Abstract
We investigated the extent to which the recognition of images is viewpoint dependent across rotations in the picture plane. The experiments employed a set of novel images that made it possible to examine the influence of the number of parts composing the image. The images consisted of from 2 to 8 cylindrical parts arranged radially within a circular space with 10 possible part locations. Reaction time to correctly identify a rotated version of a sample image revealed two main effects. First, the number of parts strongly determined RT, such that RT was fastest when there were either few (e.g. 2) or many (e.g. 8) parts, but was slowest when there were an intermediate number of parts (e.g. 5), forming an inverted U-shaped function. Second, the viewpoint dependence of the images, as measured by the slope of the RT function for an image across 30 to 120 degrees of rotation, varied strongly with number of parts. Images with few (2) or many (8) parts displayed flat slopes, revealing a lack of viewpoint dependence. Images with intermediate numbers of parts (5) displayed high positive slopes, showing strong viewpoint dependence. The inverted U-shaped functions for RT and viewpoint dependence mirror the function of total possible shape configurations for a given number of object parts. For the radial images used here, there were a smaller number of possible configurations for objects with low or high numbers of parts (e.g. 45 possible 2-part images: all combinations of 2 locations taken from a space with 10 possible part locations) while there were high numbers of possible configurations for the middle range (e.g. 252 5-part images: all combinations of 5 from 10). This suggests that image complexity, measured by the size of the “shape space” of possible forms implied by the number of image parts, influences how well the representation of an image generalizes to new viewpoints.
AC supported by a DOD NDSEG Fellowship and the Center for the Neural Basis of Cognition.