Hi John, excellent video! I also like the sense of humor, very important when talking about CIELAB :)) On the other hand, could you, perhaps, review, give us a lecture on OKLAB, this other color space? I would appreciate it, thanksss
Thank you. I'm glad it was appreciated. My quick answer about OKLab is that is lives up to its name! It fixes some of the problems with CIELAB by being based on the cones rather than the silly 1931 Standard Observer. It is nice in that it is very simple. For image processing applications, where operation count can be important, it makes sense. (Or maybe not... processors are pretty darn fast.) The big disadvantage is that they chose to use the square root as the nonlinearity function. Munsell initially tried this in 1905 and it failed. His son improved on that in the 1920s. As a result, when it comes to predicting people's perception of the difference between two colors (deltaE), it is basically on par with CIELAB. CAM16-UCS is currently the most accurate predictor of color difference, but it is on the end end of spectrum in terms of complexity. It has all the fiddly knobs that a researcher could want, but for evaluating of color in industry, a lot of those knobs just need to be nailed down. At this time, I don't know for sure what the best answer is/will be.
@@JohnSeymourTheMathGuy Hi John! Thank you very much for your detailed reply, it answers my query to a good extent, thank you! Do you think it would be possible to continue my question by email? I would write to you at johnthemathguy, is that ok?
You went into great detail about DE:76 and DE:CMC; but didn't explain how DE00 was created or what it is intended to improve over prior tolerance standards. The way it is explained in the industry sounds very similar to CMC, but besides a more complicated formula I'm assuming there are other benefits. Can you do a lecture for youtube on it at some point? Based on what you showed CMC is using the a linear hue angle like CIELAB in the formula, does DE:OO use a hue angle more aligned to a human observer?
I will provide a quick explanation here, and ponder a more complete one later. DE00 and DECMC are similar in performance. DE00 edges the other formulas out when looking against the official data sets. In my humble opinion, the improvement is marginal. This was not intentional, but the way the math works, DECMC kind of assumes that one of the colors is the target color and the other is the potential match. The difference between color 1 and color 2 is not necessarily exactly the same as the difference between color 2 and color 1. DE00 fixed this. The color difference data sets showed some weird behavior in the blue region. As a result, they developers added a band-aid. All of the ellipses in DECMC are oriented with the long axis along the hue line. One set of color difference observations showed a tilted ellipse in the blue, so DE00 added a fudge factor to fix this. (In my opinion, they messed up. I predict that there will be similar tilts in the red/orange region and somewhere in the yellow/green. These just haven't shown up in test data. The whole thing could have been fixed by fixing CIELAB rather than fixing color differences. In DECMC, the tolerance regions are true ellipsoids. In DE00, they are ovoids. They are slightly fatter at the high chroma end.
The L* value of swatch #3 in the Kodak Gray Scale chart is L*50? I'd suspect swatch M matches the 'human visual perception' of a true mid-tone, or L*50.
Video and photographers are allowed to deal with Lab intensively. A good example is the software product 3D LUT Creator. In the training video on YT titled "HSP and LAB color models in 3D LUT Creator. Part 1", about 3 minutes in, it is shown that the primary colors (RGB, CMY) in Photoshop make "loops" in the Lab space when you change their brightness, which would correspond to opening or closing an aperture. What you would want are straight lines parallel to L-axis and not loops. That's why the authors of 3D LUT Creator introduced the HSP color space to compensate for this weakness. John, what do you think about this? Wasn't that also one of your tirades?
That's a good question, but I don't feel qualified to answer it! I don't have enough experience with HSP to know if it did fix the problem. I am inclined to think that throwing more equations at XYZ is simply added a band-aid to a pimple on a wart. If HSP goes back to LMS, then they are heading in what I think would be the right direction. I welcome anyone else to offer a more informed opinion.
This is very entertaining...do I have a problem? so...a 1 unit of Delta E 2000 is still the slightest amount that human eyes could decipher. Is that correct? A delta E 2000 at 2.0 among 90% of my color patches is...good? sorry psudo color enthusiast here
A subtle distinction here ... The original work by MacAdam determined "just noticeable differences" (JND). Basically, his test subject would adjust a knob to get a color to match another again and again and again. His JND is a standard deviation of how well the test subject agreed with himself over many trials. Most of the data used to tweak the DE equations is based on human assessments of acceptability of a color match, generally among people who do this for a living. The two are obviously related, although I am not aware of a rigorous test of the relationship. One color scientist told me that 1 DE 2000 was about equal to 0.7 JND. Is 2 DE 2000 "good"? Depends on your customer and what colors you are matching. For paint in a bucket, that might not be so good. For printed stuff, I think that would generally be pretty good. But... I defer to anyone else who might like to share an opinion!
Hi John, hope you are doing well.. great video! John, I have one question regarding colour differences in offset printing though. Is it valid to say that Delta E 1 tolerance is indeed capable of delivering a Delta E 2 between repeated jobs? If I print one day a colour which is Delta E 1 from my target and the next day I print this job again and I have again a Delta E 1 deviation but in this case my print reading is pointing at the bottom end of the ellipse and the first run was pointing to the top of the ellipse.. so I am Delta E 2 from these two press runs, which yes they are a PASS for the machine BUT between run, they look unacceptable. Is there a way to avoid this? how one can communicate this in a press room? Can I match the top/bottom half of the circle only in the spectrophotometer? If that's the case, it might need to vary according to the colour I am printing which can be an absolute hassle for printers and everyone involved.. What's your take on this problem? Thanks in advance!
Yes. If one sheet is 1 DE too light, and another is 1 DE too dark, then both will be in tolerance, but they will be 2 DE apart. That's the worst case... generally the differences between two sheets will be smaller. Is there a way around this? Ummm... you could make your production tolerance 0.5 DE. (Yes, I know. Stupid answer!) One approach that the CGATS folks were advocating is kind of a "creeping target" solution. The brand owner (or whoever) decides on a target color. You get to press, and do your best at color OK. You are not exactly at the target, but you are in tolerance. The color at the press OK is now taken as the target for the rest of the run. Now you run into the same sort of issue as you mentioned. The press OK sheet is in tolerance, and the goalposts were moved (by the ref!) so that the new tolerance range is centered on the OK sheet And you can have this 1 DE + 1 DE issue. I kinda think this is hokey, but that's just my opinion. Personally, I like to think of the original target color as being the target color. The goal at press should be to match the original target color, and not the press OK sheet. Does this help?
Thanks John, yes it's a difficult task, to ask a printer or a different shift printer to match the original target and at the same time to match the previous run! It wouldn't be a big issue if tolerance is Delta E 1.. but we are working with delta E 2 for spots colours.. so difference between runs can be really noticeable. Thanks anyway, I guess it's one of those things that we need to live with..
Actually... it would be pretentious to point out that when you compare two data sets (each with one member) the Mahalanobis distance calculation is undefined. Recall that variance and covariance are calculated by first subtracting the mean from the data. This gives a variance-covariance matrix which is all zero, and cannot be inverted. If I were really pretentious, I might point out that Hotelling's T-square statistic is the square of the Mahalanobis distance, since this doesn't seem to be commented on very often. Glad to see you watched the whole video!
