DCamProf - a new camera profiling tool

[AlterEgo]

Nope, not that way

8 EV - log2(13193.07/55.471) for G - about 0.11 EV

that is if you are using your approach of bracketing shots and combining target measurements with data (spectral or otherwise) adjusted to match underexposed raws... than naturally means using the data from D1 from the EV0 and the data for D6 from EV-3...


then measured target will be (using Y from spectral data divided by 2^3) log2( 92.588 / 0.391 ) = 7.887512805 and then is not 0.11EV, but ~zero, no ? unless you state that i1pro2 can't even measure ColorChecker Passport black patch (not CCSG) correctly...

[Iliah]

> then measured target will be (using Y from spectral data divided by 2^3) log2( 92.588 / 0.391 )

I disagree with 0.391. Measured Passport with my i1Pro, Y on black patch is 0.31

[AlterEgo]

> then measured target will be (using Y from spectral data divided by 2^3) log2( 92.588 / 0.391 )

I disagree with 0.391. Measured Passport with my i1Pro, Y on black patch is 0.31

you mean you disagree not with Y = 0.391, but with Y = 3.128 from measurements ( 0.391 was derived from Y = 3.128 when I divided it by 8 = 2^3 stops to match EV-3 exposure )... well , I averaged 3 measurements a week ago (attached) - but I will repeat it again.

[Iliah]

I have Y = 94.79 for the white, Y = 3.1 for black, log2(94.79/3.1)+3 = 7.9344
However, in practical photography any variation below 1/6 EV (0.17 EV) is considered irrelevant. The best shutters are guaranteed to be within 1/12 EV, apertures within 1/10 EV accuracy and repeatability.

[AlterEgo]

The best shutters are guaranteed to be within 1/12 EV

by the way - it sounds logical that exposure time component of exposure shall be more precise with longer exposures, no ?

[torger]

Just released a quick update 0.7.3 out to fix some bad bugs introduced in 0.7.2

[AlterEgo]

0.7.3 build for Windows (mingw = dcamprof.exe + libgomp_64-1.dll) : https://app.box.com/s/bh3zagil9imxff5h3voxozgtgcwzmupa

[torger]

Been working with some elaborate workflows, seems to be one feature left to do before I can get profiles that look as they should; somehow one need to take the film curve into account. The commercial profiles do, and even some matrix-only profiles seems to be deliberately desaturated to not oversaturate when a film curve is applied. On the other hand it may just be a side-effect of using only matte targets when profiling which typically seems to lead to a slightly under-saturated profile.

This has nothing to do with reproduction style profiles which are designed to look accurate with no film curve applied. Not sure if it's suitable to rely on standard color science models at all in this case.

I think one can let highlights desaturate with the lowered S-curve contrast, but maybe one want to keep saturation in shadows, if so I need to introduce a 3D look LUT. Possibly it will look good if just compensating for the increased midtone contrast all over, that's probably what I will try first.

Adobe has a slightly different curve formula than a pure RGB curve, may need to take that into account too. If we only look at linear midtone contrast we won't though as it then is no different from a pure RGB curve.

[nertog]

Some measured spectra of the Yuji LEDs and LEDMOTIVE multichannel source:

1. High CRI blue-pumped white LED (CCT = 5800K)



2. Violet-pumped white LED (CCT = 5000K)



3. Violet-pumped LED, different phosphors (CCT = 5600K)



4. LEDMOTIVE source emulating D65 spectrum (CCT = 6504K)




Hope this helps!
Wim

[torger]

Thanks! They're not really UV-pumped it seems, but on the edge of the visible range, where both the eye and a camera has some sensitivity left. Not sure how much it will disturb though.

[AlterEgo]

manual has a small typo error - here is a piece of text :

Here follows a description of each command available.
make-target

  dcamprof make-profile <flags, with inputs> <output.ti3>

Make a target file which contains raw camera RGB values paired with reference XYZ values, and (optionally) spectral reflectance. The file format is Argyll's .ti3, with some DCamProf extensions.

[Iliah]

> They're not really UV-pumped it seems, but on the edge of the visible range, where both the eye and a camera has some sensitivity left. Not sure how much it will disturb though.

If it does not cause response in the red channel it does not matter much, not adding to metameric failures. But at least with Nikon cameras based on Toshiba sensors it does.

