Real-time salient object detection with boundary information guidance

Abstract

The low contrast between salient objects and background may result in errors in salient object detection. The low-contrast areas are likely distinguished away from salient objects by saliency information. However, salient objects near the boundary are often detected wrongly based on current methods. In this paper, a novel network with boundary information guidance is proposed to distinguish low-contrast areas near the boundary effectively. We design two separate decoding sub-networks including a saliency sub-network and a boundary sub-network to learn saliency and boundary information respectively based on multi-level feature fusion. We further design a connection module to exchange the information between two sub-networks and a fusion module to fuse information of salient objects and boundaries. The boundary sub-network supervised by boundary maps outputs the error map calculated by ground truth to focus on boundary information. The boundary sub-network is only used in training and the optimal weight of the total loss is obtained by experiments to balance the two sub-network losses. With the guidance of boundary information, salient objects are detected accurately. Extensive evaluations on six benchmark and two video datasets demonstrate that our model outperforms the state-of-the-art methods. Furthermore, the proposed model can run at more than 34 FPS.