Abstract
Parkhurst and Niebur (2003, Vision Science Society) collected data from over 1000 internet subjects, who were asked to indicate the five most interesting points in a variety of natural and artificial scenes. They found a high degree of consistency in the selected locations across subjects, particularly for early selections. Moreover, they showed that these interest points clustered around areas determined to be high in visual salience by a purely stimulus-driven computational model of attention.
We recorded eye movements from a new group of 21 laboratory subjects as they free viewed the same set of images for five seconds. The first five fixations for each subject were significantly more likely to land on locations rated as interesting by the internet subjects compared to chance levels. Consistent with previous findings, fixated areas were also higher in visual salience. In addition, we ran a number of cross correlations between internet subjects' combined interest maps, comprised of all points in each image they indicated as being interesting, and our laboratory subjects' fixation maps, comprised of all points in each image they fixated. We found that 93 out of 100 images had a higher cross correlation than would be expected by chance. Additionally, 40 out of 100 images had a higher cross correlation between their saliency and fixation maps than chance.
These results extend previous findings that showed interesting regions of images (distinct objects) are also salient (e.g., Elazary & Itti, Vision Science Society, 2006). Furthermore, subjectively determined interesting regions appear to be as strong a predictor of human eye fixations as computational models of attention, if not stronger. This may be due to the fact that these interest points reflect bottom-up factors, such as visual salience, as well as top-down factors, such as scene semantics, that influence eye movements.