Color management myths and misinformation video

[GWGill]

That sounds very reasonable. There is some debate if Lab is truly a perceptually uniform colorspace.

There's no debate - it's not a perfectly perceptually uniform space, and its many flaws have been well documented :-)

But it's much more perceptually uniform than XYZ or most device spaces, is widely accepted and understood, and there is no accepted replacement for it.

[smthopr]

This stuff all gets very complicated.

Color spaces, especially device color spaces are not necessarily uniform. A "bigger" space does not need to encompass all of a "smaller" space. This all gets too abstract to explain to novices.

I would like to suggest that you start the lesson in monochrome. Then, you could accurately describe these spaces with images of greyscale step wedges. Each containing the some number of steps, but encompassing different ranges of tone. I think this is the best analogy to begin with.

If someone would like to continue along this line, please have at it:)

[fdisilvestro]

There's no debate - it's not a perfectly perceptually uniform space, and its many flaws have been well documented :-)

But it's much more perceptually uniform than XYZ or most device spaces, is widely accepted and understood, and there is no accepted replacement for it.

Bruce Lindbloom has proposed a "Uniform Perceptual Lab" profile, you can read about it here

[Jim Kasson]

There is some debate if Lab is truly a perceptually uniform colorspace.

I don't think any color scientist would say that CIEL*a*b* is truly perceptually uniform. MacAdam-ish discrimination ellipsoids don't map to uniform spheres in Lab. Or Luv. Or any other space that I know of.

That doesn't mean that perceptually uniform spaces are not worthy Platonic ideals.

Jim

[fdisilvestro]

A "bigger" space does not need to encompass all of a "smaller" space

While this might be true about color spaces, in the case of this discussion, Abobe RGB does encompass all of sRGB.

I agree that this can become too abstract and we can get easily carried away in technicalities

A possible source for learning are the tutorials at Cambridge in Colour here. In this comparison between Adobe RGB and sRGB i find it useful the color gamut graphs at the bottom of the page, showing the difference when you use CIE xyz vs. CIE u'v', the latter giving a better approximation to what we really perceive

[MarkM]

All possible colors in sRGB can be described in Adobe RGB
Some colors in Adobe RGB are out of sRGB

Result: More colors can be described in Adobe RGB

This isn't precisely true. sRGB has greater precision that AdobeRGB so there will be sRGB values that cannot be captured by adobeRGB to the same precision. Or to put another way, there will be pairs (or more) sRGB numbers that are described by a single AdobeRGB value.

On reflection, the thermometer analogy has a serious flaw. Alan Goldhammer hit on it. Here's the problem.

Although you might say that if I take one measurement from each thermometer I have the same amount of information from each, but that may not be true - all measurements are not the same. The amount of information a measurement has depends on the precision and scale of the system (I think this is a simplification of what information theory calls the entropy of the system). Say  both thermometers go from 0-100. If one thermometer measures to a hundredth of a degree I will get a measurement like 90.75, but another might only measure to a tenth of a degree so I'll get a measurement of 90.8. There first measurement contains more information. To express it on a computer would require me to distinguish all values between 0 and 100 to hundredth of a degree precision - significantly more values than the 0-100 at a tenth degree precision.

Now if we go back to the original thermometers 0-100 and 0-200 and imagine both have a precision to 1 degree. Measurement from the 0-200 will have more information — they will require 1 more bit to store (i.e you need to distinguish between 201 values as opposed to 101). The situation changes, however if the scale changes. Rather than 1 degree precision, lets say they thermometers are each marked off from 0 - 255 (and readings between the lines is meaningless). Now a measurement from each contains precisely the same amount of information, exactly 8 bits. The 0-100 thermometer is twice as precise, but the 0-200 thermometer has twice the range.

This second scenario is exactly the situation we find ourselves in with colorimetric values. Because the amount of information contained in a pixel is given by definition of the encoding, the space is not relevant. You can have a space 50 billion times larger than proPhoto, but you still have 24 bits of information by definition.

Whether you call these individual values colors is an entirely different and mostly semantic discussion. In this case it would probably help to use a different term, such as colorimetric value, although that's not so great for explaining to beginners. But they are different concepts and some confusion is happening because we are using the same term for both. If instead of defining color encoding by bit depth we decided on a fixed unit of measurement say delta-e's from zero and used that to specify our colors, colors from the AdobeRGB would in fact have more information, but they would also require more bits to store on the computer.

[fdisilvestro]

This isn't precisely true. sRGB has greater precision that AdobeRGB so there will be sRGB values that cannot be captured by adobeRGB to the same precision. Or to put another way, there will be pairs (or more) sRGB numbers that are described by a single AdobeRGB value.

On reflection, the thermometer analogy has a serious flaw. Alan Goldhammer hit on it. Here's the problem.

