RGB-D image saliency detection from 3D perspective

With the advent of stereo camera saliency object detection for RGB-D image is attracting more and more interest. Most existing algorithms treat RGB-D image as one RGB image and one depth map, then measure saliency map independently, and last fuse them. They disregard the fact that human visual system operates in real 3D environments. The paper proposed saliency object detection for RGB-D image from 3D perspective. It regards object as three dimensional structures, and redefines boundary conception in RGB-D image, and regards space boundary including top, down, left, right, front, back plane in real 3D environment as background. It incorporates 3D compactness feature, in which salient objects typically have 3D compact spatial distributions, into color and depth feature to express similarity among supervoxels and applies manifold ranking by six boundary planes to generate six saliency maps, and then integrates them to get the RGB-D saliency map from background view. In the end it refines saliency map by high confident salient seeds from foreground view. Experiment results show that six planes of RGB-D image are superior to four sides of RGB image as background seeds and 3D compactness plays an important role in saliency measurement. Our approach outperforms other state-of-the-art algorithms on NLPR RGBD 1000 benchmark.