| ECVision indexed and annotated bibliography of cognitive computer vision publications This bibliography was created by Hilary Buxton and Benoit Gaillard, University of Sussex, as part of ECVision Specific Action 8-1 The complete text version of this BibTeX file is available here: ECVision_bibliography.bib
J. M. Lammens and S. C. Shapiro
ABSTRACT
We are working on a computational model of color per ception and color naming, which can be seen as an in stance of symbol grounding [ 4 ] in the domain of color, or as an attempt to provide an artificial intelligent agent with embodied concepts of color. This effort is part of a larger one in the field of intelligent autonomous agents [ 8, 5 ] , and will provide a perceptual grounding for some of the symbolic representations in the SNePS Knowl edge Representation formalism [ 13 ] , especially those re ferred to as ``sensory nodes'' [ 12 ] . The implemented model will allow an agent to name colors in its environ ment, point out examples of named colors in its envi ronment, and learn new names for colors. Our research draws on work in the neurophysiology of color percep tion, particularly [ 3 ] , in semantic universals for natural languages, particularly [ 1 ] , and other work in AI and Cognitive Science. We discuss two areas of the model where learning is used: learning a nonlinear mapping between two color spaces, and learning a relation between color space co ordinates and a set of symbolic color names. We have used a traditional error backpropagation learning al gorithm for the first problem, and are considering sev eral different learning paradigms for the second problem, ranging from traditional clustering techniques to an ex perimental ``space warp'' method. Using learning gives us a relatively easy way to determine a nonlinear trans formation of spaces in the first case, and increases the flexibility of the approach with respect to different ap plication needs (and languages) in the second case. We discuss the learning methods used or considered and the problems encountered. In general terms, our model has to explain (and repro duce) a signaltosymbol transition, going from light en tering a sensor to symbols representing the correspond ing perceived color. To make our problem manageable, we make some simplifying assumptions. We are only concerned with singlepoint determination of color, thus disregarding spatial interactions in color perception. We only take context into account to the extent that it is necessary for this determination. We assume foveal cone photoreceptors as sensors, and we restrict the problem \Lambda This work was supported in part by Equipment Grant No. EDUDUS932022 from SUN Microsystems Computer Corporation. to any given fixed state of adaptation of the vision sys tem, thus avoiding issues of color constancy [ 2, 14 ] . We are developing a physical implementation of an agent, using a color camera for sensing device and a robot arm for pointing device [ 8 ] , and there we will deal with these issues to some extent, but that does not concern us here. Conceptually, we break the model up into two parts. The first part takes us from the visual stimulus to color space coordinates, the second from color space coordi nates to a set of color names. |
|
| Site generated on Friday, 06 January 2006
|