Video saliency detection based on robust seeds generation and spatio-temporal propagation

This paper proposes a novel video saliency detection method for unconstrained videos with various motion patterns and complex scenes. We fuse multiple tempo-scale optical flow with discarding rule to enhance the reliability of motion information. Based on efficiently computation of motion distinction, our algorithm is able to locate the foreground and background approximately. Considering the mutuality of video frames, we regard video saliency seeds generation as the pattern mining process. With the help of robust saliency seeds, spatio-temporal propagation is performed in both intra-frame and inter-frame graphs. This provides an effective way to refine saliency maps. Quantitative and qualitative experiments are carried out on two benchmark video datasets, which show that our approach achieves state-of-the-art performance in video saliency detection.