Salient Seeds Selection Based On Multi-feature Ranking Of Object Proposals

Generic object detection has been enjoying a great popularity and making significant progress recently. Though the detection rate is continually being updated by the emerging methods, it is difficult to find the salient object accurately and completely among a large number of plausible object proposals. In this paper, we propose a salient seeds selection algorithm based on the object proposals to solve this problem. First, a set of candidate proposals are generated through an existing technique; second, two types of features respectively corresponding to region and superpixel levels are extracted and concatenated to form a high-dimensional feature vector of each object proposal; the feature vector is then input into a learned Ranking SVM model and a prediction score is output; the final seed map is merged by a weighted sum of the top K candidate proposals. Experimental results on publicly available and challenging datasets demonstrate that the seed map itself could achieve favorable performance against the state-of-the-art methods. Propagation methods can be used on the generated seed map to reach a promoted saliency map.