Abstract
Introduction: Eye Movements during scene viewing are not just directed to areas with low-level saliency but to regions that are task-relevant (Hayhoe, 2005; Koehler et al., 2015) and/or semantically meaningful (Henderson et al., 2017). A recent study (Henderson et al., 2017 ) defined semantic informativeness by measuring subjective human ratings of the meaningfulness of different regions of scenes. We propose a novel method to measure semantic content of a scene region based on its influence on the scene description. We demonstrate how small visual changes of images that greatly change the scene description can result in large changes in fixational eye movements. Methods: Twenty pairs of images were photographed with small changes in object positions or actor/s posture that altered the scene description while minimally changing the visual content. Twenty four participants were split to provide descriptions for only one set from the pair of images. Presentation time was 3 sec and eye position was measured using an Eyelink 1000. Heat maps for the fixations were generated for each observer and image. Results: Correlation of fixational heat maps across observers for the same images averaged 0.76 +- 0.05. In contrast, the correlation of fixational heat maps of the observers across image pairs (an image and its visually similar but semantically manipulated pair) was significantly lower r = 0.3 +- 0.15 (p<0.001). Conclusion: Our findings show the extent to which semantic scene properties can influence eye movements even with minimal visual changes to the images.