Investigating the role of image meaning and prior knowledge in human eye movements control

Humans sample visual information by making eye movements towards different parts of their surroundings. Understanding what guides this sampling process is an important goal of vision science, and the present thesis is a contribution to this endeavour. Chapter One provides an overview of factors influencing human eye movements, which are typically divided into bottom-up (stimulus-dependent) and top-down (observer-dependent) processes. One of the challenges in studying these factors stem from the fact that they are often difficult to operationalize in a precise, unambiguous way. This is particularly problematic for semantic information contained in visual scenes (‘image meaning’), a top-down factor which is the backbone of the recently proposed framework for understanding human eye movements: the meaning maps approach. Chapter Two evaluates this approach and demonstrates that meaning maps – a crowd-sourced method designed to quantify the distribution of meaning in natural scenes – might be sensitive to complex visual features, rather than meaning. Chapter Three builds on that finding and shows that contextualized meaning maps, the most recent variant of the original meaning maps, share the limitations of their predecessors. Chapter Four adopts a novel perspective on eye-movement control and focuses on the interactions between image features (a bottom-up factor) and prior object-knowledge possessed by an observer (a top-down factor). Specifically, it shows that the same stimuli – black and white, Mooney-style two-tone images – are looked at differently depending on whether the observer possesses object-knowledge that enables them to bind images into coherent percepts of objects. The final chapter summarizes the thesis and maps the future directions for studies on eye movements. Taken together, findings reported here indicate that while top-down factors such as prior object-knowledge play a crucial role in guiding human gaze, the tools to study them offered by the meaning maps approach still need to be improve