Detecting occlusion boundaries via saliency network

In this paper, we address the problem of detecting occlusion boundaries from video sequences. We build a bi-directed graph whose nodes are line fragments extracted from superpixels´s edges. Based on the graph, we compute a global occlusion saliency map by integrating motion, shape and topology cues into the framework of Saliency Network. Furthermore, with the structural information generated from the network, the property of structural consistency is proposed to prune the graph and refine the saliency map. Finally, we train a classifier to detect occlusion fragments combining the global saliency value and local edge strength. The detector outperforms the state-of-the-art on the benchmark of Stein and Hebert[8] by improving average precision to .80.