Fixation Prediction based on Scene Contours

Abstract

Previous works suggest that scene contours play important roles in guiding visual attention. In this study, a computational model is proposed to improve the performance in visual saliency prediction by integrating the low- and mid-level visual cues and evaluate the contribution of scene contours in guiding visual attention. Firstly, we define three kinds of Gestalt principles based on mid-level cues, including contour density, closure, and symmetry to characterize the potential salient regions. In addition, we employ the classical bottom-up methods to generate low-level saliency maps. Finally, the proposed method combines the low-level cues from natural images and the mid-level cues from the corresponding contours to improve the fixation prediction. Experimental results show that the contour-based midlevel cues can remarkably improve the performance of the bottomup models in fixation prediction.