Abstract
The distribution of fixations in natural scenes is difficult to predict. In free viewing, the proportion of variance explained by bottom-up scene effects is at best about 60%; in visual search, somewhat less. The remaining variance is usually attributed to the uncertain effects of search strategy and related cognitive factors. The aim here was to assess the variability of these top-down effects by comparing the distributions of first, second, and subsequent fixations in a difficult search task. Seven observers with normal color vision and visual acuity viewed repeated images of 20 natural scenes, each subtending 17x13 deg visual angle, displayed on a color monitor. The target was a shaded gray sphere subtending 0.25 deg, matched in mean luminance to its local surround and positioned randomly in the scene. In each trial an image appeared for 1 sec, after which the observer reported whether the target was present. Each observer performed 260 trials per scene, half containing the target and half not. Gaze position was monitored with an infra-red video eye-tracker sampling at 250 Hz. The spatial distribution of fixations was estimated by a locally weighted quadratic regression (loess) with a Gaussian kernel of standard deviation 2.5 deg, although the exact value was unimportant. As display duration was limited, only the first four fixations were analyzed. Within scenes, the first and subsequent fixations, pooled over observers, were found to be only moderately correlated with reported detections (R2 = 55%) but strongly correlated with each other (R2 = 83%), suggesting that observers maintained a constant search strategy for a given scene. This constancy was not an artifact of a central viewing bias, since across scenes fixations were relatively weakly correlated (R2 = 34%). Although top-down effects in visual search may be uncertain, they appear to be less variable than expected.
Meeting abstract presented at VSS 2013