RGB-D Salient Object Detection

Salient object detection is derived from the research of human visual saliency, and has recently been widely applied to various computer vision applications due to its ability to automatically extract conspicuous objects from background. Thanks to the development of 3D sensors, RGB-D cameras can provide us with synchronized color and depth images, thus leading to the rise of RGB-D salient object detection. Compared with RGB saliency detection, effective features and priors for RGB-D methods still remain to be explored. Moreover, there is also an urgent need to find an effective way to fuse color and depth information, and introduce RGB methods to RGB-D applications.Considering the relationship between 3D features and visual saliency, we firstly propose two global priors, which are the normalized depth prior and the surface orientation prior, and they reveal the effect of object distance and perspective of observation respectively. A Markov random field based saliency restoration scheme is also presented to reconstruct and refine saliency map globally.In addition, we introduce boundary background prior to RGB-D detection task, and improve its performance by leveraging depth information and hypergraph model. To take the advantage of depth information to find the background, a novel depth-based background prior is also proposed. Last but not least, we design a robust fusion strategy to integrate color and depth information and implement a reliable detection of salient foreground.Our experiments on two publicly available RGB-D datasets demonstrate the effectiveness of our priors and optimization methods, and our approach consistently outperforms other state-of-the-art algorithms on both datasets.