Fully convolutional neural networks (FCNs) have shown outstanding performancein many dense labeling problems. One key pillar of these successes is miningrelevant information from features in convolutional layers. However, how tobetter aggregate multi-level convolutional feature maps for salient objectdetection is underexplored. In this work, we present Amulet, a genericaggregating multi-level convolutional feature framework for salient objectdetection. Our framework first integrates multi-level feature maps intomultiple resolutions, which simultaneously incorporate coarse semantics andfine details. Then it adaptively learns to combine these feature maps at eachresolution and predict saliency maps with the combined features. Finally, thepredicted results are efficiently fused to generate the final saliency map. Inaddition, to achieve accurate boundary inference and semantic enhancement,edge-aware feature maps in low-level layers and the predicted results of lowresolution features are recursively embedded into the learning framework. Byaggregating multi-level convolutional features in this efficient and flexiblemanner, the proposed saliency model provides accurate salient object labeling.Comprehensive experiments demonstrate that our method performs favorablyagainst state-of-the art approaches in terms of near all compared evaluationmetrics.
展开▼
