Abstract
Attneave (1954) demonstrated that points of high curvature were most important for object recognition. While this conclusion has become the basis for much of modern computer vision theory, it is not consistent with psychophysical data showing that curvature, being a metric property, is not reliably perceived across viewpoints (Biederman and Bar, 1999). Having low viewpoint stability, one would not expect curvature to be a key component of a process as robust as object recognition. This weakness was further demonstrated by Foster and Gilson (2002) who, using a wire-shape discrimination task, found that changes in curvature were less discriminable than changes in the number of segments and as discriminable as segment length and angle of segment intersection (AOI). Assuming features are encoded with strength proportional to discriminability, the number of errors in a feature-dependent recall task should be inversely proportional to discriminability. This was tested using a delayed match-to-sample task in which subjects were presented with four alternatives that each differed from the target in one of the dimensions described above. Because there was no correct response, subject errors were interpreted as a reflection of the relative strength of feature encoding. Across feature contrast levels, the most errors occurred for AOI followed by length, curvature and segment number. That the relative accuracy of curvature recall is higher than curvature discriminability suggests that while human observers may not be very good at discriminating differences of curvature, information about curvature is encoded more strongly in subject memory than other features of similar discriminability.