Calculating dE

[stevekale]

Originally Posted by zoyd

If you want to calculate these yourself, Tom's thread provides a spreadsheet to do it in either CIELUV or CIELAB color systems and there is another spreadsheet in the reading charts for dummies sticky posted by Krasmuzik with the CIELUV calculations.

Hi Zoyd

It's interesting to look at these two spreadsheets and their results, particularly the calculation of dE. Being a Photoshop guy I am much more used to working in LAB. Despite Bear5K's comment here:

http://www.avsforum.com/avs-vb/showt...6#post11770736

it's interesting to read someone like Bruce Fraser say "LAB has largely replaced LUV in most practical applications, and while it isn't perfect (it exaggerates differences in yellow and underestimates them in blues, for example) it's pretty darn useful. The quest for a perfectly uniform color space continues, but thus far LAB has stood the test of time." (Real World Color Management, Second Edition.) Frankly, I don't understand Bear5K's comment because whether a colour observation is reflective or emissive shouldn't affect perceived differences. But the choice of colour space model will affect the calculations of dE and, as a result, one's own assessment of the magnitude of error.

Using Krasmuzik's Luv model with my data I get:

dE(uv) dL dC dH
Red 6.9 2.2 6.5 0.5
Green 4.5 2.2 3.9 0.0
Blue 2.7 0.6 2.5 -0.6
Yellow 2.0 1.2 1.5 -0.7
Cyan 7.2 1.6 7.0 -0.9
Mag 4.8 -0.1 4.8 0.7
White 2.2 0.0 2.2 #DIV/0!

So reasonably high overall error, most of which seems to sit in saturation rather than hue.

Using Tom's CIELAB model yields very different results:

Lightness Saturation Hue Total
dE dE dE dE
Red 16.5% 58.7% 24.9% 3.0
Green 11.7% 45.5% 42.9% 3.5
Blue 24.9% 49.2% 25.9% 0.4
Cyan 19.8% 70.4% 9.8% 1.9
Magenta 1.8% 89.6% 8.6% 0.7
Yellow 40.7% 48.6% 10.7% 0.3

(I note that neither set of dE figures are exactly what you gave me. Any idea why?)

Unfortunately Tom has locked and password protected his spreadsheet so it is more difficult to understand how to interpret his L*, saturation and hue error figures. (Any ideas?) But the (total) dE figures are, I would say, more than satisfactory.

(Oddly, one should at least be able to easily relate the two models' dL figures but I'm struggling. Take Red for example. My reading of 28.046 vs a white of 120.271 amounts to an L* of 55.4 versus a target of 53.2 and hence a dL of 2.2. I have no idea how to interpret Tom's lightness error of 16.5% in the CIELAB space.)

I'm intrigued therefore, as to the approach taken to dE in the next version of ColorHCFR. For starters, LUV or LAB?

Lastly, can you explain Krasmusik's dC to me. (Maybe I should post this question in the thread he started...) I had thought that this was "saturation (tint) error" (and that dH was "hue error"). But it seems to me when playing with the values (in particular copying target values into the measured values and noting the impact to the delta readings) that dH is both saturation and hue, combined, and that dC is something else much more dependent on L* as well.

I suspect I'm not understanding saturation properly. I always thought of it the way it is described here:

http://en.wikipedia.org/wiki/Image:HSV_cylinder.png

That is, it's a property independent of lightness. But if I copy the target x,y values to the measured part of Krasmusik's spreadsheet, while dH falls to zero, dC does not. Hence dC is dependent on Y or lightness in the spreadsheet. Confused...

Regards

Steve

[TomHuffman]

It's not surprising that you got different results. Not only is one based on CIELUV and the other in CIELAB, but the CIELUV is based on the 1976 standard and the CIELAB result is based on the 1994 formulation, where the numbers are generally smaller.

Regarding the use of CIELUV for colored lights and CIELAB being used for reflective light, this appears to be based on 2 factors, one practical and one technical:

First, CIELUV performance can be expressed cleanly on a CIE chart in which the secondaries and primaries fall on a straight line. CIELAB does not graphically plot as well.

Second, for technical reasons that I have not been able to track down, and believe me I've looked, CIELUV is better adapted to additive color, (RGB, which video uses) and CIELAB is better suited for subtractive color (CMYK, which the print industry uses).

Finally, it must also be noted that BOTH standards are explicitly device-independent: That is, they are color models that apply to the way the eye perceives color and not dependent upon the source that produces the color.

Until I read from an authoritative source that claims that using CIELAB for video is just plain wrong-headed, I will continue to use it for one reason. CIELAB is by far the more robust color model. This is evidenced by the fact that CIE updated its formula for calculating dE in 1994 to further improve upon perceptual uniformity, but it ONLY applies to LAB. Although I've been unable to find a satisfactory technical explanation for the adaptation of CIELUV to RGB and CIELAB to CMYK, EVERY source I've looked at is absolutely clear in asserting the superiority of the 1994 calculation over the 1976 calculation.

