Six things you should not do with CIELAB, Seymour, TAGA 2022 

1) Weren't approximations to the cone functions established later? Yes, the HPE transform (and other similar ones) is based on matrix transforms from XYZ to LMS, so naturally, they came after. It was my assumption that the folks involved in creating the 1931 Standard Observer just didn't have a good guess at what the cone functions were. Then I came upon a document from Troland in 1922 (L. T. Troland, "Report of Committee on Colorimetry for 1920-21," J. Opt. Soc. Am. 6, 527-596 (1922)) that had a pretty darn good guess at the cone functions. The data for these plots date back to the 1890s with work by Koenig and Dieterici to find what was called the elementary functions. I have found papers from Ives (1915), Schrodinger (1920), Adams (1923), and Hecht (1930) that give similar plots. I find it hard to believe that Guild and Wright were unaware of these. Guild certainly was. Perhaps G & W didn't think that they were really the cone functions? In 1924, Guild said "Whether these “fundamental sensations” have any real existence in the processes of vision is immaterial to the present discussion..." Hecht (1930) was also noncommittal: “Do these three excitation curves really represent the physiological primaries of Thomas Young? Are they a description of the spectral properties of three fundamental processes concerned with the reception of color? There is a tacit assumption among us that they are, though most of us would never acknowledge this publicly.” My own conclusion is that the 1931 decision was based on putting on the blinders and trying to answer the immediate question of how to standardize the tristimulus color matching devices, rather than stepping back and saying "maybe we should try to emulate the human visual system?" 2) Isn't CIECAM based on XYZ? According to theory, the cone functions are a linear combination of the color matching functions (e.g. XYZ). A simple 3X 3 matrix transform will go from one to the other. The general approach for the color appearance models is to first do that linear transform to get cone responses. After that all the nonlinear stuff can be done. CIELAB fails in that the nonlinear stuff (normalizing to illuminant and application of f ()) are done to XYZ values.

There are many efforts by lots of people. I think there needs to be one concerted effort by a group of people!

Thanks for the comment, Adrien. I don't think there is a need for four separate color coordinates. Yes, the Y tristimulus value is the definition of "luminance". When the Standard Observer was defined in 1931, one of the criteria was to use the CIE 1924 luminosity function as one of the three functions. This was based on work by Gibson and Tyndal of the Bureau of Standards in 1923. They compiled data from a number of experiments as well as their own. I ponder whether the conditions of the experiment can be generalized to our normal perception. Part of that pondering comes from consideration of the Helmholtz-Kolrausch effect, which (in layman's terms) says that saturated colors (with the exception of yellow) appear "brighter than expected". I am not sure exactly how one defines how luminous a saturated red is expected to look, but to me the effect might be an example of how the model for luminance is in need of fixing. The point I am making is that, while Y is defined as "luminance" and L* is defined as "lightness", neither one of them work all that well when it comes to saturated colors. Yes, the Y function is different than the M cone response. In my work, I have used a linear combination of L, M, and S to create the analog to Y. I used a weighting that (if I recall correctly) has 0.50 of M, 0.45 of L, and 0.05 of S. One question in my mind is whether to apply the nonlinear function to the weighted sum of the three or to each of them individual ones. I decided that I would set that question aside and try to fry the fish one at a time! It is my impression that most, if not all, modern color appearance models do not use M as the luminance function, but rather, some combination of L, M, and S. The folks working of color appearance models are also well aware of the Helmholtz Kolrausch effect. A few weeks ago, I heard a presentation from Luke Hellwig on the topic. His PhD research is on the ways to model the H-K effect in a color appearance model.

@@JohnSeymourTheMathGuy Thanks, I'll look at Luke Hellwig's work!

Very reasonable question! My own stubborn and narrow-minded approach is to continue plodding forward with my own math. :) There have been a number of LMS color spaces developed, with the various CIECAM models having the most press. Technically, these are "color appearance models", which means that they take into account the adaptive state of the eye, the intensity of the illumination, and the surrounding colors. All very worthy things if you want to understand how people see color. My gut feel is that this model might be overkill for industrial uses. All that extra capability may make it too hard/confusing. Or maybe that could be helpful? I am hoping that there can be discussions among a diverse collection of industry experts to help answer that question.

@Pranesh Kumar Tough questions... at least for me! Let me start by making it clear. To put my response in context, I claim expert-level knowledge in "what is wrong with CIELAB", but I do not claim expert-level knowledge in the alternatives. My one claim is that basing the color space on LMS rather than XYZ would solve many of the problems with CIELAB. So, any reasonable alternative must be based on LMS. I **think** that CIECAM16 might be the best approach available today, but I have not investigated it in depth. The CIECAM series includes a number of knobs for adjusting to lighting conditions and adaptive state of the user. This is obviously a good thing for the researcher and others who need the most accurate models. This is a issue, on the other hand, for the color manufacturing industry, since it adds the complication of "which knobs should I use". In my somewhat limited view of the world, I would look the Mark Fairchild as the guy. While not an easy read, his book "Color Appearance Models" is as readable as any that I have seen.

