Search guided saliency

Abstract

We propose a new type of saliency as inspired by findings from visual search studies - the searching difficulty is correlated with the target-distractor contrast, the distractor homogeneity, as well as the target uniqueness. By putting an image pixel as the target and the surrounding pixels as distractors, a search guided saliency model is designed in accordance with these findings. In particular, three saliency measures in correspondence to the three searching factors are simultaneously computed and integrated by using a series of contextual histograms. The proposed model has been evaluated over three public datasets and experiments show superior prediction of the human fixations when compared to the state-of-the-art models.