[stevekale]

Originally Posted by TomHuffman

It's not surprising that you got different results. Not only is one based on CIELUV and the other in CIELAB, but the CIELUV is based on the 1976 standard and the CIELAB result is based on the 1994 formulation, where the numbers are generally smaller.

Regarding the use of CIELUV for colored lights and CIELAB being used for reflective light, this appears to be based on 2 factors, one practical and one technical:

First, CIELUV performance can be expressed cleanly on a CIE chart in which the secondaries and primaries fall on a straight line. CIELAB does not graphically plot as well.

Second, for technical reasons that I have not been able to track down, and believe me I've looked, CIELUV is better adapted to additive color, (RGB, which video uses) and CIELAB is better suited for subtractive color (CMYK, which the print industry uses).

Finally, it must also be noted that BOTH standards are explicitly device-independent: That is, they are color models that apply to the way the eye perceives color and not dependent upon the source that produces the color.

Until I read from an authoritative source that claims that using CIELAB for video is just plain wrong-headed, I will continue to use it for one reason. CIELAB is by far the more robust color model. This is evidenced by the fact that CIE updated its formula for calculating dE in 1994 to further improve upon perceptual uniformity, but it ONLY applies to LAB. Although I've been unable to find a satisfactory technical explanation for the adaptation of CIELUV to RGB and CIELAB to CMYK, EVERY source I've looked at is absolutely clear in asserting the superiority of the 1994 calculation over the 1976 calculation.

Re the different results, I meant that neither matched the numbers Zoyd had quoted me. (Not just that the two methodologies generated different results which of course they did.)

I'm with you on LAB. Do we see companies like Gretag/X-Rite invoking LUV when they profile displays and LAB when they profile printers? No. It's LAB all the way.

But I'd go one step further than your proposal and update the dE calculation to the 2000 definition which is meant to be a marked improvement on the 1994 calculation. So I'm hoping that LAB and dE(2000) will be embodied in the next version of ColorHCFR.

Incidentally, the dE for my calibration using dE(2000) are:

R 2.2; G 1.5; B 0.7; C 1.9; M 0.7; Y 0.8

(using Bruce Lindbloom's calculator)

Regards

Steve

[I've had a go at putting the dE(2000) calculation in a spreadsheet. I still need to do some more work on it because despite matching Bruce's calculator for white to about 8 dec places I don't generate the exact same results for the other colours and despite having gone over the calculations a dozen times I can't find the error (and I'm capable enough to see I've filled the formula down from red through to white properly and it's the last one that's right!).]

[TomHuffman]

I've not started using the 2000 standard because my understanding is that it has not been widely adopted in the way the 94 standard was when it was announced and because it is apparently quite complicated to calculate (these two factors are probably related).

Originally Posted by stevekale

Re the different results, I meant that neither matched the numbers Zoyd had quoted me. (Not just that the two methodologies generated different results which of course they did.)

I'm with you on LAB. Do we see companies like Gretag/X-Rite invoking LUV when they profile displays and LAB when they profile printers? No. It's LAB all the way.

But I'd go one step further than your proposal and update the dE calculation to the 2000 definition which is meant to be a marked improvement on the 1994 calculation. So I'm hoping that LAB and dE(2000) will be embodied in the next version of ColorHCFR.

Incidentally, the dE for my calibration using dE(2000) are:

R 2.2; G 1.5; B 0.7; C 1.9; M 0.7; Y 0.8

(using Bruce Lindbloom's calculator)

Regards

Steve

[I've had a go at putting the dE(2000) calculation in a spreadsheet. I still need to do some more work on it because despite matching Bruce's calculator for white to about 8 dec places I don't generate the exact same results for the other colours and despite having gone over the calculations a dozen times I can't find the error (and I'm capable enough to see I've filled the formula down from red through to white properly and it's the last one that's right!).]

[stevekale]

Originally Posted by TomHuffman

I've not started using the 2000 standard because my understanding is that it has not been widely adopted in the way the 94 standard was when it was announced and because it is apparently quite complicated to calculate (these two factors are probably related).

It is a rather complicated formula, especially for trig novices such as myself, but if you can do one you can likely do the other. It probably just takes time for software to be updated. I've been over the formulae in my spreadsheet a couple more times and can't see any issues unless of course I simply have misunderstood Bruce's annotation. If you want I can post the spreadsheet for you and others to look at and/or correct.

[TomHuffman]

Yes, please go ahead and post this. BTW, do you have numbers for the minimal perceptible difference for the 2000 standard?

