Abstract
Depth is one of the primary visual cues which distinguish an object from its background. In recent years, salient object detection has achieved great success with the help of a convolution neural network and its corresponding depth map. Previous methods have already utilized depth map to improve the precision of the results; however, all of the previous methods are only concentrating on the available RGB-D datasets to train their network. In this paper, we used a depth estimation network to find the depth map of 2D images. That depth map has been used to train the depth-guided saliency network, which produces the intermediate depth saliency map. Finally, the depth saliency map has been fused with the coarse saliency map to obtain the final saliency map. Experiments demonstrate the effectiveness of the proposed method, which achieves state-of-the-art performance on six popular benchmarks.