I would say that a serious improvement to CIELAB is not for the faint of heart. I have spent a decade or so, in my spare time, trying to understand the causes of the foibles of CIELAB in order for them to be properly compensated for. I do not have anything quite yet that I am ready to recommend. And it what I develop may not substantially differ from CIECAM with a set of fixed parameters.

@@JohnSeymourTheMathGuy surely it would make more sense -rather than focus on the foibles of CIELAB to go back to first principles and make direct use of the present knowledge of the cone functions and modern computational techniques to develop a colour space that meets the needs of modern emissive devices, and digital (computational) devices. Build a better bike, rather than poke holes in the tires of the old bike. The old tristimulus bike’s primary utility in this process would merely be to aid in avoiding the same pitfalls. But if the new colour model accurately centric manipulation of colour. I would like to suggest the mechanism of adoption of the better colour model bike is not through the established, somewhat stagnant print industry with billions in sunk capital and industry training, but rather within the motion picture and broadcast industries (including web, and gaming industries which are based around emissive display technologies). Innovation within these industries - especially motion picture (broadcast) and gaming industries is such that consumer content and hardware purchased a decade ago is barely/ usable today (unlike the print industry in which the mechanism of reproduction has barely changed in half a century, with modern presses often being nothing more, than a bolt on unit and ink refresh added to an 80 year old press. A motion picture camera from 2012 is barely useable in a modern postproduction workflow in the absence of a cobbled together hardware/software from several generations. Innovation is baked into the motion picture industry, and I guarantee that next years blockbuster that wins the awards will not use the hardware used by todays. The post production software will likewise be radically upgraded to accommodate the camera’s capabilities, and will indeed be a generation a head of the capabilities of consumer display technologies, but will adapt the camera vision as best it can to the needs of both the director, and the consumer. Those colour professionals in the post production phase actively bounce between different colour space models (at different levels of skill and experience) in an attempt to translate the directors vision, and the media to what can be viewed on consumer displays, with the least loss of perceptual values. They are very open to better colour models that minimise the effort required to creatively express the vision, via better math. The industry is in the process for instance of adopting three primaries colour models in the processing pipeline that encompass colours outside the visible spectrum just to be able to better recompress the workspace to consumption. But of course in the workspace this wider gamut is not visible to the colourist, and the colourist must needs manipulate colour within models that can not be fully reproduced on their screens, or within their vision. However, they are well aware that some 3 suit one model rather than another. CIELAB is a model that some suggest is best used as the basis for mapping a coloured image to greyscale, so as to achieve a more natural representation of luminance. Lab is also used to generate colours in vfx compositions. But this pretty much relies on the assumption of D65 illumination, and limits the delivered product to consumers, and much of the colourists job is doing workarounds so as to trick the consumers brain into seeing colourfulness and contrast in the way the director intended within the frame of variations of RGB delivered on the basis of D65, and relatively low emissive light levels. This of course is changing rapidly, but the constraints remain. A better colour space model that in the short term enables more efficient colour “correction” will facilitate wider adoption as an industry standard in the motion picture /broadcast/web/gaming industries far far sonnet than in the print industry which will continue to use CMYK and Pantone (maybe not Pantone now that it’s use is moved to a rental model) for the our foreseeable lifetime. I suggest refocusing from documenting holes in CIELAB, to building a model based on cone values. One whose math can be efficiently linearly mapped to the wide gamut colour spaces used in ACES and DaVinci Wide Gamut workflows. Having better (even if still imperfect) human centric colourspace that’s actually works correctly, adopted by the motion picture industry would deliver the fastest uptake, and widespread critique imaginable, leading to further development. Especially if it could be adopted into contemporary display technologies with little industry effort to sell more consumer units. A royalty on consumer devices of 1 cent per thousand consumer units would make you a very wealthy man in your retirement. I suggest collaboration with the colour scientists within the motion picture/broadcast industries who are developing DaVinci Wide Gamut (Blackmagic Design) and ACES (The Potion Picture Academy) -they are both actively engaged with the development of next years hardware and software that will set the standard, and all of those in that industry are obsessed with delivering better colour reproduction to consumers, and it is they who are leading the broader consumer manufacturing industry to higher standards. Meanwhile the print industry, is merely obsessed with maximising per page profits, by reducing the cost of the inks, and “papers” having settled on a long since standardised hardware regime. Any improvements in colour reproduction in the print industry in recent decades has been on the back of the developments in film, and digital technologies used in the motion picture industries, but knee capped by the limitations of delivering to a physical medium which hasn’t much changed for centuries.

@@johnprivate6625 I appreciate the insight into an industry that I have little expertise in!