Originally Posted by stevekale

It is a rather complicated formula, especially for trig novices such as myself, but if you can do one you can likely do the other. It probably just takes time for software to be updated. I've been over the formulae in my spreadsheet a couple more times and can't see any issues unless of course I simply have misunderstood Bruce's annotation. If you want I can post the spreadsheet for you and others to look at and/or correct.

[stevekale]

Originally Posted by TomHuffman

Yes, please go ahead and post this. BTW, do you have numbers for the minimal perceptible difference for the 2000 standard?

Ok here you go. I've got the dE figures correct now. (I had misinterpreted Bruce's annotation for the inverse tangents - what on earth are these? don't answer - and had the a,b coordinates around the wrong way.) My conversions from xyY to LAB seem to be slightly off but I think this is simply inaccuracy in the definition of the reference white to D65.

As to the minimum perceptible difference, here's an extract from Real World Color Management, Second Edition (Bruce Fraser):

"In theory, LAB is perceptually uniform, and in theory one delta-e unit is the smallest difference perceivable by humans with normal color vision, but in practice it falls a little short of that goal. In the saturated yellow region, for example, even viewers with very discriminating color vision may be hard-pressed to see a 3 delta-e difference, while in midtone neutrals a difference of 0.5 delta-E may be possible." No reference is made to which delta-e calculation they're referring to.

Now of course the work continues to take account of these points and that's why we have the newer 1994 and 2000 delta-e calculations. (From what I've read thus far, the 1994 work was much more of a quantum leap than the 2000 work.) Nonetheless, I think one can take away the view that a perfect delta-e calculation, if it existed, would mean that 1 delta-e point difference is the smallest perceivable difference. Work on the delta-e formulae refine to this goal. (While I assume work on the development of alternative, more accurate perceptually uniform colour spaces also continues.) ("Understanding Color Management" by Sharma has a brief discussion of why the 1994 and 2000 calculations improved on the earlier ones with a discussion of ellipticals versus spheres that goes way over my head!)

Regards

Steve

[TomHuffman]

Here's an unlocked version of my CIELAB 94 spreadsheet. Feel free to check the math.

[stevekale]

I just noticed an error in the calculation for dE versus Rec709 which I've now corrected. I've amended my earlier post to include the new file. (I also started playing with greyscale calcs but have run out of steam for now.) Let me know if there are any other issues.

Regards

Steve

[Bob Sorel]

I have moved a few off topic posts from another thread to a thread of its own here.

[stevekale]

I did a bit more digging on this subject and asked some questions of the colour management/science gurus on the Apple Colorsync list. They were very helpful.

In summary, it's not the uniformity of the colour space model (LAB versus LUV etc) that we need be so concerned about in discussing dE estimations but rather the accuracy of a particular dE formula to measure perceived colour difference. But for the record, the newer CIECAM02 colour space is regarded as a much more perceptually uniform than CIELAB.

There is no doubt in my mind, based on the responses from the Colorsync list and all that I've read, that CIEDE2000 is a much more robust measurement of perceivable colour difference than dE(1994) or LUV-based calculations.

There is a useful overview of the historical development of colour difference equations here:

http://www.iscc.org/jubilee2006/abst...uoAbstract.pdf

The commentator that referred me to the above went on to say "in this paper, it's clear that CIELUV stopped and new formulas based on CIELAB were
developed, like CIEDE2000 and others. I believe CIELUV is rarely used
nowadays."

Another explained in more detail:

"CIELuv was brought up during the last CIE meeting in Ottawa, last year, and
while the consensus among the attendees was that CIELuv wasn't really needed anymore (all current color science development revolved around CIELab and CIECAM worldwide) it was agreed to keep it around for 'emissive' people. Seems to serve those in the broadcast or television industries well.

But the prevalent model is CIELAB, no question."

"CIE Luv is a linear transformation of XYZ. As such, it does not have the
cube-root compression that CIELab has. Plus, it does not have at its root
the Munsell color system. So, perceptually, it is not better to describe
color differences than, say, CIE xy. dE94 and dE2000 are both 'weighted'
color differences metrics."

I very much hope that the ColorHCFR crew decide to include a CIEDE2000 calculation in their software.

Regards

Steve

[zoyd]

Originally Posted by stevekale

Using Krasmuzik's Luv model with my data I get:

dE(uv) dL dC dH
Red 6.9 2.2 6.5 0.5
Green 4.5 2.2 3.9 0.0
Blue 2.7 0.6 2.5 -0.6
Yellow 2.0 1.2 1.5 -0.7
Cyan 7.2 1.6 7.0 -0.9
Mag 4.8 -0.1 4.8 0.7
White 2.2 0.0 2.2 #DIV/0!

