Graph-based Clustering with Spatiotemporal Contour Energy for Video Salient Object Detection

In this paper, we propose a novel graph-based clustering model with spatiotemporal contour energy for video salient object detection, which can preserve the salient object and suppress the irrelevant surrounding background regions effectively. In order to estimate the salient object robustly in spatiotemporal domain, a novel spatiotemporal contour energy is modeled by exploiting optical flow, spatial contour and gradient flow field, which can enhance the energy inside the salient object and weaken it outside the salient object. Then, we estimate the saliency degree of superpixels by computing the geodesic distance of spatiotemporal contour energy between each superpixel node and border background nodes on a superpixel level graph model. The comprehensive experiments show that the proposed model outperforms the state of the art models, which are evaluated on two challenging datasets by three widely used performance metrics. We also applied the saliency map of the proposed method as a prior knowledge to unsupervised video object segmentation, showing that the proposed method can improve the segmentation performance of unsupervised video object segmentation.