Although you might say that if I take one measurement from each thermometer I have the same amount of information from each, but that may not be true - all measurements are not the same. The amount of information a measurement has depends on the precision and scale of the system (I think this is a simplification of what information theory calls the entropy of the system). Say  both thermometers go from 0-100. If one thermometer measures to a hundredth of a degree I will get a measurement like 90.75, but another might only measure to a tenth of a degree so I'll get a measurement of 90.8. There first measurement contains more information. To express it on a computer would require me to distinguish all values between 0 and 100 to hundredth of a degree precision - significantly more values than the 0-100 at a tenth degree precision.

Now if we go back to the original thermometers 0-100 and 0-200 and imagine both have a precision to 1 degree. Measurement from the 0-200 will have more information — they will require 1 more bit to store (i.e you need to distinguish between 201 values as opposed to 101). The situation changes, however if the scale changes. Rather than 1 degree precision, lets say they thermometers are each marked off from 0 - 255 (and readings between the lines is meaningless). Now a measurement from each contains precisely the same amount of information, exactly 8 bits. The 0-100 thermometer is twice as precise, but the 0-200 thermometer has twice the range.

That might be true if you are limiting to 8 bit Jpegs, but not if you use 16 bit Tiffs (or 15 bit as Photoshop really works).
The abstraction of color spaces per se does not limit the amount of digits you can use

[MarkM]

That might be true if you are limiting to 8 bit Jpegs, but not if you use 16 bit Tiffs (or 15 bit as Photoshop really works).
The abstraction of color spaces per se does not limit the amount of digits you can use

It doesn't matter if you are using 16 bits or 10 billion bits — sRGB will always be more precise for a given bit depth because those bits will be spread across a smaller area than adobeRGB. Maybe, I'm misunderstanding you on this…

[GWGill]

Bruce Lindbloom has proposed a "Uniform Perceptual Lab" profile, you can read about it here

There have been many efforts to produce more perceptually uniform colorspaces. Many of them are of a similar form to a cLUT profile - ie. numerically optimizing against CIEDE2000, McAdams elipses, CIECAM02 space etc., but this form is not very useful for general use - a simple set of equations is what's needed. Something like IPT space or DIN99 comes close to that (though neither are as simple as L*a*b*), but so far nothing seems to have hit the "sweet spot" of being noticeably better than L*a*b*, simple enough to be widely useful, and widely or officially accepted.

[fdisilvestro]

It doesn't matter if you are using 16 bits or 10 billion bits — sRGB will always be more precise for a given bit depth because those bits will be spread across a smaller area than adobeRGB. Maybe, I'm misunderstanding you on this…

Do not rely on the quantity of digits, we could add bits if necessary, 100, 1000, 1Million bits? Yes you will have more density or "precision" for a smaller space but it will be practically a waste of digits, there is a point where adding numerical precision does not lead to any practical difference. Once we hit the noise floor, there is no point in adding precision. You could just generate random numbers for the least significative bits and the result will be the same.

[digitaldog]

sRGB has greater precision that AdobeRGB so there will be sRGB values that cannot be captured by adobeRGB to the same precision.

With the same encoding, because the colorimetric distance is father apart in the wider gamut space? (the pixels are farther apart  ).

Although you might say that if I take one measurement from each thermometer I have the same amount of information from each, but that may not be true - all measurements are not the same. The amount of information a measurement has depends on the precision and scale of the system (I think this is a simplification of what information theory calls the entropy of the system). Say  both thermometers go from 0-100. If one thermometer measures to a hundredth of a degree I will get a measurement like 90.75, but another might only measure to a tenth of a degree so I'll get a measurement of 90.8. There first measurement contains more information. To express it on a computer would require me to distinguish all values between 0 and 100 to hundredth of a degree precision - significantly more values than the 0-100 at a tenth degree precision.

Analogous to the encoding, 24 bit/48 bit finer possible ways to divide up the data?

This second scenario is exactly the situation we find ourselves in with colorimetric values. Because the amount of information contained in a pixel is given by definition of the encoding, the space is not relevant. You can have a space 50 billion times larger than proPhoto, but you still have 24 bits of information by definition.

Makes sense.

[Eyeball]

But that's pretty important. As already pointed out, color, is a perceptual property. So if you can't see it it's not a color. A coordinate in a "colorspace" outside the spectrum locus is not a  
color. Color is not a particular wavelength of light. It is a cognitive perception, the excitation of photoreceptors followed by retinal processing and ending in the our visual cortex, within our brains. As such, colors are defined based on perceptual experiments. And from that, we get deltaE.

I understand what you're saying and probably agree with most, if not all, of it.

It is hard for me to put together a coherent position on this so let me just state some things that bother me about depending on the "delta-e definition" of what a "color" is:

- I know from some study that a low delta-e (in particular <1) means that a human won't be able to distinguish the difference between two near-by colors.  The thing that impressed me though is that those gamut volumes from the Profile Inspector were much lower than even 8-bit color could convey.  Would I really NOT see a difference if I took an 8-bit image (16.8 million potential color values) and uniformly reduced "rounded-down" the values to only 1.2 million (aprox. delta-e volume for AdobeRGB)?  Makes sense on the one hand but hard to believe on the other.  By the way, Bruce Lindbloom calculated only 2.9 million for ProPhoto and I suspect that includes the "imaginary" colors that ProPhoto contains.

