Color Constancy

The light $l(\lambda)$ the eye receives is the product of the illumination $i(\lambda)$ of the light source and the reflectance $r(\lambda)$ of the surface:

$$L(\lambda)=R(\lambda) I(\lambda)$$

But usually we can tell the color of an object, $R(\lambda)$, relatively independently of the many different possible illuminations $I(\lambda)$ (e.g., overhead sun, setting sun, overcast sky, indoor tungsten light or fluorescent light, etc.). For example, a banana looks yellow under all of these lighting conditions. This is called color constancy, which can be more convincingly demonstrated by the follow experiment by Land (1959).

In a dark room the audience was asked to look at a multicolorred display made of patches of paper of different color, pasted together. Such a display is called a color Mondrian because of the resemblance to the paintings of the Dutch painter Piet Mondrian. The color Mondrian was illuminated by three projectors each projecting filtered light of narrow spectrum: red (630-700 nm), green (525-565 nm) or blue (400 - 470 nm) on the display with adjustable intensity. These projectors were first turned on one at a time with its intensity adjusted so that the the reading of a photometer receiving the light reflected from one particular patch, say a green one (under some white illumination), was always 1 for each of the three projectors. When all three projectors were turned on the audience reported that the patch was green (instead of gray).

Next a patch of a different color, say, orange, was selected and the above process was repeated. The intensities of the projectors were readjusted so that when each one of them was on, the reading of the photometer receiving reflection from the orange patch was again 1, the same as before. When all three projectors were turned on, the audience reported that the patch was still orange, in spite of the fact that the physical light reflected from the orange patch was the same as what the green patch had reflected earlier!

If the same procedure was repeated with exactly the same setting except the green and orange patches were viewed in isolation without the surrounding patches (``void condition''), the audience reported the patches looked the same color, say, light gray.

The Land's Mondrian experiment demonstrates clearly the color constancy of human vision. That is, human color vision has the ability to perceive the the surface, relatively independent of the illumination. Similar experiments were carried out to test goldfish and the same color constancy resulted.

Land's experiment can be intuitively explained as follows. Suppose the first patch looks reddish under normal lighting. In order for this patch to reflect the same amount of light (same reading on the photometer) under each individual illumination light (R, G, B), the green and blue projectors have to be adjusted so that these lights have higher intensity than the red light. In other words, when all three projectors are tuned on, the overall illumination of the scene has a bluish/greenish tone. Similarly, when the process is repeated on a patch which looks bluish under normal lighting, the projectors will be so adjusted that they produce a yollowish (mixture of red and green) illumination. When looking at such a multicolor display, one's visual system senses not only the light from some specific (local) patch, but also the overall (global) tune of the illumination, which is biased in one way or another. Due to some unknown process of spatial comparison, the visual system is able to extract the color information of a local region, relative to the overall background. As the result the two patches in the experiment appear as different colors perceptually, while their reflect identical lights physically. To conclude, we see that it is this overall tune of the illumination from various sources (overhead sun, setting sun, overcast sky, indoor tungsten light or fluorescent light) that carries some extra information in addition to the light reflected locally from a specific object, and thereby the color constancy.