Purchase this article with an account.
Mary M. Conte, Sela Han, Jonathan D. Victor; Processing of image statistics with and without segmentation cues. Journal of Vision 2005;5(8):605. doi: 10.1167/5.8.605.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Statistical aspects of images are cues for texture discrimination and segmentation. In pre-segmented patches (VSS 2003), discrimination of local 1st order structure (luminance: LUM) and local 4th order structure (EO: even/odd isodipole textures) is much more efficient than that of non-local 2nd order structure (SYM: mirror symmetry), despite its visual saliency. This and other evidence suggests that symmetry detection uses a different computational substrate than processing of local statistical structure. Here we compare the relationship of these three statistical image classes to segmentation.
Stimuli consisted of four 8×8 arrays of black and white checks. In each trial, the target array deviated from that of the three distractor arrays in LUM, EO, or SYM. Arrays were positioned either 4 deg from fixation along the cardinal axes (fixed) or in “jittered” locations about the fixation point to introduce positional uncertainty. We used three display backgrounds, providing either a large segmentation cue (GRAY: uniform gray background), no cue (SAME: random checks of the same size as the targets) or an intermediate cue (HALF: random checks half the size of those in the targets). Practiced observers (N=5) were asked to identify the target in a 4-AFC task (stimulus duration: 100ms).
For LUM and EO, fraction correct without segmentation cues (LUM/SAME and EO/SAME) was 0.72 and 0.44, and increased when a segmentation cue was provided (LUM/GRAY, 0.98; EO/GRAY, 0.70). A modest segmentation cue (HALF) produced intermediate performance. Detection of SYM without a segmentation cue (SYM/SAME) was at chance, but increased to 0.46 for SYM/GRAY. There were no differences in fraction correct due to target position (fixed vs jittered) for any condition. Symmetry, while visually salient, does not support segmentation by itself. Moreover, segmentation effects are graded - even for image statistics that support segmentation, additional segmentation cues further enhance performance.
This PDF is available to Subscribers Only