Can Computers Learn From Humans to See Better? Inferring Scene Semantics From Viewers' Eye Movements

摘要

This paper describes an attempt to bridge the semantic gap between computer vision and scene understanding employing eye movements. Even as computer vision algorithms can efficiently detect scene objects, discovering semantic relationships between these objects is as essential for scene understanding. Humans understand complex scenes by rapidly moving their eyes (saccades) to selectively focus on salient entities (fixations). For 110 social scenes, we compared verbal descriptions provided by observers against eye movements recorded during a free-viewing task. Data analysis confirms (ⅰ) a strong correlation between task-explicit linguistic descriptions and task-implicit eye movements, both of which are influenced by underlying scene semantics and (ⅱ) the ability of eye movements in the form of fixations and saccades to indicate salient entities and entity relationships mentioned in scene descriptions. We demonstrate how eye movements are useful for inferring the meaning of social (everyday scenes depicting human activities) and affective (emotion-evoking content like expressive faces, nudes) scenes. While saliency has always been studied through the prism of fixations, we show that saccades are particularly useful for (ⅰ) distinguishing mild and high-intensity facial expressions and (ⅱ) discovering interactive actions between scene entities.