Previous: Theoretical Evaluation of the Category Model Up: From Color Space to Color Names Next: A Sketch of a Developmental Model

Comparing Performance for Different Color Spaces

In the preceding section we have looked at differences in performance of the categorial model with respect to different underlying color spaces. We now take a closer look at why those differences might arise, i.e. if there are any intrinsic properties of the color spaces that might explain the differences in performance.

First of all it is important to note that both the XYZ and the L*a*b* are psychophysical color spaces, defined by the Commission Internationale de l'Eclairage (International Lighting Committee), based on color perception experimental data averaged over large populations. In contrast, the NPP color space is based on neurophysiological recordings made from a single individual (and a monkey at that). As such it should be representative of a particular individual, but not necessarily of a population mean. The color naming data of Berlin and Kay is also averaged over a number of experimental subjects, so it is likely that theirs would be a better fit with the CIE spaces than with the NPP space, when considering only the way both color spaces and naming data were constructed or collected.

The categorial model itself is of course also a psychological model, based on experimental observations over a wide range of tasks and sizable populations [Shepard 1987], so again we would expect similarly constructed color spaces to be at an advantage. The model works best on the L*a*b* space, which is meant to be a perceptually equidistant space, meaning that a fixed distance between two points anywhere in the space should correspond to the same magnitude of perceptual difference (e.g., the perceptual distance between the top two gray scale samples in Figure should be the same as the distance between the next two). This is not surprising giving the categorial model, which assumes a Euclidean distance measure on the underlying space, i.e. the perceptual differences should be equal in all directions and in all areas of the space.

The reason the NPP space does not perform better may be related to the scaling issues discussed in Section . For instance, the green region in the NPP space seems to be compressed and low on the brightness dimension, compared to the L*a*b* space (Figure ). This means that the NPP space is not perceptually equidistant. The reason for this could be either that at the LGN stage of color perception, there is no perceptual equidistance yet, and equidistance is an effect of a ``higher'' stage in the neurophysiology, or it could be that using a different scaling method might result in a more equidistant space. But the latter could itself be seen as modeling a higher stage in neurophysiology, of course. I have not investigated this question any further yet, but it is certainly an interesting one.

Finally, the error metric as developed in the preceding sections does give us a quantitative basis for comparing the performance of different color spaces, but it is a less than perfect one. For instance, it does not take into account how ``serious'' a categorization error is: categorizing very pale blue as white is not as serious as categorizing green as red. It is difficult to include such qualitative judgments into the error metrics, but doing so might give somewhat different results.

lammens@cs.buffalo.edu