Inferring task-relevant image regions from gaze data

A number of studies have recently used eye movements of a user inspecting the content as implicit relevance feedback for proactive retrieval systems. Typically binary feedback for images or text paragraphs is inferred from the gaze pattern. We seek to make such feedback richer for image retrieval, by inferring which parts of the image the user found relevant. For this purpose, we present a novel Bayesian mixture model for inferring possible target regions directly from gaze data alone, and show how the relevance of those regions can then be inferred using a simple classifier that is independent of the content or the task.