just create a patch #151 with a perfect neutral reflectance
-
Welcome to the new forums! Please read this first. For known issues we are working to resolve, click here.
Camera color accuracy outside of the training set
- Thread starter JimKasson
- Start date
There are 190 patches already, but that’s what occurred to me. GMTA.just create a patch #151 with a perfect neutral reflectance
right, #191, my bad ... never was good with counting stuff on mondays (or ever) !There are 190 patches already, but that’s what occurred to me. GMTA.just create a patch #151 with a perfect neutral reflectance
- Error bars colored with corresponding patch colors.
- White balanced to perfect gray patch; added as 191st patch.
- Added chromaticity plot for errors: green is under 0.5 DE2000, blue is between 0.5 and 1.0, and red is over 1.0.
the last word the last word - Photography meets digital computer technology. Photography wins -- most of the time.
Photography meets digital computer technology. Photography wins -- most of the time.
blog.kasson.com
Jack Hogan
Veteran Member
Very good Jim. Average DeltaE00 of 0.7 is unheard of with current CFA dyes. What does the matrix look like?
- Error bars colored with corresponding patch colors.
- White balanced to perfect gray patch; added as 191st patch.
- Added chromaticity plot for errors: green is under 0.5 DE2000, blue is between 0.5 and 1.0, and red is over 1.0.
Jack
PS. I had a look and the containers I remembered for reflected colors: MacAdam Limits (theoretical) and Pointer's Gamut (practical). Something similar to the latter was formalized as ISO/TR 16066 : "Standard Object Colour Spectra database for color reproduction evaluation (SOCS)". It is supposedly made up of 50000+ samples as described and compared to Pointer's Gamut here. I was not able to find the database online though apparently the standard comes with a CD-ROM.
TheUnblinkingEye
Member
- Messages
- 49
- Reaction score
- 8
The first row of the chart is all memory colors; are those helpful in increasing color accuracy or could they just as easily be replace by different samples?
Forward Matrix to XYZ D50/2 =Very good Jim. Average DeltaE00 of 0.7 is unheard of with current CFA dyes. What does the matrix look like?
- Error bars colored with corresponding patch colors.
- White balanced to perfect gray patch; added as 191st patch.
- Added chromaticity plot for errors: green is under 0.5 DE2000, blue is between 0.5 and 1.0, and red is over 1.0.
0.9802 -0.1032 0.1521
0.0659 0.9870 0.0111
0.0053 -0.0026 0.8606
--
the last word the last word - Photography meets digital computer technology. Photography wins -- most of the time.
Photography meets digital computer technology. Photography wins -- most of the time.
blog.kasson.com
Better with XYZ:Very good Jim. Average DeltaE00 of 0.7 is unheard of with current CFA dyes.
- Error bars colored with corresponding patch colors.
- White balanced to perfect gray patch; added as 191st patch.
- Added chromaticity plot for errors: green is under 0.5 DE2000, blue is between 0.5 and 1.0, and red is over 1.0.
I need to go back and recompute the optimum Gaussian set. The Gaussians at the top of the page were computed to optimize the Macbeth cc24.
--
the last word the last word - Photography meets digital computer technology. Photography wins -- most of the time.
Photography meets digital computer technology. Photography wins -- most of the time.
blog.kasson.com
Here it is with a real CFA dye set.
I changed the circle coloring in the chromaticity chart so that 0 error is blue, the highest error is red, and the colors are mixed in proportion to the error.
I'm using the SMI formula, but it's not really SMI because the patch set is different. This camera gets a cc24 SMI of about 89.
set.
--
https://blog.kasson.com
Last edited:
Here's the Arri D-21 that you know well:
--
the last word the last word - Photography meets digital computer technology. Photography wins -- most of the time.
Photography meets digital computer technology. Photography wins -- most of the time.
blog.kasson.com
I am sorry - but that database for example has 30K+ colors from "graphic prints"... that is how many different pigments ? well - how about you get 20-30 real pigments spectra and then simply generate synthetic from that... no worseVery good Jim. Average DeltaE00 of 0.7 is unheard of with current CFA dyes. What does the matrix look like?
- Error bars colored with corresponding patch colors.
- White balanced to perfect gray patch; added as 191st patch.
- Added chromaticity plot for errors: green is under 0.5 DE2000, blue is between 0.5 and 1.0, and red is over 1.0.
Jack
PS. I had a look and the containers I remembered for reflected colors: MacAdam Limits (theoretical) and Pointer's Gamut (practical). Something similar to the latter was formalized as ISO/TR 16066 : "Standard Object Colour Spectra database for color reproduction evaluation (SOCS)". It is supposedly made up of 50000+ samples as described and compared to Pointer's Gamut here. I was not able to find the database online though apparently the standard comes with a CD-ROM.
