What dE to use when calibrating?

greetings,

Historically, grayscale work used dE 76 and color stuff used dE 94.

It is still taught that way. although like gamma used to be, it also depended a lot on which expert you were talking to.

Regards

Sent from my Nexus 7 using Tapatalk 2

you may also want to use dEuv for grayscale, I know many professionals do,

and de94 for gamut as Michael posted...

CalMANv5 will use dE2000 by default for everything because it's more uniform (for color errors) than the older formulas.

uv (and uvL) is still probably the most perceptually "reasonable" color space, so using dEuv makes a lot of sense. I use dEuv for both grayscale and color. But I also use dE94 (as an additional line of data in the "raw data" table in CalMAN) as a sanity check to avoid large errors that are missed by the dEuv calculations.

Every time a "newer" method of calculating perceptual errors comes out, it is touted as being "better" but every time, it doesn't take too long for someone to find flaws in the newer method. To this day, there's not a single color space option that is perfect/accurate/reliable for every possible measurement.

uv/uvL space relies on hue, saturation, and Luminance to define colors (and to calculate errors). It's difficult, but not impossible, to encounter perceptual errors in uvL space BUT... uvL has the advantage of behaving most like human visual perception in most circumstances. You can think of uvL space as a cylinder with dark (black) on the end of one axis, white at the other end of the center axis, saturation running radially from the center axis to the outside, and hue running around radially. It avoids all sorts of non-linear functions that have to be applied to other spaces to get the "other" dEs (1994, 2000, etc.).

So if you want to be thorough, use dEuv PLUS an additional calculation of your choice... I use 1994, you may want to use something else.

CalMANv5 will also support dELuv in addition to the current dEuv.

CIELUV is now 35 years old. There has been an enormous amount of research by color scientists since 1976. Since 76 the dE formulas have become increasingly complicated in an effort to address shortcomings discovered with the earlier formulas. All of the dE methods approved by CIE since 1976 have been based on the Lab color space.

The Luv color space offers a few advantages for grading video performance. First, it offers a linear chromaticity chart, which Lab does not. This factor was decisive in CIE approving Luv as a Lab alternative in 1976. Second, it is somewhat more sensitive than Lab to small errors in gray, so it presumably offers a little more granularity. Third, it weights red errors somewhat more than Lab, which can be important for getting skin tones correct.

However, the 1976 (Lab or Luv) formulas have some rather profound shortcomings regarding their treatment of luminance errors in color. Thus, I would only use CIELUV for grayscale analysis. For color, use CIE94 or CIEDE2000.

One thing to note is that there are two types of uv.

CIE u'v' which is a more uniform 2D colorspace similar to xy.
and
CIE L*U*V* which is very similar to CIE L*a*b* .

The dE uv in CalMAN uses the CIE u'v' space so it is only chromaticity that is reported as error with no regard to luminance or perception.

but v5 should have that latter as well? (L*U*V*)

Yes in v5 we try to be as clear as possible about the colorspace used in the dE formula

CIE u'v' is just a useful convention we can use to describe the CIELUV color space minus the luminance component. I have never seen this referred to as a native color space in any color science text I have looked at, or at least I don't recall seeing it.

You would never use u'v' for color. It's value lies in grayscale analysis only, where the luminance performance is better treated as gamma response. For the reasons I mention above, I would also not recommend using CIELUV (or CIELAB) for color either, though this is merely a suboptimal idea as opposed to using u'v', which would be a terrible idea.

BTW, the subject of the various color difference formula and their relative performance is a subject that I find endlessly interesting, and strongly influenced by decision to get into this field. CIE endorsed a new color space CIECAM02, but has not yet settled on a color difference formula that uses it. You can find several examples on the web showing how this might be done.
Edited by TomHuffman - 8/19/12 at 5:43pm

I believe this is the root of most of the misunderstanding.

CIE u'v' is a coordinate system where every chromaticity has a fixed value, d65 = 0.1978, 0.4683 u',v'

CIE L*U*V* is based on a white point, so the reference point for white is 0,0.


But if the white point is D50, the u'v' coordinates shift to 0.2092,0.4881, while L*U*V* is still 0,0 since it's coordinate system is in reference to white.

One is a competitor to L*a*b*, where you actually get perceptual saturation effects, and u'v' is the replacement for x,y where you are simply describing chromaticity without respect to perceptual saturation.


The saturation of colors based on luminance is very significant as well.

For Red the u'v' targets are always 0.4507, 0.5229
In L*U*V* at 100% they are 53.2, 175.01, 37.75
at 30% they are 12.7, 41.85, 9.02

So as drop in target luminance the ability to perceive saturation drops as well
Edited by sotti - 8/19/12 at 6:18pm

The only dE formulas that CIE has ever endorsed use the Luv or Lab color spaces. They do NOT use u'v', which are chromaticity coordinates useful for plotting color on a CIE chart, but useless for dE calculation. I assumed that you were using u'v' as shorthand for Luv without the L, but this reads like you are referring to the 1976 chromaticity coordinates. To give a very specific grayscale example, if the test color is x0.319, y0.329, and the reference is x0.3127, y0.329 then the CIELUV dE would be 5.8.

=SQRT((100-100)^2+(0.0 - 5.705)^2+(0.0 - 1.216)^2)

Are you suggesting using the u'v' coordinates u'0.1948, v'0.4683 and u'0.2022, v'0.4963 instead?
Edited by TomHuffman - 8/20/12 at 6:38pm

I've always found it odd that calibrators haven't embraced the progress made over the last 35 years. Why wouldn't one simply use CIEDE2000? To not do so suggests that CIE development has headed off in the wrong direction.

