New project: camera color profile calculation

lks

I forgot that you did not take the calibration photo in daylight -- I'm guessing the ACR calibration program works best on a color checker chart shot in daylight.

At any rate, the JPG/CRW for my calibration is at dngpsrawin.zip - 10.24MB.

Now that I know it is possible to get 22 areas, I will try some other time to take a better calibration picture. My guess is that one has to push the exposure to the right enough so that the white patch is 245+ in intensity. I can just bracket a few exposures and see if any work better.

lukg

lks
I forgot that you did not take the calibration photo in daylight -- I'm guessing the ACR calibration program works best on a color checker chart shot in daylight.

Thanks for your files.
The problem for me is that ACR calibration script expects to compare colors taken by an uncalibrated source with colors of real gretag charts (those colors are fixed and known). But without a real gretag is useless (we are lucky you have one), because the colors of my palette shot depends on two variables: colors fidelity of my LCD, and color fidelity in capturing of my camera. So I have 2 non-calibrated sources mixed together (actually we have another variable we must face to: the ACR calibration is done after we apply matrix to DNG. If the matrix is not good, with the ACR calibration we obtain optimal settings for that image and nothing else, and that's pretty useless too)
For what I saw in your JPG, the camera is already good calibrated by Canon. So ACR calibration is not our main problem (since our DNG are not "bad calibrated" from Canon, they're not calibrated at all). The problem is finding a perfect transformation matrix, but as I already said, I'm worried that there are millions of good matrix, but only few are good for all the pictures made by that camera. :'(

lukg

Ok, I did some crazy experiments. I took 10 pictures (-1.1/3exp) in continuos mode. Averaged RAWs on camera, and jpgs using Photomatix. This is the results of the first calibration (at 1.00 brighness):

RESULTS
White area: 1
Errors:
Zone 1: 11,8%
Zone 2: 10,6%
Zone 3: 11,4%
Zone 4: 11,8%
Zone 5: 8,5%
Zone 6: 11,3%
Zone 7: 9,4%
Zone 8: 5,1%
Zone 9: 5,7%
Zone 10: 11,8%
Zone 11: 10,5%
Zone 12: 3,5%
Zone 13: 11,6%
Zone 14: 12,3%
Zone 15: 9,6%
Zone 16: 6,4%
Zone 17: 6,9%
Zone 18: 8,7%
Zone 19: 9,7%
Zone 20: 11,2%
Zone 21: 7,5%
Zone 22: 4,2%
Zone 23: 9,2%
Zone 24: 11,8%
### Max error is greater then 10%, RESULTS ARE POOR ###
Camera -> XYZ matrix:
1,829391 0,412710 0,792508
0,235154 1,880082 -0,035845
0,024875 -0,741887 5,160420
XYZ -> Camera matrix:
0,568353 -0,159643 -0,088393
-0,071335 0,553391 0,014799
-0,012995 0,080328 0,196336

All 24 patches found. That's a good start i think.

lukg

I have an idea for Artden, after I saw how the ACR calibration script works.
I don't know what are the mathematic calculations that are behind dng4ps, but it should be usefull in my opinion, having the possibility to choose a weight value to assing for each areas in the calibration process (like ACR script).
Here the screenshot of what i am talking about:

ACR Calibrator L

I think it could be usefull in finding a good compromise between minimizing the total color error (all color of the palette = same importance) and what color are "more" important than other ones (eg. red, green, blue patches).

Is it an insane idea?

lukg

Thanks Matt, what you think about my latest matrix with 24 patches found. It's the best matrix I used till now (still a blue dominance, but I noticed that in all my calibrations).
Anyway, are you saying that it is "impossible" (and useless) trying to get the DNG closer to the jpg (in normal situation) because of tone mapping?

lukg

ok ok! Thanks Matt for your replies!
My averaging experiment was meant to reduce the overall standard deviation, since I saw that in calibration picture the noise was still there even at 80iso. The purpose was to obtain uniform colors and help the program in the calculation (since I think the program calculates mean in each area).
Averaging is not a real manipulation, since it applies a simply math mean to every pixel obtained from the 10 identical pictures taken (in the jpg and in the raw too of course).
I think this is a better way to take a calibration picture cause a mean is surely "more" realistic/reliable of a single value (noise is variable).
Since I suppose (I hope) that averaging is a "linear" operation, i don't think that this could influence negatively on the results, since it is applied on both raw and jpg in the same way. At the contrary it gives me very good results (compared to previous matrix).

