Deep gated attention networks for large-scale street-level scene segmentation
作者：
Highlights：
• A novel spatial gated attention mechanism is proposed in the context of pixel-wise labeling tasks.
• A multi-scale feature interaction mechanism is proposed for hierarchical feature aggregation.
• Different levels of features are re-weighted with the local structure and contextual information.
• The proposed GANet can better capture the scene layout and multi-level information of street-view images.
• State-of-the-art performance on three challenging large-scale scene segmentation benchmarks is achieved.
摘要
•A novel spatial gated attention mechanism is proposed in the context of pixel-wise labeling tasks.•A multi-scale feature interaction mechanism is proposed for hierarchical feature aggregation.•Different levels of features are re-weighted with the local structure and contextual information.•The proposed GANet can better capture the scene layout and multi-level information of street-view images.•State-of-the-art performance on three challenging large-scale scene segmentation benchmarks is achieved.
论文关键词：Scene segmentation,Fully convolutional network,Spatial gated attention,Street-level image understanding
论文评审过程：Received 19 July 2018, Revised 10 December 2018, Accepted 16 December 2018, Available online 17 December 2018, Version of Record 24 December 2018.
论文官网地址：https://doi.org/10.1016/j.patcog.2018.12.021
