Co-saliency detection via sparse reconstruction and co-salient object discovery

Abstract

Co-saliency detection aims at discovering common and salient objects in a group of related images, which is useful to variety of visual tasks. We propose a novel co-saliency detection framework via sparse reconstruction and co-salient object discovery. By taking advantage of the common background in-formation, we first reconstruct images with the common background bases and computer sparse reconstruction error. Second, we discover the common salient objects using high-level and low-level features. Then the reconstruction errors are refined using co-salient object information to get the superpixel-level co-saliency. Third, pixel-level saliency is computed by an integration of multi-scale superpixel-level co-saliency maps, with the help of intra-saliency propagation and Gaussian refinement. The quantitative and subjective experimental results on two benchmark datasets show that our method outperforms both the state-of-art saliency detection methods and co-saliency detection methods.