RGB-D Saliency Detection: Dataset and Algorithm for Robot Vision

Abstract

Saliency detection is an active research field in computer vision in recent years. As RGB-D sensors are more and more widely used in a robot system, the demand of corresponding saliency detection datasets and algorithms are growing rapidly. In this paper, we built a RGB-D saliency detection dataset NJUSTDS1000 contains 1000 real RGB-D scenes. The labeling of saliency ground truth of this dataset is based on color and depth fixation map. Then we propose a spectral and spatial analysis based RGB-D saliency detection model. It uses quaternion to present multi-channel features and dose fast saliency detection based on spectral analysis. Then a two-steps scale adaptive saliency fusion process is carried out in spatial domain, which are scale adaptive superpixel based saliency smoothing and multi-layer cellular automata based saliency maps fusion. We validate the proposed model on NJUSTDS 1000 and MSRA10K.