Abstract:

Saliency detection is an important part of computer vision and many creative studies have been proposed. However, most existing methods focus on detecting salient objects in 2D images and they cannot be used to detect salient objects in RGBD images. In the paper, a new RGBD saliency detection method based on multiple perspective fusion is proposed. The method considers multiple complementary relationships including color and depth, global and local as well as foreground and background to detect salient objects. Firstly, it combines color feature and depth feature to construct the graph model. Secondly, it computes the global and local compactness feature based on the graph model to get the initial saliency map. Thirdly, it uses the boundary connectivity based on color and depth to calculate the background probability as the weight on initial saliency map and then produces the background optimization saliency map. Finally, it extracts the foreground inquiry node from background optimization saliency map and then uses the manifold ranking algorithm to get the final saliency map. Experiments show that the fusion policy of multiple perspectives can effectively improve the accuracy of saliency detection and outperforms other state-of-the-art algorithms in RGBD1000 benchmark.

Key words: saliency detection, multiple perspectives fusion, compactness, boundary connection weight, manifold ranking