Learning image saliency from human touch behaviors

Shaomin Fang, Yijuan Lu, Xinmei Tian

Abstract: The concept of touch saliency was recently introduced to generate image saliency maps based on human simple zoom behavior on touch devices. However, when browsing images on touch screen, users tend to apply a variety of touch behaviors such as pinch zoom, tap, double tap zoom, and scroll. Do these different behaviors correspond to different human attentions? Which behaviors are highly correlated with human eye fixation? How to learn a good image saliency map from various/multiple human behaviors? In this work, we design and conduct a series of studies to address these open questions. We also propose a novel touch saliency learning approach to derive a good image saliency map from a variety of human touch behaviors by using machine learning algorithm. The experimental results demonstrate the validity of our study and the potential and effectiveness of the proposed approach.