The visual identification of relational categories

An experiment was performed to investigate the ability of human observers to identify configural relations among three dots. Four stimulus categories were defined on the basis of whether or not the dots were arranged collinearly and whether or not the central dot was equally spaced relative to the two flanking dots. Observers were initially trained with feedback to identify these categories at a single orientation with a fixed uniform background, and then they were tested with variable orientations and backgrounds without feedback. The results revealed almost perfect generalization. We also simulated the same task using a recent feature hierarchy model (J. Mutch & D. G. Lowe, 2008) that is among the most successful for object recognition. This model performed well for fixed orientations and backgrounds, but it could not accurately identify these categories with variable orientations and backgrounds even when given training with those conditions. These findings suggest that feature hierarchy models represent the spatial relations within an image quite differently than human observers.