Examination of the relation between affective content of images and gaze behavior

Saliency maps show the likelihood of each pixel on the image being looked at. They are often computed considering a neutral human subject, who is considered to be simply observing the image without any particular motivation (e.g, searching, registering etc.) or feelings (e.g, excited, fearful etc.). In this study, we focus on the emotional aspect and investigate whether there is any need to adjust these saliency maps to account for the emotions that they may induce in the viewers. To that end, we choose a set of images from an emotional image data set and display them to human subjects. We register their eye gaze and compute how well empirical gaze data matches the saliency maps. We quantify this based on two distribution-based (AUC, IG) and two location-based saliency metrics (CC and SIM). We see that although there are some parallels in some of these metrics, the evidence is not strong enough to claim a significant contribution due to affective content.