- Being able to differentiate two colors and being able to "perceive" them seem to me to be to be slightly different things.  For example, I have color X and color X'00001 and they are so close in hue, chroma, and lightness that I can't tell them apart.  Given that, which one then is THE "color" and which one is not?  Now the easy answer is "since color is human perception and I can't tell them apart, they are the same "color" no matter what instrumentation may tell me." OK.  Now I look at X'00001 and X'00002 and determine the same thing.  But wait, I can SEE a difference between X and X'00002 and what happens if I started my comparison somewhere else?  How do I determine what instrument readings get disqualified as "color" and which don't?  I have a feeling that there is an important distinction between "colors" and "uniquely perceptible colors".

- If you do accept though, that there is SOME finite limit to the precision that color can be measured (either by a human or by an instrument), then I guess you could state that "AdobeRGB can contain a greater number of colors than sRGB".  Why?  Because the precision would be the same for both color spaces but the volume of one is greater - therefore the larger-gamut color space could hold more "colors".  Notice though that I am still not saying that AdobeRGB "has more colors";  I am still just making a statement of volume although wording it differently.

[Slobodan Blagojevic]

... I guess you could state that "AdobeRGB can contain a greater number of colors than sRGB".... Notice though that I am still not saying that AdobeRGB "has more colors"...

Hmmm...  "a greater number" does NOT equal "more"?

[fdisilvestro]

I would really like to see a real example where thanks to the "increased precision" of sRGB it is possible to differentiate two colors that otherwise will appear as one color in AdobeRGB due to the "lower precision" of using the same number of digits in both spaces.

[Eyeball]

Hmmm...  "a greater number" does NOT equal "more"?

Substitute "more" for "a greater number". No problem.  The important difference for me is between "has" and "can contain".

[Slobodan Blagojevic]

Substitute "more" for "a greater number". No problem.  The important difference for me is between "has" and "can contain".

Ah, OK. So, when does "can contain" transition to "has"? When the subject has colors outside of sRGB?

[Tony Jay]

Just to respond to several posts and streams of debate:

Colour, as Andrew has already intimated, is NOT colour until it is perceived.
In addition, even if the conditions are constant i.e. the same wavelengths of light are involved invoking the same reflective and transmissive properties of a the substance or surface(s) being viewed several different individuals may, in fact, not perceive the viewed colours as identical.
An, obviously pathological, example of this phenomenon is colour-blindness but on a population-wide basis we all have subtle variations in how our cones (not rods) perceive colour.
That does not take into account the vast amount of post-processing that takes place first in the retina itself (the retina is not just a dumb image receptor but in fact an outgrowth of grey matter from the brain) and then in the visual cortex (this occupies a large part of the occipital cortex of both hemispheres of our brains) that eventually results in our perception of colour.

One, but not unique, link with the whole point of colour management is before the issue of accuracy of colour reproduction (and then, by extension perception) can be addressed the issue of consistency needs to be sorted. This is the essential reason for the existence of colour management in the first place. If the same "colour" is displayed or printed on a specific paper and viewed in specific lighting conditions then whatever each of us actually perceives should be consistent for each of us irrespective of the accuracy of that colour compared to some standard.
Once that is sorted one can then look to "accuracy".

Because the issue of consistency of colour across different platforms and output modes (display versus print etc) is the key theme in colour management and actually the number one concern of nearly every photographer at any and every level of expertise, colour management is an intimate and inextricable part of any photographic workflow (this is not always appreciated), not just those who choose to shoot RAW and use a so called "fully colour managed workflow".
Even individuals who attach their camera straight to a printer and just press 'print' are unwittingly invoking colour management principles whether they are aware of it or not.
Those who do understand the relevant colour management principles will still get better results in that situation than those who don't.

So photography in general cannot be divorced from a fundamental understanding of colour management.

Tony Jay

[digitaldog]

The important difference for me is between "has" and "can contain".

Indeed. So we go back to the image itself, the encoding, other possible factors. Meaning it's probably not a good idea to say "Adobe RGB has more colors than sRGB" any more than "New Mexico has a larger population than New York".

[Slobodan Blagojevic]

..."New Mexico has a larger population than New York".

Soon, soon...

[Eyeball]

Ah, OK. So, when does "can contain" transition to "has"? When the subject has colors outside of sRGB?

Perhaps a more correct way to say it is that an image in the AdobeRGB color space can contain more colors than one in the sRGB color space.
The color space does not "have" the colors, the image does.
The "can" then changes to "has" when you have identified a particular image and determined if it does or does not.