Previous: Naming Color Samples Up: Putting It All Together: From Visual Stimuli to Color Names, and Back
Next: Choosing Objects by Color
The complete process required for pointing out an object (or a blob) of a
named color in one's field of view is schematized in
Figure .
Compared to the color naming process, we also need a pointing device (a
color monitor is suggested in the diagram, but other pointing devices like
a robot arm could be used), a complete color image (frame) rather than just
a blob, and an index into the image for each blob to be categorized. In
Chapter , I present an application along these lines,
and further implementation detail is provided there.
When given a color name to point out an example referent of, we sample the
image using a certain sample (blob) size, collecting the averaged device
RGB values and the image coordinates of (the center of) each blob, and
transform the RGB values to the color space of choice, which gives us a set
of points in color space with corresponding image
coordinates:
where is a point in color space,
is the
position vector of the corresponding (center of the) blob in the image
(field of view), and
is the total number of samples (blobs).
We then determine the membership (goodness) value of the samples in the appropriate category:
where is the normalized Gaussian as before, and
is the category
with which the name
is associated, with its corresponding parameters
and
.
Next we select all samples with a membership value exceeding the threshold
for category membership :
These are all candidate referents for the name N.
As before, we sort the candidates by decreasing membership value:
If we just want any referent for the category named N, we pick an arbitrary
element from this tuple (or from set B), and if we
want the best example we take the first element
.
After selecting a referent , all that is left to do is point it out in
the image (or in the world giving rise to the image, which is harder to
do), using the index vector
.
As is the case with color naming, we can do forced choice experiments if we
use a zero threshold, in which case a referent will always be selected
regardless of how good an example it is, or we can do free choice
experiments using the standard threshold , in which case a referent
will only be selected if it is a ``good enough'' example of the category in
question.