Deeply supervised group recursive saliency prediction.

Fully convolutional neural networks (FCNs) have shown great effects in the saliency detection task. In this paper, we present an FCN based model to aggregate multi-level convolutional features for salient object detection. The main contribution of this work is to make full use of the information extracted by the network while preserving the characteristics of each level. We first propose a novel Intra-stage Feature Aggregating (IFA) module to enhance the detailed information by integrating feature maps from different layers within the same stage of the ResNet-50. We then design an Inter-stage Feature Fusion (IFF) module which exploits contextual information by combining the features from neighboring stages to make each feature map more discriminative. Finally, we propose a Deeply Supervised Group Recursive Prediction (DGR) module to refine the details of the predicted saliency maps and generate the final saliency map. Our approach performs favorably against state-of-the-art methods on five popular datasets and achieves a real-time speed of 30 fps. [ABSTRACT FROM AUTHOR]