Saliency Density Maximization for Efficient Visual Objects Discovery

摘要: Detection of salient objects in an image remains a challenging problem despite extensive studies visual saliency, as the generated saliency map is usually noisy and incomplete. In this paper, we propose new method to discover object without prior knowledge on its shape size. By searching sub-image, i.e., bounding box maximum density, formulation can automatically crop various sizes spite cluttered background, capable handle different types maps. A global optimal solution obtained by proposed density-based branch-and-bound search. The apply both images videos. Experimental results public dataset 5000 show that our unsupervised detection approach comparable state-of-the-art learning-based methods. Promising are also observed for videos with good potential video retargeting.