科研之友
微信
新浪微博
Facebook
分享链接
摘要
As an important topic which is emerging in the field of computer vision, co-saliency detection is aimed at finding the salient target among several related images. The existent methods are usually to formulate co-saliency map by designed clues or initialized and direct forward pipeline. However, these models are in lack of an improved scheme for loop scheme. At the meantime, lots of methods only focus on RGB images while ignore depth cues of RGBD images. In this paper, an iterative RGBD co-saliency method will be introduced. It utilizes the existing single saliency maps as initialization and another kind of refinement-cycle model to generate a final RGBD co-saliency map. The proposed model adopted three schemes, including addition scheme, deletion scheme and iterative scheme. Besides, this paper also proposed a new descriptor in addition scheme: Depth Shape Prior (DSP). Putting introduced depth cues into DSP can enhance recognition ability for co-saliency target, which will eventually achieve the transformation from two-dimension saliency detection to co-saliency detection based on RGBD images. Experiments show the effectiveness of described in this paper.
