Tobias Thomas^1,2 (), David Pietras¹, Constantin A. Rothkopf^1,2; ¹Technische Universität Darmstadt, ²Hessian Center for Artificial Intelligence (hessian.AI)

Humans move their eyes multiple times every second and behind every movement is a decision, where to direct gaze. Past research has identified factors influencing this selection, including the stimulus, the task, the scene context, but also oculomotor factors such as the ”inhibition of return” (IOR), which describes the effect that saccades to previously visited locations are subsequently inhibited. IOR has also been prominent in many computational models of attentional selection. However, while some studies have found spatial biases away from recent gaze targets consistent with IOR, other studies have found the bias only in some task conditions, and some studies instead even found the opposite bias. Here we conducted an eye tracking experiment using the free-viewing paradigm with 60 natural images, 20 of which were from a previous study (Bays & Husain, 2012), while the other 40 were chosen from 5 distinctly different image categories, including faces, landscapes and fractals. We have analyzed resulting eye movement data with respect to their (relative) amplitude, directions, but most importantly their change in direction. The results show significant differences with respect to the previously mentioned eye movement features between the different image categories. These differences are so pronounced, that using only the proportion of forward and return saccades as features for clustering is sufficient to assign most of the images back to the image category. To quantify this effect and add predictability for new images and categories, we related our empirical results to information theoretic measures of spatial properties of the images’ saliency maps. Taken together, our experiment reveals that IOR is not solely an oculomotor bias but instead actively depends on the stimulus itself as a key factor influencing the presence and strength of the IOR and show potential underlying causes using an analysis of the spatial distribution of image features.

Acknowledgements: The authors acknowledge the support of the cluster project "The Third Wave of AI" as part of the Excellence Program of the Hessian Ministry of Higher Education, Science, Research and Art