Abstract
Part decomposition is a fundamental process in the representation of complex shape. However, parts are not all created equal: they are emphasized to varying extents in shape representation depending upon their internal geometric properties and their relations to the whole shape. In recent work, we used a segment-verification task to study the relative contributions of different types of curvature extrema in shape segmentation (Cohen & Singh, OPAM, 2003). This method was premised on the logic that segments that constitute natural units in visual representation should be easier to verify. The current study developed a 4AFC segment-identification task to investigate the continuous nature of part representations. On each trial, observers viewed a smooth 2D shape, followed by a contour segment (defined by negative minima of curvature) taken from the shape's boundary. Their task was to indicate the location of the segment on the original shape, by selecting one of the 4 quadrants presented on the screen. The turning angles at the negative minima (hence the strength of the part boundaries) and the lengths of the contour segments were manipulated. The results demonstrated a systematic increase in identification accuracy with both part-boundary strength (defined by turning angle) and segment length. The results allow us to quantify the contributions of turning angle and segment length to part salience, as well as those of other measures that can be derived from these—such as part protrusion (Hoffman & Singh, Cognition, 1997). More generally, the segment-identification task provides a general performance-based method for measuring the graded representation of parts in visual shape, and their geometric determinants.
Supported by NSF grant BCS-0216944.