2.15. Colorspaces

'Color' is a very complex concept and depends on physics, chemistry and biology. Just because you have three numbers that describe the 'red', 'green' and 'blue' components of the color of a pixel does not mean that you can accurately display that color. A colorspace defines what it actually means to have an RGB value of e.g. (255, 0, 0). That is, which color should be reproduced on the screen in a perfectly calibrated environment.

In order to do that we first need to have a good definition of color, i.e. some way to uniquely and unambiguously define a color so that someone else can reproduce it. Human color vision is trichromatic since the human eye has color receptors that are sensitive to three different wavelengths of light. Hence the need to use three numbers to describe color. Be glad you are not a mantis shrimp as those are sensitive to 12 different wavelengths, so instead of RGB we would be using the ABCDEFGHIJKL colorspace...

Color exists only in the eye and brain and is the result of how strongly color receptors are stimulated. This is based on the Spectral Power Distribution (SPD) which is a graph showing the intensity (radiant power) of the light at wavelengths covering the visible spectrum as it enters the eye. The science of colorimetry is about the relationship between the SPD and color as perceived by the human brain.

Since the human eye has only three color receptors it is perfectly possible that different SPDs will result in the same stimulation of those receptors and are perceived as the same color, even though the SPD of the light is different.

In the 1920s experiments were devised to determine the relationship between SPDs and the perceived color and that resulted in the CIE 1931 standard that defines spectral weighting functions that model the perception of color. Specifically that standard defines functions that can take an SPD and calculate the stimulus for each color receptor. After some further mathematical transforms these stimuli are known as the CIE XYZ tristimulus values and these X, Y and Z values describe a color as perceived by a human unambiguously. These X, Y and Z values are all in the range [0…1].

The Y value in the CIE XYZ colorspace corresponds to luminance. Often the CIE XYZ colorspace is transformed to the normalized CIE xyY colorspace:

x = X / (X + Y + Z)

y = Y / (X + Y + Z)

The x and y values are the chromaticity coordinates and can be used to define a color without the luminance component Y. It is very confusing to have such similar names for these colorspaces. Just be aware that if colors are specified with lower case 'x' and 'y', then the CIE xyY colorspace is used. Upper case 'X' and 'Y' refer to the CIE XYZ colorspace. Also, y has nothing to do with luminance. Together x and y specify a color, and Y the luminance. That is really all you need to remember from a practical point of view. At the end of this section you will find reading resources that go into much more detail if you are interested.

A monitor or TV will reproduce colors by emitting light at three different wavelengths, the combination of which will stimulate the color receptors in the eye and thus cause the perception of color. Historically these wavelengths were defined by the red, green and blue phosphors used in the displays. These color primaries are part of what defines a colorspace.

Different display devices will have different primaries and some primaries are more suitable for some display technologies than others. This has resulted in a variety of colorspaces that are used for different display technologies or uses. To define a colorspace you need to define the three color primaries (these are typically defined as x, y chromaticity coordinates from the CIE xyY colorspace) but also the white reference: that is the color obtained when all three primaries are at maximum power. This determines the relative power or energy of the primaries. This is usually chosen to be close to daylight which has been defined as the CIE D65 Illuminant.

To recapitulate: the CIE XYZ colorspace uniquely identifies colors. Other colorspaces are defined by three chromaticity coordinates defined in the CIE xyY colorspace. Based on those a 3x3 matrix can be constructed that transforms CIE XYZ colors to colors in the new colorspace.

Both the CIE XYZ and the RGB colorspace that are derived from the specific chromaticity primaries are linear colorspaces. But neither the eye, nor display technology is linear. Doubling the values of all components in the linear colorspace will not be perceived as twice the intensity of the color. So each colorspace also defines a transfer function that takes a linear color component value and transforms it to the non-linear component value, which is a closer match to the non-linear performance of both the eye and displays. Linear component values are denoted RGB, non-linear are denoted as R'G'B'. In general colors used in graphics are all R'G'B', except in openGL which uses linear RGB. Special care should be taken when dealing with openGL to provide linear RGB colors or to use the built-in openGL support to apply the inverse transfer function.

The final piece that defines a colorspace is a function that transforms non-linear R'G'B' to non-linear Y'CbCr. This function is determined by the so-called luma coefficients. There may be multiple possible Y'CbCr encodings allowed for the same colorspace. Many encodings of color prefer to use luma (Y') and chroma (CbCr) instead of R'G'B'. Since the human eye is more sensitive to differences in luminance than in color this encoding allows one to reduce the amount of color information compared to the luma data. Note that the luma (Y') is unrelated to the Y in the CIE XYZ colorspace. Also note that Y'CbCr is often called YCbCr or YUV even though these are strictly speaking wrong.

Sometimes people confuse Y'CbCr as being a colorspace. This is not correct, it is just an encoding of an R'G'B' color into luma and chroma values. The underlying colorspace that is associated with the R'G'B' color is also associated with the Y'CbCr color.

The final step is how the RGB, R'G'B' or Y'CbCr values are quantized. The CIE XYZ colorspace where X, Y and Z are in the range [0…1] describes all colors that humans can perceive, but the transform to another colorspace will produce colors that are outside the [0…1] range. Once clamped to the [0…1] range those colors can no longer be reproduced in that colorspace. This clamping is what reduces the extent or gamut of the colorspace. How the range of [0…1] is translated to integer values in the range of [0…255] (or higher, depending on the color depth) is called the quantization. This is not part of the colorspace definition. In practice RGB or R'G'B' values are full range, i.e. they use the full [0…255] range. Y'CbCr values on the other hand are limited range with Y' using [16…235] and Cb and Cr using [16…240].

Unfortunately, in some cases limited range RGB is also used where the components use the range [16…235]. And full range Y'CbCr also exists using the [0…255] range.

In order to correctly interpret a color you need to know the quantization range, whether it is R'G'B' or Y'CbCr, the used Y'CbCr encoding and the colorspace. From that information you can calculate the corresponding CIE XYZ color and map that again to whatever colorspace your display device uses.

The colorspace definition itself consists of the three chromaticity primaries, the white reference chromaticity, a transfer function and the luma coefficients needed to transform R'G'B' to Y'CbCr. While some colorspace standards correctly define all four, quite often the colorspace standard only defines some, and you have to rely on other standards for the missing pieces. The fact that colorspaces are often a mix of different standards also led to very confusing naming conventions where the name of a standard was used to name a colorspace when in fact that standard was part of various other colorspaces as well.

If you want to read more about colors and colorspaces, then the following resources are useful: poynton is a good practical book for video engineers, colimg has a much broader scope and describes many more aspects of color (physics, chemistry, biology, etc.). The http://www.brucelindbloom.com website is an excellent resource, especially with respect to the mathematics behind colorspace conversions. The wikipedia CIE 1931 colorspace article is also very useful.