Heading recovery from optic flow: Comparing performance of humans and computational models

Recovery of heading from optic flow (OF) has been studied extensively by experimental assessment of human performance and by building computational models capable of heading recovery. However, relatively little work has made direct comparisons between models or between models and human performance. Here, we undertake such comparisons investigating heading recovery when OF density and dot direction noise are manipulated. Participants and a range of computational models viewed radial, limited lifetime dot OF fields and made 2AFC judgements about whether heading was to the left or right of a target in the scene. There were four possible horizontal target locations (±2, ±4 deg from centre of display) and 10 possible horizontal focus of expansion offsets (±0.2, ±0.5, ±1, ±2, ±4 deg relative to target). Dot motion orientation was corrupted by additive, zero mean Gaussian noise with standard deviation at one of three levels (0, 7.5, 15 deg). Dot density was varied by changing the number of dots in the field (5, 50, 100, 200). Thresholds for human observers dropped most sharply (by 50–75%) as number of dots increased from 5 to 50 but then performance stabilised. Furthermore, human observers showed some robustness to noise; when there were at least 50 dots in the display performance in the no noise and 7.5 deg noise conditions was similar but was degraded slightly in the 15 deg noise condition. Performance for the models tested varied greatly. Of these models, Longuet-Higgins & Prazdny (PRSL:Series B; 208(1173); 1980) model performed particularly poorly over the dot density range and showed little robustness to noise. In contrast, the Perrone (JOSAmA; 9(2); 1992) model was considerably more robust to noise and showed a qualitatively similar pattern of dependence on dot density to that seen in humans.