Jack Hogan
Veteran Member
Why are you sorry zzip? Do you know of a good database available online?
Your comment is appropriate, the link above was included exactly so that folks could have a sense of what's in there, here is one of the relevant tables for those so inclined. Many repetitions, though there appear to be several thousand useful spectra there to add to the current much smaller mix.
Spectra part of ISO/TR 16066, for Color Reproduction Evaluation, from the link above
Although, as mentioned earlier, if we were interested in evaluating how well linear color generalizes with specific SSFs, it would be useful to set up a synthetic training set with samples evenly spaced on the surface and/or inside the solid of, say, the gamut of all natural and man-made colors (e.g. Pointer's ). Then test it with real reflectances from databases like the one above. Then do it all again for specific subsets of spectra for different applications (e.g. studio portrait, repro, landscape, ...)
Jack
Last edited:
I think the question is, what do you need colour accuracy for ?
If your application is photographing works of art, such as paintings or colour woodcuts, then the training set should include artists' pigments, especially the traditional earth colours (Siennas, Ochres, Umbers, Venetian Red, Indian Red). It should also include earth colours mixed with white, and the same painted as a pure coat and then overlaid with a white glaze (Tyndall Effect). Also the same pigments mixed with white.
For ceramics, the training set needs to include common glaze colours.
For fabrics and tapestries, a range of natural and synthetic dyes.
Do you have a good local art shop ?
Don Cox
If your application is photographing works of art, such as paintings or colour woodcuts, then the training set should include artists' pigments, especially the traditional earth colours (Siennas, Ochres, Umbers, Venetian Red, Indian Red). It should also include earth colours mixed with white, and the same painted as a pure coat and then overlaid with a white glaze (Tyndall Effect). Also the same pigments mixed with white.
For ceramics, the training set needs to include common glaze colours.
For fabrics and tapestries, a range of natural and synthetic dyes.
Do you have a good local art shop ?
Don Cox
Last edited:
I believe it is common, in fields that highly prize accurate color reproduction, to train with a patch set whose spectra approximately replicate those of the material to be digitized. If it's not, it should be.
--
https://blog.kasson.com
Last edited:
I think you put too much faith into what is a problem with very limited data. After all, those spectra are projected to a 3D space. You want to map them to another 3D space with some matrix. That matrix has 9 entries, and accounting for the luminance, there are 8 left. Whatever you do, you compute 8 numbers. You just put different weights on different spectral by playing with the training set; not necessarily the way you want. For example, if you have too many reddish pigments, you skew it to the red side.
Now, you can try a nonlinear mapping. You can make it perfect on some specific (but not all) set of spectra. You have no idea however what it does to spectra not in your training set. If you want good control over that, you might be back to a linear or an almost linear map.
probably off-topic - but I assume when you do repro of such targets you use polarized light, etc - so just whatever spectra floats somewhere might not be collected with a proper filter on the spectro-instrument ? did I say something stupid ?
Jack Hogan
Veteran Member
Explain how to drop from 9 to 8 unknowns JACS? [Edit: I got it]
Depending on usage, assuming we need to know the CCT/WP of the illuminant as we mostly do in photography, we can in practice reduce the number of unknowns to just 6 since not hitting the neutral tones correctly can result in objectionable color casts. So we mostly want the sum of the rows of the Forward Matrix for a given illuminant to be equal to its White Point coordinates in XYZ, which gets rid of three unknowns.
I normally still keep 9 unknowns in order to have the most degrees of freedom and therefore get the best fit for the available data - but I always check the White Point of the FM to make sure it is not taking off in the wrong direction. It is often surprisingly close, if not it usually means that there is something wrong.
The optimized Forward Matrix has pretty well the same CCT White Point as the Illuminant
Jack
PS From here https://www.strollswithmydog.com/open-raspberry-pi-high-quality-camera-raw/
Last edited:
If actual paint pigments are used, the spectra should be very close.
Winsor and Newton have sample cards made from actual paints (not from CMYK simulations). These are not easy to get hold of, but if you are a good customer of your local art shop, they might be able to get you one.
Back in the 1970s, somebody with access to a spectrometer made me some curves from paint samples. Unfortunately I no longer have his original curves, but I do have the tidied up versions that I made for lecture slides.
A set for recording colours of tomatoes for advice to farmers would no doubt need more reds, oranges, and yellows. and few samples of blues.
Don
Last edited:
I think the polarizing filter (if used) would in effect modify the colour of the lighting. So it's a matter of white balance.
Don
It's more than white balance. Specularity produces desaturation. If I were making a profile to be used for polarized or cross-polarized light, I'd photograph the training set under the same light.
Keyboard shortcuts
- f
- Forum