| Summary: | Given a group of images, co-saliency detection aims at highlighting the common and salient foreground regions. To optimally explore the complementary information among images, we propose an effective propagation mechanism for RGBD images. First, we design a depth optimization map guided by image global saliency, which generates a superpixel-level saliency propagation label to express the primary saliency propagation confidence. Then, we further get the superior saliency propagation confidence based on the corresponding probabilities of the external contrast between images and the image internal regions. Finally, the primary and superior saliency propagation confidence are integrated to optimize the saliency propagation and get the final co-saliency value. The proposed method enables the complementary information among images to be reasonably propagated in a group of images. The relevance of depth information is enhanced and the co-salient objects are closer to the truth values. Experiments on two RGBD co-saliency datasets demonstrate the effectiveness of the proposed model.
|
|---|