Abstract:

How can robot programmers best express algorithms for parsing the visual world? Even a simple color-segmented planar world poses many

interesting perceptual problems. I will argue that the most natural approach uses two types of representations: iconic (pixel-based) and

algebraic (shape-based). This "dual-coding" approach to vision, in which the two representation types can interact when necessary, was

inspired by Alan Paivio's dual-coding theory of mental representations. My students and I have implemented the dual-coding

idea in the Tekkotsu robot programming framework and used it to teach robot vision programming to undergraduates. An important factor in

our success has been an interactive GUI tool with which students can examine the robot's mental representations on the fly. Our approach

can potentially be extended in several interesting ways, drawing inspiration from Shimon Ullman's proposal for visual routines. Two

examples are incorporating gestalt perception operations, and adding mechanisms for control of visual attention.

Dual-coding representations for robot vision programming in Tekkotsu