Ataol Burak Özsu¹, Marcel Linka^1,2, Diana Kollenda^1,2, Benjamin de Haas^1,2; ¹Experimental Psychology, Justus Liebig University, Giessen, Germany, ²Center for Mind, Brain and Behavior (CMBB), Marburg and Giessen, Germany

Computational models of gaze in natural scenes evolved to consider ever more intricate features, from local contrast (Itti et al., 1998) to local meaning (Henderson & Hayes, 2017). However, even state-of-the-art models (Kümmerer et al., 2022) are blind to context and interrelations. In contrast, humans can extract global meaning from a single glance (Fei-Fei et al., 2007) and are sensitive to context effects (Murlidaran & Eckstein, 2024). Here, we ask whether human scan paths trace meaningful relationships between elements of a scene. We re-analyzed a large-scale dataset (n > 6,500; Linka et al., 2024) of children and adults free-viewing 40 complex scenes (Linka & de Haas, 2020). For each observer and scene, we simulated scan paths by randomly shuffling the order of empirical fixations, with or without spatial biases (Clarke et al., 2017). This simulation informed us about the frequency of transitions between scene elements to be expected under a (spatially biased) random walk. Adult observers were much more likely than this model to make direct saccades between certain pairs of scene elements. To understand the nature of these saccades, we asked a separate set of participants to describe the same set of scenes (Kollenda et al., 2024). Crucially, direct saccadic transitions were biased to trace relationships picked up in the descriptions. For instance, observers showed a strong tendency to make direct saccades from agents to patients of transitive actions featured in the descriptions. Finally, children’s gaze was closer to the random walk model and systematic deviations slowly increased up until early adulthood. We conclude that humans learn to trace meaningful relationships with their gaze. This likely reflects and aids the understanding of global meaning (Murlidaran & Eckstein, 2024) and formation of event models (Loschky et al., 2020, 2024), which are informed by experience-dependent priors.

Acknowledgements: This research was supported by ERC Starting Grant INDIVISUAL to BdH and an Erasmus+ Traineeship to ABÖ