Here a comparison between the DNG and the JPG of the camera (white point obtained picking up the bottom-left square)

I am sure that there will be a better matrix outside, but this work is driving me crazy 😁 😁

lukg

mmm I understand what you mean....I don't know how CHDK does the average (maybe it's just a RAWsum/n.picture) 😀
...but for the jpg I did it by hand in photoshop and compared with the one I did in photomatix. I think averaging is a simple "arithmetic/geometric/harmonic mean" operation, I hope it is a "standardized" operation 😁

The results look OK so maybe the end justifies the means, but I'm still think someone should check we aren't introducing suble issues into the matrix.

I agree with you..for the moment I'll use my new matrix, since I have nothing better. 😁

lks

The new matrix is quite good in daylight conditions. Significantly better than any of the past matrices, including the 720is.

In lower temperature light (2800 to 4000 degrees) the colors and white balance of the new matrix get out of whack. I will try again to make a calibration under Tungsten light -- it hasn't worked well the first few times.

lukg

lks
The new matrix is quite good in daylight conditions. Significantly better than any of the past matrices, including the 720is.

In lower temperature light (2800 to 4000 degrees) the colors and white balance of the new matrix get out of whack. I will try again to make a calibration under Tungsten light -- it hasn't worked well the first few times.

Glad that you find the matrix good! I agree that we need a matrix for lower temperature!

Sess

Let me see if I understand the purpose of this project. By comparing the raw and jpg, you can figure out how each camera interprets the colors in the raws and converts them to jpg. With this information, you can make DNG4PS-2 better interpret the colors in the raw files when it creates a dng. Is that right?

The big downside for me using raw files is that when I open the subsequent dng files in Lightroom or ACR, they auto pick a white balance that is way off from what the camera chose. Would calibration help programs like Lightroom and ACR pick a better white balance when they open these dng files?

oversky

Sess, lightroom can choose WB as shoot, tone curve as linear.

Sess

oversky
Sess, lightroom can choose WB as shoot, tone curve as linear.

The CHDK raw files have no white balance info and neither do jpgs used when making dng files. So when you set the white balance in Lightroom to "As shot" it is really just doing "Auto" and making its best guess based on the photo. The problem is the best guess in Lightroom for the raws from my A650 is always way off from what the camera chose for the jpgs.

From reading this thread it sounds like DNG4PS-2 needs to be properly calibrated for each camera, otherwise the colors in the dng files it makes will be skewed. In other words, Lighroom is choosing the correct white balance for the photo it is given, it's just that the photo it is given has skewed colors to begin with.

lukg

Hi Dawmatt, I'm using this matrix:

Camera -> XYZ matrix:
1,943902 0,442380 0,815924
0,215457 2,022005 -0,092312
0,021941 -0,780416 5,443674
XYZ -> Camera matrix:
0,531769 -0,148074 -0,082215
-0,057135 0,513726 0,017275
-0,010334 0,074246 0,186507

It's surely better than the one included in dng4ps2, simply because there's no S5 matrix in DNG4PS2 yet (only S5is label but the matrix from a different camera afaik) 😁

ArtDen

lukg, 0.2.2beta-26 contains test matrix for S5. Please test it and tell which matrix is better (mine or yours)

lukg

Ok, thank you! I'm going to test it.

lukg

I did a quick test using my sample images, and the matrix in dng4ps2 is reddish for almost all the pictures.

Here are 2 examples.

Right = my matrix
Center = dng4ps2 matrix
Left = jpg

http://img410.imageshack.us/img410/1789/catturamv9.jpg (look at the wood)

http://img410.imageshack.us/img410/3664/cattura2mw6.jpg (here all is reddish)

ArtDen

Have I include your matrix into dng4ps2?

lukg

That would be good for me! Thanks!

ArtDen

Done: zSHARE - dng4ps2-0.2.2beta-26.exe

lukg

Thank you, Artden! Very appreciated!

« Previous Page Next Page »