Temporally smooth online action detection using cycle-consistent future anticipation

作者：

Highlights：

• We propose a novel network that anticipates future and applies temporal smoothness for online action detection.

• The future anticipation is trained by an unsupervised method with a novel cycle-consistency loss function.

• Our proposed temporal smoothing architecture can be applied in the online setting.

• Our framework achieves the state-of-the-art performance among the recent methods of online action detection.

摘要

•We propose a novel network that anticipates future and applies temporal smoothness for online action detection.•The future anticipation is trained by an unsupervised method with a novel cycle-consistency loss function.•Our proposed temporal smoothing architecture can be applied in the online setting.•Our framework achieves the state-of-the-art performance among the recent methods of online action detection.

论文关键词：Online action detection,Cycle-consistency,Temporal smoothing,Video understanding

论文评审过程：Received 1 April 2020, Revised 6 October 2020, Accepted 18 March 2021, Available online 24 March 2021, Version of Record 5 April 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.107954