Congratulations on being the ninth person (so far) to bring this to my attention! :) Thanks for sharing. A great deal of my life in color science has dealt with the practical question in color manufacturing processes of "is there a good match?" I routinely deal with small color differences. In the article, they are dealing with huge color differences -- is there a large contrast between these two colors when seen side-by-side? So, I don't consider myself an expert on huge color differences! I know that they are not easy to measure. With small differences, the eye/brain can kinda gauge whether two colors are indistinguishable, very, very close, or just very close. With color differences about 10 DE, it's really hard to tell. (My interpretation) The article says that the two measures of distance are incompatible, which is interesting. If you build a grayscale ramp with 20 steps each of 2 deltaE2000, the perceived difference between the first and last won't be 40 DE, but something smaller. Or maybe something bigger? In 1967, the Optical Society of America got together, and looked at the available data. They said it was bigger. (Now it get complicated) They tried to fill a slice of color space with the same lightness (a plane) with ellipses that were all the same perceptual distance from the center. They came to the conclusion that it can't be done. (More complicated now) Deanne Judd likened it to a fan. In order to get the ellipses of tolerance into a plane, you had to fold the plane into one of those oriental fans. If you flatten out the fan, you don't get a circle - you get a "circle" with more than 360 degrees. I have had a superficial look at more recent data, and I don't think there is a problem -- when it comes to making tolerances ellipses fit. I admit to not paying the fee to read the paper, so all I have to go on is the abstract. If I understand correctly, they are saying that the "fan", when laid out, is less than 360 degrees. Maybe the space of the plane is a cone? In my somewhat ignorant opinion, their research is 1) not ready for prime time, and 2) not directly related to what I do. Will it effect CIELAB in the future? Likely not in my lifetime.

@@JohnSeymourTheMathGuy thank you very much for the explanation, very detailed. For sure I will read it other three times to better understand the explanation. If I get that paper I will contact you. About evaluating color small differences: today I went to a client new business to "sample" the main material colors used by the Interior Designer, to use them in the Art Direction project. Premise, no spectrophotometer or spectroradiometer was used, no fancy stuff, unfortunately. I sampled some colours using a "mix" of struments that, thanks to the work of Color Scientists, are provided to us mere mortals. A bit of trickery, I've: - used some thick white paper as background, I couldn't bring there a photobox to isolate the context (also I need to paint some surfaces with neutral response ink to these purposes); - illuminated the context with a *odox M1 light at 5000K (97 CRI and TLCI); - used a *antone color match card to capture the colours and a pretty recent *antone coated paper palette to verify the color and compare with the original material. Under the 5000K light 3 of 5 results were pretty close, considering that I had also wood with veins to sample (so the final colour was an approximation). I don't know the Color Science behind the *antone color match card (maybe cause it's proprietary, duh - indeed I hope that's not CIELAB) but it's a great little device to us Visual Designers. I will make some tests to see how it works under different colour temperatures, and color contexts. Unfortunately I don't own, so far, a spectrophotometer so I can't measure how the *odox M1 SPD changes with color temperature changes. But I will make this experiment anyway. I would ask you if a light that is at 5000K or 6500K can be named as a D50 or D65, respectively, illuminant or the latters are a specific characterized illuminant, with specific characteristics? Thank you again.

@@rafograph854 1) There isn't a lot of color science behind the Pantone books.... Well, it took a bit of deep math to do the original formulations and it takes a bit to make sure the color is right, but the organization of the book is a haphazard accretion over the decades. 2) Some definitions here: The Planckian locus is the collection of all spectra that can be emitted by a mass that is emitting light strictly because it is really hot. Positions on the Planckian locus are designated by how hot something has to be to create that spectrum. They are always measured in Kelvin, which is 273.16 degrees different from Celsius. 5000 degrees K is equal to about 4,727 degrees C. (The term 5000K is ambiguous, cuz K could mean 1,000 or it could mean degrees Kelvin. That's why we say D50 and D65.) A light can have a "correlated color temperature" of 5000K. This means that it the closest *color* on the Planckian locus is a mass at 5,000 degrees Kelvin. It could be an unnatural shade of green and still have a correlated color temperature. Note the use of the word "color". The spectrum can be whatever, so long as the color of the light looks something like D50. This is normally what we mean when we say a light is 5000 degrees K. D50 and D65 are further tighter definitions. The designation is only applied if the correlated color temperature is sufficiently close to the Planckian locus, and if some other specs are met. I don't recall them specifically (they have changed in the past ten years), but they can be found in ISO 3664.

@@JohnSeymourTheMathGuy thank you very much for the detailed information. Best regards

Hi Prof. John, Thank you for the fascinating talk! I'm an AI researcher working on building cogntively plausible models of human visual perception. Most of our work so far has been focused on using geometric information like point clouds for our perception algorithms. However, it's apparent that color is a huge element of our visual world. I am just now stepping outside the small little world of sRGB :) I was wondering if it's possible that a construction of an empirically consistent color appearance could benefit from considering the explicit representation of lighting sources, and even surface properties of the object being viewed? Does that sound like overkill?