2PESNet: Towards online processing of temporal action localization

作者：

Highlights：

• The new task which aims to detect instance-level actions instantly from the streaming videos.

• 2-pass action instance detection method that can be applied in the online setting.

• Two-stage action end detection module to gain the accurate action end point with low delay of response.

• Extensive experiments on THUMOS’14 and ActivityNet v1.3 that show effectiveness and robustness of the model.

摘要

•The new task which aims to detect instance-level actions instantly from the streaming videos.•2-pass action instance detection method that can be applied in the online setting.•Two-stage action end detection module to gain the accurate action end point with low delay of response.•Extensive experiments on THUMOS’14 and ActivityNet v1.3 that show effectiveness and robustness of the model.

论文关键词：Online video understanding,Temporal action localization

论文评审过程：Received 24 June 2021, Revised 17 February 2022, Accepted 21 June 2022, Available online 23 June 2022, Version of Record 28 June 2022.

论文官网地址：https://doi.org/10.1016/j.patcog.2022.108871