Using your Steve608v4.chc measurements, HCFR (which uses the LUV formulation) yields for rec.601:

R 7.5 (7.7)
G 5.4 (5.3)
B 3.7 (3.9)
Y 2.8 (2.9)
C 5.6 (5.3)
M 4.7 (4.9)
W 2.4

The numbers in parenthesis are what you get if you enter the values manually to 4 decimal places and the slight differences are the display/export problem KI_ talked about in the HCFR thread.

1st question, why do these differ from values I posted in the other thread?
answer: These are different from the ones I posted in the HCFR thread for this data file because I had replaced your white measurement in the primaries/secondaries window with the 75% value from the grayscale window (not sure why I did that).

Using Krasmuzik's spreadsheet with SMPTE-C target values the numbers are:
R 7.7
G 5.5
B 3.9
Y 2.8
C 5.5
M 5.5
W 2.4

I don't know why you calculated different numbers using that spreadsheet but these agree pretty closely with what HCFR is generating although magenta is off by 0.6 and I'm not sure why. Are you sure you used the SMPTE-C target chromaticities and not rec.709?

Lastly, can you explain Krasmusik's dC to me. (Maybe I should post this question in the thread he started...) I had thought that this was "saturation (tint) error" (and that dH was "hue error"). But it seems to me when playing with the values (in particular copying target values into the measured values and noting the impact to the delta readings) that dH is both saturation and hue, combined, and that dC is something else much more dependent on L* as well.

dC contains the magnitude of the chromaticity errors (du and dv) but it is weighted by the lightness L*. It is proportional to the magnitude of the distance between white and the (u',v') coordinate and also L*. What it tells you is that an error in saturation can be enhanced (or subdued) depending on the lightness. This is why "saturation" is a difficult concept to visualize.

[TomHuffman]

Steve:

Thanks. Very helpful post.

BTW, a quote in your post

"CIELuv was brought up during the last CIE meeting in Ottawa, last year, and while the consensus among the attendees was that CIELuv wasn't really needed anymore (all current color science development revolved around CIELab and CIECAM worldwide) it was agreed to keep it around for 'emissive' people. Seems to serve those in the broadcast or television industries well.

But the prevalent model is CIELAB, no question."

and a comment in the referenced paper

"The formulae in the family of linear transformation from XYZ have been widely used for additive colour mixing such as that involving coloured lights and emissive phosphor displays. Some earlier formulae were developed including the CIE U*V*W* apace.7 In 1976, it was refined to become CIELUV.
CIELAB and CIELUV have been widely used, mainly because it is relatively easy to relate colours as seen with positions on the diagram. The ΔE values are calculations of the distance between the standard and sample in these spaces. They are used for industries concerned with subtractive mixture (surface colorant) and additive mixture of coloured light (TV), respectively."

are typical of the sort of offhand comments I ran into again and again.

Just EXACTLY what is it about CIELUV that makes it better adapted to additive (RGB) color and CIELAB better adapted to subtractive (CMYK) color? And, is this difference important enough to override the obvious advantages carried by CIELAB in general and the more recent color difference formulas associated with it in particular?

[zoyd]

Originally Posted by TomHuffman

And, is this difference important enough to override the obvious advantages carried by CIELAB in general and the more recent color difference formulas associated with it in particular?

There is another aspect of this you have to keep in mind. While improving the measurement of color differences (especially small ones which the post 1994 formulations appear to do) is theoretically a good thing to do, it may not have much practical significance for display calibration. There are four main reasons for this, environmental "pollution" and display artifacts, variances in individual perception, probe error, and lack of good color management controls. So what are the errors introduced by environmental factors such as room lighting and wall color? Display artifacts such as the color shifting seen in LCD off-angle viewing? Also, all of the color system dE values can be traced back to color-matching experiments which are averages over a group of observers, the peak-to-peak variance in these experiments can be as high as 10-15 dE with standard deviations of ~5 dE. Probe errors for the good ones are in the 2-3 dE range. Lack of color management controls leave a majority of displays at the 10 dE level for one or more primary colors. When you add up all these external error sources at the end of the day the color perceived by any individual brain is likely to be far outside the precision of any of the dE systems currently used. It's one of those "dirty little secrets" but the truth is that the tools (and formulations) we use to do display calibration are limited not by their own internal imperfections but by external factors like the HVS, environment, and lousy display controls.

[TomHuffman]

Zoyd:

Another excellent post. I agree entirely.

I guess at this point I am more interested in a correct theoretical model than I am in whatever practical benefit that theoretical model would deliver in practice.

