Abstract
Mirror-symmetry is a salient visual feature that plays an important role in object recognition by providing the skeleton or canonical axis for the representation of shape. Here, we examine the extent to which segmentation, and changes to the local orientation, position and shape of the axis of symmetry affect symmetry detection and axis localisation. Stimuli were dot patterns containing different amounts of mirror-symmetry about the vertical axis. The symmetry axis was either continuous or broken into segments of a particular length. We used five segment lengths and varied their local orientation and position. There were three symmetry-axis distortion conditions: (a)‘position only’–the vertical axis segments were randomly jittered in a horizontal direction, (b)’orientation only’–positional aligned symmetry-axis segments were of random orientations, (c)‘sinusoidal-shaped axis’–both the positions and orientations of the symmetry-axis segments were modulated following a sinusoidal function. We varied the amount of symmetry by changing the proportion of symmetrical dots and measured both symmetry detection thresholds using a 2IFC procedure, and perceived location of symmetry-axis using a point-and-click task. We found that symmetry detection thresholds: (i)were higher for all segmented-axis compared to continuous-axis patterns, with the highest thresholds in the sinusoidal-shaped axis followed by position-only and orientation-only conditions; (ii)increased gradually with decreasing segment-length in the sinusoidal-shape axis and position-only conditions, but not in the orientation-only condition; (iii)increased with the amount of position and orientation jitter. Errors in the axis localisation task were affected by the amount of position and orientation jitter, but not by the segment length. While both the number of axis segments and their degree of positional alignment affect symmetry detection, the localisation of symmetry axis is primarily affected by the positional alignment. The results have implications for symmetry detection models based on early spatial filters relying on the alignment of oriented filters and AND-gate combinations of symmetric filters.
Acknowledgement: This research was supported by a Wellcome Trust Investigator grant (WT106969/Z/15/Z) given to EG