Bayesian Modeling of Visual Attention

The mechanism in the brain that determines which part of the multitude of sensory data is currently of most interest is called selective attention. There are two kinds of attention cues, stimulus-driven bottom-up cues and goal-driven top-down cues determined by cognitive phenomena like knowledge, expectations, reward, and current goals. In this paper, we propose a Bayesian approach that explains the optimal integration of top-down cues and bottom-up cues. The top down cues include appearance feature, contexts, and locations of a target. The bottom up attention (saliency) is defined as the joint probability of the local feature and context at a location in the scene. The feature and context is organized in a pyramid structure. In this way, multiscale saliency is easily implemented. We demonstrate that the proposed visual saliency effectively predicts human gaze in free-viewing of natural scenes.