Local Saliency Estimation and Global Homogeneity Refinement for Video Saliency Detection

Abstract

Saliency detection aims to segment the object of interest from the rest of the scene. While there has been a big number of saliency detection methods in still images, video saliency is in its early stages. In this paper, we propose a two stages video saliency detection method using local saliency estimation and global homogeneity refinement. Starting from a patch, the problem of saliency detection is modeled as a growing region which starts from a patch with high saliency information to the background. Local saliency is measured by combining spatial priors presented by local surrounding contrast with temporal information issued from the motion estimation feature. Temporal and spatial information are fused and then used to label each patch as foreground and background patches and produce the final saliency maps. Finally, Global homogeneity refinement is used to refine the saliency results by evaluating the foreground and background probabilities ratio propagated from the patches. Experiments have proved that the proposed method outperforms state-of-the-art methods over two benchmark datasets.