Abstract

Some models of 3-D object recognition rely on the detection and classification of viewpoint-invariant geometrical features of 2-D images. In an investigation of the extent to which the human visual system is differentially tuned to viewpoint-invariant features vs viewpoint-dependent features, we measured the statistical efficiency with which observers can classify curved contours (circular arcs of fixed arc length) into pretrained categories. Optimal tuning to viewpoint-invariant features predicts that higher efficiencies would be obtained when one experimental category included contours with zero or near-zero curvature (i.e., near-straight lines). Prior to experimental trials, observers were shown sample curves from two curvature categories, defined as sets of curves with Gaussian distributions of different mean curvature. In a forced choice procedure, they then indicated the category of origin for curves randomly selected from the two categories. Efficiency was taken as the ratio of the observer’s d-prime and the d-prime of an ideal detector. Typically, efficiencies were greatest (as large as 80%) when one of the category means was close to zero curvature. When the categories were moved from zero curvature, efficiencies decreased by as much as 30%. These results are consistent with the exploitation of viewpoint-invariance differences in object recognition.

© 1991 Optical Society of America