ColorMatrix v2.3

[Steve56]

Yes that's true, I noticed that too late. His plugin contains optimization code, so it might be faster. (btw, hdtv isn't always Rec.709 as suggested in his docs ...) [/B]

ah, good to know ;-) I did now try to check with GSpot, if a certain MPEG-2 HDTV 1080i stream (Astra 1080 Konzert1), is encodet with the BT709 or BT601 matrix. But I don't know where to read this information from?!

Steve56

[Boulder]

Open the clip in GSpot v2.52b01.

[Wilbert]

1) That info is also available in 'DGMPEGDec 1.0.13 beta 11'. Using


mpeg2source(info=true)

2) Use latest beta of GSpot and look in the upper right corner. In the video section, the five boxes next to NVOP. The fives boxes are from left to right:

1) I709 (= Rec.709 coefficients)
2) FCC (almost the same as Rec.601)
3) I470 (= Rec.601 coefficients [an updated version of Rec.470-6, but coefficients are exactly the same])
4) S170 (= SMPTE 170M; exactly the same as Rec.601)
5) S240 (= SMPTE 240M; almost the same as Rec.709)

see table 6-9 page 56 of http://le-hacker.org/hacks/mpeg-drafts/is138182.pdf

If the boxes remain "black" then the used coefficients are not present in the header, and the default (which is Rec.709) should be assumed.

I hope it's clear now

[Steve56]

Wilbert: Thanks for the comprehensive answer!

Before that, I just did a color conversion judged by the look of it, when turned on or off. ;-)

[tritical]

Wilbert, is the source included in the zip linked to at the beginning of this thread the latest? I tried doing some simple optimization and got it running about 40% faster (about 210fps vs 145fps in YV12 on my comp) still with plain c code, but I'm wondering about a few things looking through it... in the constructor it checks mode against "Rec.709->Rec.601" and "Rec.601->Rec.709", yet in the main part it checks mode against "mpeg2->mpeg1" is this intended? In the YV12 mpeg1->mpeg2 it uses the "21" conversions and not the "12" conversions, but in the YUY2 code the mpeg1->mpeg2 does use the "12" conversions? Also, could you tell me or point me to a source that has the numbers for how the coefficients used were derived? I'm just curious, I tried to calculate them myself using coefficients from a few different websites, but they never work out to be exactly the same (usually off by 20-40 after multiplying by 65536). Thanks in advance.

Hm, must be because:

mpeg2source()
clip1 = last.colormatrix(mode="rec.601->rec.709")
clip2 = last.colormatrix(mode="rec.709->rec.601")
subtract(clip1,clip2)
coloryuv(analyze=true)

shows both clips to be identical .

[Wilbert]

in the constructor it checks mode against "Rec.709->Rec.601" and "Rec.601->Rec.709", yet in the main part it checks mode against "mpeg2->mpeg1" is this intended? In the YV12 mpeg1->mpeg2 it uses the "21" conversions and not the "12" conversions, but in the YUY2 code the mpeg1->mpeg2 does use the "12" conversions?

I corrected all that (v1.6). Thx very much. It's a bit stupid that I didn't notice this

Also, could you tell me or point me to a source that has the numbers for how the coefficients used were derived? I'm just curious, I tried to calculate them myself using coefficients from a few different websites, but they never work out to be exactly the same (usually off by 20-40 after multiplying by 65536).

I included a text file (coefficients.txt) which contains all calculations (if you know maple i can send you the worksheet). I hope that's clear enough.

Btw, some websites use slightly different coefficients, perhaps that could be the problem.

I tried doing some simple optimization and got it running about 40% faster (about 210fps vs 145fps in YV12 on my comp) still with plain c code

Great! Could you add those changes to v1.6?

[North2Polaris]

Originally posted by Wilbert
Yes that's true, I noticed that too late. His plugin contains optimization code, so it might be faster. (btw, hdtv isn't always Rec.709 as suggested in his docs ...)

@Wilbert,

Other than speed, any advantage to using ColorMatrix over BT709ToBT601?

North

[Wilbert]

Other than speed, any advantage to using ColorMatrix over BT709ToBT601?

Rec.601->Rec.709 conversion (useful for capping -> mpeg2) is possible in ColorMatrix

[Boulder]

Originally posted by Wilbert
Rec.601->Rec.709 conversion (useful for capping -> mpeg2) is possible in ColorMatrix

However, if you use CCE for encoding, no need to use ColorMatrix as it doesn't specify any coefficients for the encoded file for some reason?

[Wilbert]

However, if you use CCE for encoding, no need to use ColorMatrix as it doesn't specify any coefficients for the encoded file for some reason?

Only if you use CCE Remember pal/ntsc uses Rec.601 coefficients. TMPGEnc/QuEnc add fcc (approx Rec.709) coefficients to the mpeg-2 header.

[tritical]

I included a text file (coefficients.txt) which contains all calculations (if you know maple i can send you the worksheet). I hope that's clear enough.

That was exactly what I was looking for , thanks.

Great! Could you add those changes to v1.6?

Yep, here: [link removed]... most the changes required rearranging things a good bit. You can just skim through it and see if you want to use anything from it. I also made mmx routines for the YV12 conversions (mainly stolen from tom's filter, with some minor changes). On my system the YV12 c code runs around 210-220fps, the mmx versions average somewhere around 350fps). However, due to rounding differences, the mmx and c output are not exactly the same. The maximum difference on the Y plane is +-2 and for UV its +-1. I included an mmx parameter to disable using mmx if you want to though (mmx=false).

[Wilbert]

Great! I updated the first post. Thanks very much

[Didée]

However, due to rounding differences, the mmx and c output are not exactly the same. The maximum difference on the Y plane is +-2 and for UV its +-1.[/B]

So, which one is closer to the "exact" value? With the actual performance the filter has reached by now, I don't care for the remaining speed difference (you know my scripts ) - but the value differences could be of importance (ask jorel ... BTW, what happened to him?)

edit: forgot to say: Great work, thank you!

[Wilbert]

So, which one is closer to the "exact" value? With the actual performance the filter has reached by now, I don't care for the remaining speed difference

The c output, thus ColorMatrix(..., mmx=false).

I've done a screenshot using mmx and without mmx and I can't spot any difference between using mmx and not using mmx. Are the rounding errors just insignificant? Here are two screenshots.

MMX=false



MMX

[Wilbert]

The maximum difference on the Y plane is +-2 and for UV its +-1.

The difference is minor, you won't see it by visual inspection

[len0x]

Can this filter safely be used after resizing (for speed reasons) or better to put it before?

[len0x]

I just grabbed a bunch of DVDs I have at hand and actually only 50% of them are using default mpeg2 coefficients. There are considerable number of them that uses s170 and i470. Also I had one that uses FCC. So I guess its no good to just apply this filter to any DVD you have.

P.S. Is it possible to write a filter that returns coefficient type, so that conditional usage of ColorMatrix is possible?

[Didée]

Originally posted by Wilbert
The difference is minor, you won't see it by visual inspection

OOOoooohh. Yes you will.

interleave(last.colormatrix(mmx=false),last.colormatrix(mmx=true))

Step back and forth, and tell me you see no difference.


edited stupidity ... grrr!

[tritical]

What about having dgdecode set a global variable colormatrix could read? To make it multi instance compatible the name would need to depend on the d2v file, so colormatrix would need the name of the d2v file as input as well. Good or bad idea?

Wouldn't it be:

interleave(last.colormatrix(mmx=false),last.colormatrix(mmx=true))

for Wilbert's point .