Originally Posted by zoyd

There is another aspect of this you have to keep in mind. While improving the measurement of color differences (especially small ones which the post 1994 formulations appear to do) is theoretically a good thing to do, it may not have much practical significance for display calibration. There are four main reasons for this, environmental "pollution" and display artifacts, variances in individual perception, probe error, and lack of good color management controls. So what are the errors introduced by environmental factors such as room lighting and wall color? Display artifacts such as the color shifting seen in LCD off-angle viewing? Also, all of the color system dE values can be traced back to color-matching experiments which are averages over a group of observers, the peak-to-peak variance in these experiments can be as high as 10-15 dE with standard deviations of ~5 dE. Probe errors for the good ones are in the 2-3 dE range. Lack of color management controls leave a majority of displays at the 10 dE level for one or more primary colors. When you add up all these external error sources at the end of the day the color perceived by any individual brain is likely to be far outside the precision of any of the dE systems currently used. It's one of those "dirty little secrets" but the truth is that the tools (and formulations) we use to do display calibration are limited not by their own internal imperfections but by external factors like the HVS, environment, and lousy display controls.

[krasmuzik]

A very important point zoyd makes that it is the tolerance for errors in video why CIELUV has not been updated.

When CIEDE2000 was approved for trial - CIE said only for dE<5. Even the above paper says they improved small to medium errors - what it does not say is at the expense of large error accuracy - the type you see in video calibration.

chromaticity diagrams are important to video which is why CIELUV has been retained in the standard. They are a direct representation of how light adds to make new colors. Perception of light is more strongly weighted by its luminance - this is why errors are directly proportional to it in CIELUV. Surfaces do not even have a luminance - they proportionally return the ambient light it receives. This is why modeling for lightness is entirely different between the two standards.

CIELAB is more suited to surface color - in which color difference, appearance and specification is extremely important. Product branding and trademarks requires reproducing the exact colors under stringent QA - even if imperceivable. While modern error measures have focused on improving accuracy for CIELAB - that is hardly a concern for video for reasons zoyd pointed out - nobody cares about 5dE error in video - but that is considered a huge error for surfaces. Consider the improvements were done using psychovisual experiments on surface colors. Do you pick the coefficients for textiles or paints for video? What relevance does ambient light on the surface have to video? With FP none - ambient washes out the video so all you see is the underlying surface. For RP and directview - also none - all you see is the light from the display. For surfaces ambient light is absolutely important to perception - to the point that dE is replaced entirely with CAM (chromatic adaptation or color appearance modeling)

You can conduct your own psychovisual experiments with video - run all the various versions of DE - especially the LCH versions. Simply ask yourself does this error magnitude rank match what I see?. I have done that with many people using CIELCH based on CIELUV - they all agreed it does work. None of them cared if dE should be 0.895 or 0.864 - they cared about was pre/post calibration improved by 15-20dE - something none of these CIELAB dE formulas really care about. Customers care about does that display have a 50dE or 5dE for green.

The real irony is that HT calibrators seem to be a lot more anal about dE than do their professional counterparts - who likely know nothing even about dE - and no surprise they are not active in the CIE process to update CIELUV. When I have to have different calibration presets for each HDTV broadcast channel - it is obvious this is the least of a professional broadcast engineers concerns.

[krasmuzik]

Steve

read the thread from which you obtained the spreadsheet, as well as the references listed. I don't need to be explaining CIELCH derived from CIELUV here when it is done better elsewhere.

[TomHuffman]

Originally Posted by krasmuzik

nobody cares about 5dE error in video

Although I am actually sympathetic to this spirit of this statement, but I can't resist the temptation to point out that you and Zoyd made exactly the opposite point when I claimed that the difference between Rec. 601 and Rec. 709 was really not big enough to be terribly concerned about.

http://www.avsforum.com/avs-vb/showp...&postcount=330

and subsequent posts.

[zoyd]

Originally Posted by TomHuffman

Although I am actually sympathetic to this spirit of this statement, but I can't resist the temptation to point out that you and Zoyd made exactly the opposite point when I claimed that the difference between Rec. 601 and Rec. 709 was really not big enough to be terribly concerned about.

http://www.avsforum.com/avs-vb/showp...&postcount=330

and subsequent posts.

Not the same point at all, you should try and minimize the error sources you have control over, especially if they are anywhere near the same magnitude as what the display will allow (e.g. 601/709 red difference on CMS enabled sets). The question of whether you use dE(LUV) or dE(something else) to do it is separate, Kras's argument is that dE(LUV) is the best system for measuring typical video dE's in the emissive domain, you are searching for a definitive statement of "How much better?" and why. I don't know the answer to that one.

[krasmuzik]

A >10dE difference between SD/HD is actually well outside the recommended use of latest dE updates focused on small dE improvements. Another great example of why the original definitions should be used.

I do care about SD/HD Red's - as they are well over 5dE(uv)- most people I know can see the diff. But as I pointed out to you in that thread - you should care more about SD/HD blues because it has a larger error in dE(ab) than red. I personally find it difficult to see the blue difference other than side by side comparisons of test patterns - less than 5dE(uv) explains why. Maybe you are red blind like 10% of males and cannot see SD/HD red is surely the worst difference than SD/HD blue? Though I realize you are pointing to average dE as why they should be insignificant - I look more at individual differences to show SD/HD can be perceived - not the average.

