Color Constancy, Adaptation, and Why Colors Change with Context

Color constancy is not one correction. It is a negotiated estimate: cone gains discount the illuminant, spatial comparisons stabilize surfaces, learned memory colors pull familiar objects toward their expected appearance, and a camera-style white balance tries to do the same thing with explicit channel multipliers. This workbench puts those mechanisms in one interactive model.

The light reaching your retina from any surface is the product of two things: reflectance (a property of the surface, fixed across time) and illumination (a property of the light source, hugely variable across a day). The visual system has only the product to work with - but somehow recovers, most of the time, an approximately stable estimate of the underlying reflectance. That stable estimate is what we call the object's color.

Without color constancy, every change in lighting would look like a change in object color. The world would be visually chaotic. With color constancy, the apple stays red as you carry it from the kitchen to the patio, and the white shirt stays white as the sun rises and sets. The system is imperfect - measured constancy is roughly 60-85% across typical illuminant changes - but good enough that surface identity dominates over lighting in conscious experience.

Color constancy is not one mechanism. It is at least three, layered on top of each other, operating at different stages of the visual pipeline and on different timescales.

Together, these three levels produce constancy on timescales from milliseconds (lateral inhibition) to half an hour (full dark adaptation). They also explain why constancy can be partly defeated by removing context - look at a surface through a tube and its color shifts noticeably as the illuminant changes, because the cortical context machinery has lost its scene-wide cues.

The simplest and oldest model of chromatic adaptation comes from the German physiologist Johannes von Kries (1902). It says each cone class rescales its gain independently so that the chromatic statistics of the current scene approximate "white." The cone response under the new illuminant is multiplied by a constant - the gain - chosen so the system's perceived white matches the new white point.

The model is crude but works surprisingly well as a first approximation. Modern color-management systems use refined versions of it for cross-illuminant transforms (Bradford, CAT02, CAT16), where the gains are computed in carefully chosen "sharpened" cone bases rather than the raw L, M, S directions.

Cones rescale to global averages. But the retina also performs local rescaling through lateral inhibition - each cell's output gets pulled toward whatever its neighbors are doing. The consequence is simultaneous contrast: a patch surrounded by a bright region looks darker; surrounded by a saturated red, it looks more greenish. The eye is reporting contrast against the surround, not absolute lightness or color.

In the 1970s, Edwin Land - inventor of the Polaroid camera - mounted a striking demonstration. He projected three black-and-white slides of a colorful "Mondrian" scene through red, green, and "white" filters, and varied the intensities so that the light from any one patch could be made physically identical across very different filter conditions. Observers still saw the underlying colors correctly. The final color of a patch did not depend on its local light - it depended on the relationships across the whole scene.

Land's Retinex theory (1977) was the first serious attempt at a computational model of color constancy. It proposed that the visual system computes each cone-class lightness independently by integrating edge ratios across the scene - effectively reconstructing reflectance from a network of local comparisons. The model still influences computer-vision algorithms half a century later.

Modern color-management systems implement chromatic adaptation as a 3×3 matrix transform between XYZ tristimulus values under different illuminants. The most widely used transforms - Bradford, CAT02, and CAT16 - all share the same structure: convert XYZ into a "sharpened" cone basis, apply independent von-Kries-like gains, then convert back to XYZ.

Color constancy works by making assumptions about the scene. When those assumptions are wrong, the conclusions are too. The most famous color illusions are not failures of the visual system but consequences of its normally correct reasoning applied to unusually crafted images.

Constancy gets help from learning. Skin, sky, grass, bananas, and Coca-Cola red have memory colors the visual system has internalized over a lifetime of viewing. When light is ambiguous, the brain biases its interpretation toward the memorized typical color. Show a person a desaturated banana and ask them to adjust it until it looks gray - they will overshoot toward blue, because their internal model "knows" the banana should be yellow.

Three conditions reliably defeat color constancy. Knowing them is useful for both designers (avoiding them) and illusion designers (exploiting them).

CIELAB gives you device-independent coordinates for a color under standard reference conditions. But it does not predict how a color will actually appear when conditions change - viewing angle, surround, background, luminance level, or partial adaptation. Color appearance models (CAMs) fill this gap.

The current standards are CIECAM02 (2002) and its successor CIECAM16 (2016). Both take an input stimulus plus a description of viewing conditions and output six appearance correlates: lightness, brightness, chroma, saturation, colorfulness, and hue angle. These are designed to capture the way a color actually looks, not just where it sits in XYZ space.

Camera sensors do not have constancy. They record what hits them. To make photographs look natural, every camera system performs white balance - an explicit, software-driven version of biological chromatic adaptation. White balance multiplies the R, G, and B channel gains so that something the camera thinks is white actually comes out white in the final image.

Six questions on constancy, adaptation, and context. Wrong answers come with brief explanations.