Previous: Pointing Out Examples of Colors Up: Putting It All Together: From Visual Stimuli to Color Names, and Back Next: Semantics, Grounding, and Truth

Choosing Objects by Color

When combined with an image segmentation algorithm, we can use the same techniques to choose objects by color. For instance, if a robot is instructed to ``get the red foozle'', and it does not have the capability to distinguish foozles from non-foozles, it can still make an educated guess provided it can segment foozles from the background, and there aren't too many red objects around. Color provides another constraint for determining the referents of expressions, and while it may not be sufficient to determine the referent uniquely, it may provide enough constraint to enable a unique determination in combination with other constraints such as shape and size, even if none of these alone would be sufficient. Interestingly, some non-basic color names are so specific that they practically provide all the information necessary to pick out the intended (class of) referent(s). Consider for instance a term like ``blond'', which is applicable to very few object classes only (mainly hair, possibly beer too). If one has a perceptual category for such a term, one can pick out the intended referent by color only.

Another interesting observation in this respect is that the robot's color perception mechanism need not be as good to perform this task (discrimination) as to perform the naming or pointing tasks. The categories may be considerably ``wider'' and more overlapping, or the robot's idea of what constitutes a particular color may vary to some extent compared to the agent issuing the request, as long as the colors of potential referents are distributed widely enough throughout the color space. This task is probably best performed with , to avoid quibbles over whether or not a particular object color is a good enough example of the requested kind (or in other words, to be maximally cooperative).