Predictive video saliency detection

Abstract

Human Visual System has the characters of focusing on salient regions when seeing images or videos. The method which finds the irregularity and unpredictability of images or videos by simulating human visual system is called saliency detection. In this paper, we propose a novel video saliency detection method based on temporal consistency. The traditional video detection approaches fall into two main groups. One processes a video frame by frame independently without considering motion information, and the other regards optical flow only as a part of features without taking account of consistency of video saliency between consecutive frames. In the proposed method, the temporal consistency constraint is enforced by using motion vectors. By constructing correspondences via motion information, the saliency map of each frame can be predicted by the result of its previous frame. By combining the predicted results and the traditional approaches, our algorithm can achieve better video saliency maps. © 2012 Springer-Verlag.