And that's why it's the default dE formula in CalMAN5.

Of course we make it easy to switch formulas if you have a personal preference.

how different are dE1994 and dE2000 for gamut? for grayscale?

There are small differences. For example, the test color I reference above x0.319, y0.329 has a CIE94 of 3.5 and a CIEDE2000 of 4.6. an oversaturated green of x0.296, y0.675 is CIE94 7.2 and CIEDE2000 6.0. If you add a luminance error as well, say 0.7842, then CIE94 8.5, CIEDE2000 6.8.

I mostly agree, especially with respect to the odd attraction to CIELUV. I am reluctant to wholeheartedly endorse CIEDE2000 I guess for philosophical reasons. I am skeptical that the very small difference in uniformity is worth the truly immense increase in complexity. The complete CIEDE2000 formula is almost comically complex.

does the complexity issue affect the end user at all or just the software developers?

Just the developers.

And I wouldn't say that programming dE 2000 is that complex. If I can manage it in an Excel spreadsheet, a programmer will find it a breeze. CIECAM02 looks like a different kettle of fish...although perhaps the dE formulation ends up being easier, I have no idea.

I guess my main question would be less in relation to the complexity of the calculation, but rather given a particular dE2000 result whether the hunt for a better result is made more complex (if that makes sense)...

Just so others can see what what we are talking about

dE 1976



dE 1994



where,



dE2000



where,



Here's the added performance of each successive formula



As I said, my only point was philosophical. With CIE94, you get a large increase in performance for a modest increase in complexity. With CIEDE2000 you get a small increase in performance with a huge increase in complexity. CIECAM02 seems to offer little over CIEDE2000, which is why I suspect that it has not been formally adopted. For the user, this added complexity means nothing. Even for the developer, once it is in place all the work is done.

I just have an affinity for Ockham's Razor.

Years ago when no one seemed to want to implement dE2000 I made this spreadsheet to help get myself through this formula:

https://dl.dropbox.com/u/70685392/deltaE%282000%29%202.xls

Are the dE targets still as follows, or are there other recommendations?

CIELUV : white (4.0), color (4.0)
CIELAB : white (3.0), color (3.0)
CIE94 : white (3.0), color (1.5)
CIE2000 : white (3.0), color (1.5)

Well I'm sure they at chromapure believe the same thing, as all de methods are selectable, and the manual clearly explaines the benefits and shortcommings of the different formula's.

I think by now we all get that these teams are very competitive and that each program is excellent in it's own right but...

it would honestly be great if threads like this were not turned into a 'measuring of swords'.

Those of us that know enough to be capable also know that while the fundamentals are, more or less, in place there are so many varied 'professional' opinions that the term science often fades from view.

Jason

What everybody is forgetting is that we are calculating errors for 2, 10 or 20 well defined grayscale steps and (usually) 75% colors for the 3 primaries and 3 complimentary colors (they are NOT secondary colors, secondary colors come into play on the printed page or crayons).

uv (and uvL) are REALLY all you need if you are using calibration measurements and you are using the calibration software to display an appropriate luminance graph so you can see if there's a possibly understated Luminance error... it will be right there in the software's graph. Most of the differences in 1931 vs 1972 vs 1994 vs 2000 fix oddball "mis-calculations" in unusual colors and aren't necessarily focused on making red, green, blue, cyan, magenta, and yellow calculations particularly better... though they do tend to improve.

Someone mentioned calibrators not taking advantage of "newer developments" or "newer science"... if we were working with some esoteric colors that had historic problems being measured accurately... certainly, some "newer science" might be worthwhile. But we're measuring white (and gray) and 6 other highly defined colors that are the 6 colors ANY error calculation method has to get reasonably right in order to not be laughed out of existence.

In short, if you know what you are doing and you use the calibration software correctly (so you can see how close your measured points are to the reference coordinated and so you can see how accurate (or inaccurate) your luminance values are, you really don't even NEED and error calculation unless you simply need to know if +1 green produces a higher or lower error than +2 green (for example). You could achieve the same end calibration result just by zooming-in on the reference point so you see your measurement and the reference with more resolution. In fact, if the cal software supports that capability, I'd venture to say a good calibrator could produce essentially "perfect" results for a given display without a single dE calculation as a guide (by "perfect" I mean get the display as accurate as it can be made given the controls, brand, and model).

With out a doubt that is true. dE's are good to let you know when it doesn't matter any more, but finding optimal settings doesn't require dE at all.

Not necessarily. This issue goes way beyond "esoteric colors". What if you have a CMS with an inadequate range of adjustment, so you are unable to remove all of the error from an extremely oversaturated color? Or perhaps you are working with a gamut with one or more undersaturated colors that cannot be corrected? Or, better yet, what if you have no CMS at all? All you have a main color control. In these cases, simply showing the luminance error is not sufficient because there may very well be a specific mix of luminance and chromaticity values that minimizes overall visible error that is different from simply getting the correct level of luminance. Furthermore, the 1976 dE formulas make very different recommendations in this regard compared with the 94 and 2000 formulas. They can't all be correct.

Also, the Lab-based formulas and Luv weight the contributions of RGB error differently, so if there is a residual error (you can't get it perfect) it is useful to have a dE number to tell you whether that error meets reasonable tolerances. Obviously, because the formulas weight the RGB contributions differently, you may get different results depending on the dE formula you use.

Of course, if you get a really low dE value when you are finished, then it doesn't matter which formula you use. A dE of 1 is going to be a very accurate color regardless of the dE method used to calculate it. It is when the result is less than perfect and compromises have to be made where these sorts of decisions make a real difference.