RGB-D Saliency Object Detection Based on Adaptive Manifolds Filtering

Abstract

The extensive works on saliency object detection for decades concentrated on RGB images. Since depth cameras are widely used recently and the obtained images provide users with a higher viewing experience, it is important to specially study RGB-D images. Inspired by good performance of adaptive manifolds filtering, we propose a new RGB-D saliency object detection method to preserve fine details in the saliency maps, combining both the depth characteristics and local edge-preserve ability of filters. The proposed method contains two stages. Firstly, color enhancement map and depth smooth map are calculated respectively, whose goal is to enhance the color contrast and reduce the influence of defected depth data. On this basis, depth enhanced approach is adopted to improve the saliency detection ability. The experiments demonstrate that proposed method achieves good visual quality and structure measure scores.