Salient object detection via proposal selection

The scoring mechanisms in existing proposal algorithms do not work well and are usually unable to accurately evaluate the proposals generated by other algorithms because they come from the respective generation processes of proposals. In this paper, we re-examine the characteristics of proposals and present an unsupervised saliency detection method via proposal selection. First, we re-define the evaluation indicators of objectness, based on which some good region proposals are coarsely selected. We employ the top-scoring ones to produce an initial saliency result. Second, we self-train a structural ranker across a group of images to rank the proposals and obtain the proposal-level saliency map for each image. Different from traditional rankers, which balance the accuracy of the full list, this ranker prefers the high-quality proposals to be ranked at the top regardless of the rest. After that, we refine the saliency result by combining the finer processing based on superpixels. Experimental results on four benchmark datasets demonstrate that the proposed method performs favorably against the state-of-the-art methods.