Part-aligned pose-guided recurrent network for action recognition

作者：

Highlights：

• The novel end-to-end architecture can improve the accuracy of action recognition efficiently.

• Introducing the part alignment into action recognition can capture spatio-temporal evolutions of actions.

• The part-based hierarchical pooling approach can learn a robust and discriminative feature.

• Our method obtains the state-of-the-art results on two important benchmarks of action recognition.

摘要

•The novel end-to-end architecture can improve the accuracy of action recognition efficiently.•Introducing the part alignment into action recognition can capture spatio-temporal evolutions of actions.•The part-based hierarchical pooling approach can learn a robust and discriminative feature.•Our method obtains the state-of-the-art results on two important benchmarks of action recognition.

论文关键词：Action recognition,Part alignment,Auto-transformer attention

论文评审过程：Received 8 December 2018, Revised 27 February 2019, Accepted 16 March 2019, Available online 20 March 2019, Version of Record 1 April 2019.

论文官网地址：https://doi.org/10.1016/j.patcog.2019.03.010