A Biologically-Inspired Visual Saliency Model to Test Different Strategies of Saccade Programming

Saliency models provide a saliency map that is a topographically arranged map to represent the saliency of the visual scene. Saliency map is used to sequentially select particular locations of the scene to predict a subject's eye scan-path when viewing the corresponding scene. A saliency map is most of the time computed using the same point of view or foveated point. Few models were interested in saccade programming strategies. In visual search tasks, studies shown that people can plan from one foveated point the next two saccades (and so, the next two fixations): this is called concurrent saccade programming. In this paper, we tested if such strategy occurs during natural scene free viewing. We tested different saccade programming strategies depending on the number of programmed saccades. The results showed that the strategy of programming one saccade at a time from the foveated point best matches the experimental data from free viewing of natural images. Because saccade programming models depend on the foveated point, we took into account the spatially variant retinal resolution. We showed that the predicted eye fixations were more effective when this retinal resolution was combined with the saccade programming strategies.