[GWGill]

Thanks! They're not really UV-pumped it seems, but on the edge of the visible range, where both the eye and a camera has some sensitivity left.

Might make for more fluorescent whitener response than a "real" light source.

[torger]

Experimented with some different methods to adjust saturation when using the profile with a film curve. Adjusting chroma of the XYZ reference values via CIECAM02 Jab/JCh seems to work well. JCh is more linear concerning hue (that is less problems of changing hue when chroma is scaled) and overall more perceptually uniform than Lab/LCh, and a suitable adjustment seems to be in the 1-2 units range, that is very small, so hues stay stable enough. I was thinking of making a Munsell corrected space (like Bruce Lindbloom's UP Lab http://www.brucelindbloom.com/index.html?UPLab.html), but as adjustments are so small and often with more saturated colors than exist in Munsell's book of colors it's not worthwhile, and as said Jab is already a bit better than Lab.

Instead of adjusting the XYZ reference values the other alternative is to apply the curve on the RGB values, but then matrix fitting becomes badly nonlinear and I don't think the result is better. Just like the curve itself this type of adjustment is more about look-and-feel than accuracy.

I'll include possibility to adjust chroma it in the next release, and voilà DCamProf has its first subjective look-and-feel adjustment.

This is a good read regarding perceptually uniformity of various color models by the way: http://www.handprint.com/HP/WCL/color7.html

[AlterEgo]

any ideas how to come up with some good heuristic to speed up (or rather try in a proper directions) trying possible combinations within 28 measured SSFs to "calculate" the best "virtual" SSF for my camera ? a simple brute force is too slow because calling dcamprof executable hundreds of thousands times (to test how SSF profile generated by Matlab matches /dcamprof test-profile/ the real measured and shot target) is not an option...

so far brute force (I can only try all combo up to 5 cameras out of 28 on my notebook in several hours) gives me the around the following match :



not bad, but I want better.

dcamprof test-profile ("dcamprof test-profile -i ill.sp -C -B ccpp.ti3 ssf-profile.json", where ccpp.ti3 has real/actual raw RGB and real/actual measurements) gives for this profile :

Native LUT patch match
average DE 0.67, DE LCh 0.25 0.35 0.44
median DE 0.62, DE LCh 0.27 0.39 0.38
p90 DE 1.24, DE LCh 0.45 0.60 0.95
max DE 1.31, DE LCh 0.68 0.79 1.23

results of conversion in ACR (process 2010, etc, with WB off D2 patch ) are naturally worse than math and not matching dcamprof test-profile 1:1

[torger]

Nelder-Mead simplex can do such stuff, but it requires lots of finetuning of error function and start values to get good results.

It would be interesting to integrate an SSF estimation function into DCamProf but it's probably not happening for at least a few months. I'm planning to code less during the summer which is starting here now :-)

[AlterEgo]

Nelder-Mead simplex can do such stuff, but it requires lots of finetuning of error function and start values to get good results.

It would be interesting to integrate an SSF estimation function into DCamProf but it's probably not happening for at least a few months. I'm planning to code less during the summer which is starting here now :-)

what might be the optimization parameters, I am using max dE as the top one, then with max dE being equal I am thinking shall I consider an average/mean dE or go for maximum chromacity error (not paying attention to luminocity errors) ?

[torger]

what might be the optimization parameters, I am using max dE as the top one, then with max dE being equal I am thinking shall I consider an average/mean dE or go for maximum chromacity error (not paying attention to luminocity errors) ?

For Nelder-Mead to find a good minimum it needs a quite "well-behaving" error function, if it's too random it's hit and miss. That is if it changes a parameter it must see some sort of trend if it got better. Using max dE may be worse than mean for this reason. I don't have any general good guide though, trial and error is the method I've used.

[AlterEgo]

How could I possibly know the chances?
This thing is important:
How do I reset my Spectrolino?

To reset your Spectrolino, hold down the measurement button on the top of the instrument for approximately 10 seconds. During this time you will hear a sequence of four tones, then two tones in rapid succession that are repeated after a pause. Continue holding the measure button until you hear a fifth long tone, indicating that the reset is complete.

so I shall find out tonight... whether I can measure the black patches on CCSG or not

[torger]

Just released 0.7.4. A very small update adding the subjective chroma adjustment possibility.