Saccadic eye movements shift the fovea between objects of interest to build a visual percept. In humans, saccades are predominantly executed along the cardinal axes, particularly in the horizontal direction. It is unknown how this horizontal saccade bias could arise mechanistically, though previous work suggests contributions from neural, image-based, and ocular motor factors. Here we used two publicly available eye movement datasets to first investigate which image features-spatial frequency, saliency, and structural content-relate to the horizontal saccade bias. Among the three image features, we found that orientation anisotropies in saliency content best predicted the strength of the horizontal saccade bias. Based on this result, we next implemented a saccade target selection model combining allocentric biases aligned with image orientation and egocentric biases aligned with eye or head orientation, independent of image content. As in prior work, this combination successfully replicated human saccade distributions during free viewing of upright images. When applied to tilted images, the model produced effects of image tilt and saccade size that were correlated with prior empirical findings, though with reduced amplitude, suggesting that current saliency models do not fully capture image effects. Taken together, these results suggest that saccade generation reflects both the allocentric biases present in the structure of natural scenes and the egocentric biases present in the saccade generation system itself. An open question is why the egocentric saccade bias exists, but our results suggest that it is adaptive in response to regularities in the world and our typical upright orientation.
Reeves et al. (Thu,) studied this question.