Those worrying the most about the perfect dE number - don't seem to be comparing it to what their eyes see with gross errors and fail to realize that dE(ab) improvements don't care about those gross errors. Compare the numbers against what you see.

[TomHuffman]

Originally Posted by zoyd

Kras's argument is that dE(LUV) is the best system for measuring typical video dE's in the emissive domain, you are searching for a definitive statement of "How much better?" and why. I don't know the answer to that one.

I don't either. However, if I understood Kras correctly, at least one of the REASONS he gave for why dE(LUV) was better was that any advantages offered by the alternatives were important only for very small dE differences and not as useful for the larger (i.e., over 5 dE) differences we typically care about with video.

I want to stress that this is a largely academic debate with not a lot of practical consequence, which is fine.

To get back to the SD/HD difference, there is one practical consequence I know of: I'm as much of a color Nazi as anyone I know and I'm too lazy to setup separate SD/HD profiles on my CMS, simply because SD material (broadcast TV mostly) doesn't look obviously wrong to me in the HD color space, though HD material looks a little desaturated in the SD space. I just run it at Rec. 709 and leave it there.

[TomHuffman]

Originally Posted by krasmuzik

But as I pointed out to you in that thread - you should care more about SD/HD blues because it has a larger error in dE(ab) than red. I personally find it difficult to see the blue difference other than side by side comparisons of test patterns - less than 5dE(uv) explains why. Maybe you are red blind like 10% of males and cannot see SD/HD red is surely the worst difference than SD/HD blue? Though I realize you are pointing to average dE as why they should be insignificant - I look more at individual differences to show SD/HD can be perceived - not the average.

The eye is much more sensitive to red, especially under bright conditions than blue and we are especially sensitive to red anyway because the impact it has on skin tones.

And BTW, one of the consequences of the dE94 formulation was to substantially lessen the blue weighting of LAB.

[zoyd]

Originally Posted by TomHuffman

The eye is much more sensitive to red, especially under bright conditions than blue and we are especially sensitive to red anyway because the impact it has on skin tones.

All the dE formulations fold in the HSV responsivity because they're all designed to fit experimental color-matching data generated by people choosing between reference colors and target colors. The fact that when I look at primaries calibrated using dE(LUV) for HD and SD and see that first red, then green, then blue are most noticeably different and that the dE numbers correspond to my perception tells me that at least in a relative sense dE(LUV) is doing at better job than dE(LAB) in this particular case.

[TomHuffman]

Let me add some additional data to this. I have a particular interest in the JVC RS1 projector. I have been critical of it since release because of its extreme color inaccuracies, and I have been excoriated on the Projectors forum for doing so. I've calibrated three of them now, and in two of the cases I had an external CMS to help reign in the primaries. Anyway, I'm pretty familiar with it. Using the numbers I got from my most recent experience, here's the xyY data and the dE weightings I got using various calculations. Feel free to check my math.

R: 0.667---0.327---0.233
G: 0.311---0.673---0.766
B: 0.148---0.055---0.085

CIELAB76
R: 26.1% 24.2
G: 54.8% 50.8
B: 19.1% 17.7

CIELUV
R: 26.6% 9.8
G: 45.9% 17.0
B: 27.5% 10.1

CIELAB94
R: 28.4% 5.9
G: 53.4% 11.1
B: 18.1% 3.8

CIELAB2000
R: 34.0% 6.2
G: 48.3% 8.8
B: 17.8% 3.2

LCH(uv)
R: 39.6% 26.4
G: 36.1% 24.1
B: 24.3% 16.2

The fact that there's an external CMS available is helpful because it provides some objective measurement that backs up the subjective impressions, which are that green is worst, followed by red, and then blue. Lumagen is planning to increase, yet again, its CMS's range of adjustment only because what's provided simply isn't enough to fully reign in the RS1 green primary.

First, notice the very large differences in the absolute numbers. This shows the need for clear standards for minimal perceptible error and maximum acceptable error relative to the calculation you use.

Second, this display provides a large-scale dE error by any definition. But notice that in this instance, it's CIELUV that gives a too-large blue error. And the LCH(uv) calculation seems just plain wrong. Red simply is not the biggest problem with the RS1. The CMS adjustments show this and just about everyone who has ever noticed its color errors focuses on the nearly glowing greens.

Third, compared to the differences between SD and HD dE differences (which are relatively small), this suggests that neither CIELUV nor (especially) LCH(uv) does better with large scale errors.

