Saliency Detection by Unifying Regression and Propagation

This paper introduces a novel saliency detection method by incorporating logistic regression into the label propagation framework, along with a principled weight computation for saliency fusion. First, the initial map is generated by computing objectness and backgroundness. Second, we unify logistic regression and label propagation to predict saliency labels. Last, we fuse the predicted result and initial map and further refine the fused map across multiple scales. Moreover, we use clustering random forest to learn the pairwise affinities between superpixels for backgroundness computation and saliency prediction. Extensive experiments on three large benchmark datasets demonstrate the proposed algorithm performs well against the state-of-the-art methods.