Abstract
Previous studies on object recognition have drawn different conclusions regarding the importance of specific object features, such as convexities, concavities and intermediate points. Some studies found evidence for a predominant role of convexities, whereas others favored concavities or intermediate parts. However, most of studies have employed familiar objects or simple geometric shapes not necessarily containing curves (polygons) as their stimuli. Here we present a novel set of shapes with well-defined convexities, concavities and points between convexities and concavities. The shapes were composed of the sum of three different radial frequency (RF) components with random phases, segmented to remove all but variable lengths of contour centred on the feature of interest. Observers were required to accurately match the segmented test shape to one of two subsequently presented non-segmented and re-scaled match shapes. Results show that for very short (dot-sized) segment lengths, performance was significantly higher for convexities than for either concavities or intermediate points. Performance for convexities remained constant as a function of segment length, and although performance improved with segment length for concavities and intermediate points, it only reached convexity performance at the largest lengths tested. No significant differences between concavities and intermediates were found. These results suggest that for this class of closed shapes, shape is encoded from the positions of convexities, rather than from the positions of either concavities or intermediates. This is in line with current neurophysiological results proposing a sparse coding scheme based on regions of maximum contour curvature.