Originally Posted by zoyd

All the dE formulations fold in the HSV responsivity because they're all designed to fit experimental color-matching data generated by people choosing between reference colors and target colors. The fact that when I look at primaries calibrated using dE(LUV) for HD and SD and see that first red, then green, then blue are most noticeably different and that the dE numbers correspond to my perception tells me that at least in a relative sense dE(LUV) is doing at better job than dE(LAB) in this particular case.

[dlarsen]

Originally Posted by TomHuffman

I'm as much of a color Nazi as anyone I know and I'm too lazy to setup separate SD/HD profiles on my CMS, simply because SD material (broadcast TV mostly) doesn't look obviously wrong to me in the HD color space, though HD material looks a little desaturated in the SD space. I just run it at Rec. 709 and leave it there.

+1. I believe I can be a little discriminating too and I can’t perceive much downside to rendering everything to sRGB/Rec.709 gamut and levels.

Dave

[krasmuzik]

Lipstick, sunburn should be accentuated with REC601 material rendered as REC709 - just not as much as one would see with red pushed video or an excessive gamut - which appears to add rather than accentuate lipstick/sunburn. I do think one needs trained to see this SD/HD difference - the gamut difference is certainly less an issue than the decoding errors.

A lot of the RS1 pics that sfogg posted as Lumagen correcting - are of lipstick and sunburn where they do not belong - purely gamut issues as there are no decoding/hue issues.. The complaints are likely not as vocal about red simply because people are used to red pushed video more than they are green. Even when pointed out they say how do we know the actor was not sunburnt and was not wearing lipstick. But when grass is way off - they find it more annoying - be it in the source as in SeaBiscuit or in the PJ or both. And a lot of the complaints came more from seeing the CIE xy gamut diagrams - as if it indicates severe green push - when it is well known in color science that green perceptual ellipses are very large on a CIE xy. Combine that with the well known SD/HD green brightness error when signals are crossed...

Perceptual tests are not the same as annoyance tests - perceptual is can you see a difference and how much difference so it is an objective test, annoyance is how much do you let it bother you which is a subjective test.. Perceptual is valid more for groups - annoyance is valid more for individuals. I let SD reds cast as HD reds bother me - but I was correct in guessing it is around a 10dE at worst - never having done the math before. Most people would not find a 10dE annoying - but they certainly can perceive it. 3-4dE has been established as imperceivable.

You argue in defense of the modern dE from CIE - yet you ignore CIE themselves who say don't use them if the old dE is over 5 - that is not what the newer dE is for.

The CIELUV and CIELCH dE should be the exact same numbers - the only difference is the component breakdown. I figure you must be referring to the dE version that is not a standard -sometimes used for greyscale assuming gamma tracks and it is not valid for colors as it ignores luminance. I don't know what your percentage numbers are (oh I see - percent of RGB dE total - a meaningless number) ...but since the LCH numbers in the last column match the spreadsheet dE I have - I will assume that is dE. Are you seriously suggesting the modern dE is more correct in assigning only 8 or 11 dE to the RS1 Green? About the same magnitude as SD vs. HD Red? But ignoring the magnitude for dE2000 as being way different - it is saying the Reds and Green error is within 3dE of each other and thus imperceivable which is worst error - if you want to assume it is the latest greatest correct standard.

[zoyd]

Originally Posted by TomHuffman

Second, this display provides a large-scale dE error by any definition. But notice that in this instance, it's CIELUV that gives a too-large blue error. And the LCH(uv) calculation seems just plain wrong. Red simply is not the biggest problem with the RS1. The CMS adjustments show this and just about everyone who has ever noticed its color errors focuses on the nearly glowing greens.

What is the difference between CIELUV and LCH(uv)? LCH(uv) (and 1976 CIELAB) errors >=20 dE appears more consistent with your statement of "extreme" color errors with the RS1 compared to the more recent formulations which show errors closer to the SD/HD difference. In your judgement is the 1976 Lab result (green twice as bad as red) closer to what you see than the LCH(uv) (green and red equally as bad)? and how much is the green "annoyance" factor described by Kras playing a role? Also, the Lab76 result of 50 dE is huge, especially if this is with respect to rec.709 (larger than my plasma green primary by 20 dE). From those numbers, the LCH(uv) and Lab76 look closer to RS1 "reality" than the newer formulations. From what I've read from the papers cited earlier, the main difference between the new and older formulations is how they parameterize the effect of lightness errors and from your result it looks like they are de-emphasized relative to the older ones. This is consistent with the difference between emissive and reflective sources. Colored reflective sources absorb incident light (hence subtractive) and reflect a much broader range of light than a corresponding emissive primary, so you can think of them as being desaturated (spectrally) to begin with, the printer's gamut is smaller. That's why prints never look quite as "colorful" as what you see on the screen in photoshop. So my guess is that these newer dE formulations, because they try and model reflective, less saturated, color changes are less well-suited for video displays with larger gamuts.

