Eigenobject-wise saliency detection based on manifold ranking

Saliency detection that utilizes graph model has achieved considerable progress during the past years. However, few methods consider object cues. We propose a novel manifold ranking based graph model that estimates the saliency of the image elements via their relevances to object seeds. An “eigenimage” selection algorithm dependent on the solved eigenvectors of the normalized Laplacian matrix is proposed to generate the object-wise seeds. Meanwhile, we propose a foreground border blanking approach to settle the failure of boundary prior saliency when object regions touching the border. Extensive experiments on benchmark datasets indicate that our algorithm could further improve the performance of representative graph-based saliency detection methods.