered with old matrix and GOP settings. I got smaller filesize with my testing using older GOP (1-12-3-1-0) than with newer GOP (1-18-3-1-0). I thought newer GOP settings would create smaller file size. Could it just be the sample 10 min clip I'm doing??

Thanks!!

You are correct!. The older GOP produced smaller file size, but also produces a flashing effect on the video with long scenes, because of the longer GOP. That's why the last value was changed from 0 to 48, so that the GOP's would be shorter.
The flashing effect had been confimed by many people who used the old GOP. The new 1-18-3-1-48 fixes that.

kwag

Kwag I still use the original KVCD template as I think it gives better quality than any of the later versions. I have never seen this "flashing" that you describe. Is it very noticable?

Yes it is!. In the sample I had from "Under Siege 2", it was very visible.
It all depends on the movie and the scene.

kwag

ANDREAS Matrix
 

Hello every one,

this matrix have a history. First, the matrix is develop for a MPEG2 video stream for a standard SVCD an ~120min with 2pass on 2cds. Second, the old TMPEG versions have many problems with the B-frames. From the second B-frames the encoder had not enougth bitrate for the other B-frames. The 99 is only to save blockness in the movie. Detail and motion are not so good. My tests (renew article :) ) have shown, that the 99 Matrix is not so good for ths KVCD. I take only the standard matrix from TMPEG and if I test other solutions I have only develop a new matix for 448 x 320 and 320 x 432 motion area. For the new generation of TMPEG 2.54 and higher I thing we have develop a new solution for the MPEG2 standard SVCD.

I have encode my first whole movie. It is "rush hour 2" in PAL with 84min playing time. The result is :

CQ68, min 650kbps, max 2550kbps, P and B frame with 0

In the first step I used virtual dub with filter and frame server and secound TMPEG CQ. The amount of the file is 832000 MB by 126795 frames. The quality is better as a VHS tape but a little bit more worse then a very good S-VHS tape (copied from DVD PAL with an macrovisionsdecoder to a very good S-VHS video-tape recorder).

ANDREAS

Kwag,

You misread my post. The old template I used was ( 1-12-3-1-0 ) compared to ( 1-18-3-1-0 ). I haven't tested ( 1-12-3-1-48 ) versus ( 1-18-3-1-48 ) yet though.

Thanks!!

Re: ANDREAS Matrix
 

Hi Andreas !
Welcome !
I would to ask u to help me to make clear how exectly matrix is working ?
If he affects only quality or file size too ?
When u are talking about resolutions and matrix performanses for them
I would like to understand better what u mean .
I just don't have needed background to make this clear for me .
If u can give me any link with good explanation will be great !
bman

OK, I'll try it with the "48" instead of "0" and see if it makes any difference. Just encoded Jewel Of The Nile from a noisy DVD (glad I didn't buy it!!) with a CQ of 66 and the file size was 760Mb. Am about to encode Romancing The Stone (from a similar noisy DVD). As they are both the same running time (101 mins) I will see if this change will affect the final file size or not. As long as it is under 795 mgs NERO will burn it OK.

@bman

well the matrix. The matrix is most of the complicated things in an encoder. On whitch way they are working (intra and non-intra matrix) is the know how and secret of the developper from the encoder. We understand the whole theme in this time not 100% exactly. At this http://tan.informatik.tu-chemnitz.de...mpeg_tech.html and this http://www.kom.e-technik.tu-darmstad...cs/basics.html Link you have a very good overview about MPEG codiering. Go on a searching mashine and tipe "intra matrix, non-intra matix, mpeg, DCT and iDCT". With this words you found a lot of informations in the net. If you speak german you can visit this site http://www.edv-tipp.de/gastbeitraege/kika001_dct.htm. It is an artikel about intra matrix from one of our mods.

ANDREAS

addition:
http://viswiz.gmd.de/DVP/Public/deli...g/hi_str01.htm

@ Andreas again !
THX for so quick replay !
I'll check all those links and hope it'll help me to acheve some knowledge
in this complicated theme . :wink:
THX again and I'm really glad u joined us ! :D :D :D
bman