[stevekale]

I'm by no means an expert (although I don't think I'm completely stupid and I try to look at these things from an intelligent but lay person's point of view), but it seems to me that a bunch of different things are being jumbled together here.

The first is a debate about the usefulness of CIELUV versus CIELAB as a perceptually uniform model. This I care a lot less about. If there is a linearity about the CIELUV model that makes it a computationally or presentationally easier space for video then so be it. However, I've never seen an adequate defence of the statement "CIELAB is more suited to surface color", forgetting about the calculation of colour differences for the moment. I just don't see how fundamentally one model can be better suited to emmissive rather than reflective colour unless it factors in some behavioural aspect of the eye or brain that functions differently in each situation. I suspect that video people find it computationally easier to work in CIELUV for whatever reason but that doesn't mean that that colour model is better suited for emmissive media. Comment?

Now to colour differences. I frankly don't care in which model a colour difference is calculated but rather do care about the accuracy of the resultant calculation, or rather how well such a colour difference formula stands up to emperical testing. It's my understanding, that as people far smarter than me have attempted to improve colour difference formulations their work has been conducted in the LAB space and that testing of these models has been done for both reflective and emmissive media. They could have chosen LUV but didn't. So the improved colour difference models are LAB based and using them represents deployment of best-known colour science.

I agree that a debate over colour differences ignores completely differentials between a particular viewing environment and the standard observer conditions in which the models were formulated (Zoyd's point). All that says, though, is that we may well be wasting out time calibrating in the first place because our calibration models do not and can not factor in other aspects of the viewing environment. So we manage what we can (including obvious viewing environment recommendations).

I think the comment regarding perceptual rather than annoyance differences is simply a case of deciding, when compromises are required, as to which way one should lean and that is simply personal preference.

"You argue in defense of the modern dE from CIE - yet you ignore CIE themselves who say don't use them if the old dE is over 5 - that is not what the newer dE is for." When you say "modern", "them", "old" and "newer", what exactly do you mean? It's interesting to think about the > or < 5 argument in light of the results of the two different computational methodologies to my current calibration: d(uv) has many results > 5 whereas CIEDE2000 doesn't have any.

Perhaps the practical best adage might be to "choose one colour difference method, stick to it and minimise it". But which one.....

At any rate, I would like to see a choice available in ColorHCFR.

Regards

Steve

[zoyd]

Originally Posted by stevekale

I'm by no means an expert (although I don't think I'm completely stupid and I try to look at these things from an intelligent but lay person's point of view), but it seems to me that a bunch of different things are being jumbled together here.

The first is a debate about the usefulness of CIELUV versus CIELAB as a perceptually uniform model. This I care a lot less about. If there is a linearity about the CIELUV model that makes it a computationally or presentationally easier space for video then so be it. However, I've never seen an adequate defence of the statement "CIELAB is more suited to surface color", forgetting about the calculation of colour differences for the moment. I just don't see how fundamentally one model can be better suited to emmissive rather than reflective colour unless it factors in some behavioural aspect of the eye or brain that functions differently in each situation. I suspect that video people find it computationally easier to work in CIELUV for whatever reason but that doesn't mean that that colour model is better suited for emmissive media. Comment?

The question of perceptional uniformity and the ability of a color model to predict color differences are intimately tied together. A perceptually uniform model that covers the entire HVS gamut does not exist. All these models are based on color matching experiments that test different portions of the HVS. The fact that you get better model/data agreement under different conditions tells you that the HVS non-linearities are different in different regimes. One model will not fit all conditions and it appears that there is enough of a difference between reflective and emissive gamuts within present-day print and display technology that some dE formulations will better represent errors in one regime than they do in the other. The only way to decide which is "better" in this type of situation is through experience and what your eyes tell you.

[stevekale]

But there do exist models which cover the gamut of our displays... Even if those models weren't very uniform, if the dE formulae took those uniformity weaknesses into account then the result would be fine. Sure, a perfectly perceptually uniform space would make that task easier but it's not a condition precedent.

I still have trouble with the argument that one model is bettersuited to emmisive versus reflective. With reflective we standardise the illuminant for this very reason (and view at an angle to avoid specular highlights) and our video spec contains a definition of white's colour. All ignore externalities such as surrounding media.

In lay terms, when the director films a red blouse (reflective media) what we care about is that that red blouse is perceived to be the same color red by the average viewer. We just need to define what colour that red is and it's useful to use a device independent space to characterise it. A particular red in LAB has a direct equivalent in CIELUV. Whether that's a red blouse lying on the table, a printed photo of the red blouse or displayed image of the red blouse on the screen, the goal is the same: to generate the same colour. Knowing that nothing will be perfect we then desire to measure perceived error to help us minimise it.