Annular Feature Pyramid Network for Salient Object Detection

Abstract

With the rapid development of deep learning, some attempts based on the fully convolutional networks have shown outstanding performance in salient object detection. Visual multi-context information is beneficial to detect salient regions so that how to better integrate multi-level convolutional features becomes essential. In this paper, we develop an annular feature pyramid network to augment information flow and enhance feature hierarchy. Our network contains a top-down path with lateral connections to help shallow layers locate salient regions and a bottom-up path with lateral connections to help deep layers retain fine object boundary. These feature maps at each resolution are combined to generate a final saliency prediction, which can take full advantage of high-level semantic features with low-level fine details. Comprehensive experiments demonstrate that our network performs favorably against state-of-the-art algorithms in term of different evaluation metrics.