Previous: Color Naming
Up: Problem Definition
Outline of the Model
I will define a computational model of human color perception and color
naming, i.e. construct an algorithmic mapping 
as defined in
Section 
, which is based in part on existing data about the
neurophysiology and psychophysics of color perception, and which can
explain existing anthropological and linguistic data on color naming as
well. The model should allow an artificial cognitive agent (e.g. a
GLAIR-agent [Shapiro \& Rapaport 1987][Hexmoor et al. 1993b][Hexmoor et al. 1993c][Lammens et al. 1994]), when
equipped with the necessary sensors and actuators, to
- Name colors in response to a visual stimulus, and provide a
confidence or ``goodness of example'' rating of its judgment; this requires
evaluation of
and some kind of thresholding on the resulting
pairs 
.
- Point out examples of named colors in its environment,
and provide a confidence rating, and as a derivative of this capability,
pick the best example of a named color from a set of color samples,
or from its environment in general; this requires evaluation of
over the whole visual field, and some kind of maximization on the
resulting pairs .
The performance on these tasks must be consistent with human performance on
the same tasks, as described in Chapter . In particular,
this requires that
- The model place the foci of basic color categories, as
described in [Berlin \& Kay 1969] and elsewhere, in the same regions of the color
space as human subjects do
- The model place the boundaries of basic color categories in
the same regions of the color space (cf. Chapter ) as
human subjects do.
A more precise definition of ``same region in the color space'' will be
given, to determine success in this area.
The model deals only to a limited extent with a number of important issues
in color vision, most notably the effects of surrounds on perceived color,
or in general the relation between spatial and color vision, and color
constancy. These issues are dealt with only in as far as they are relevant
to color naming.
The model has been integrated into a vision system capable of interacting
with its environment as described above. I will refer to this system as the
Color Labeling Robot (CLR),
after the
Color Reader Robot described as a thought experiment in
[Hausser 1989]. The model has also been tested on simulated data, and
the results compared to known data about human color naming.
A system that exhibits the behavior described above is an example of a
(partly) embodied [Kay \& McDaniel 1978][Lakoff 1987] or grounded
[Harnad 1990] system, or it can be seen as an instance of
situated cognition [Suchman 1988].