Natural Scene Categorization from Conjunctions of Ecological Global Properties

Human scene understanding is remarkable: with only a brief glance at an image, an abundance of information is available spatial layout, scene function, semantic label, etc. Here we propose a scene-centered model of rapid human scene understanding that uses a vocabulary of global, ecological scene properties that combine to categorize natural landscape images. Behaviorally, we show human observers are sensitive to the underlying distributions of these global properties for use in basic-level categorization. An ideal observer trained only on the distributions of these properties predicts human scene categorization performance (r=0.90) and human errors.