Video action detection by learning graph-based spatio-temporal interactions

作者：

Highlights：

• Video action detection is addressed using spatio-temporal graphs.

• A single graph can handle spatial and temporal relationships.

• Improvements over robust backbones and state-of-the-art results are presented.

• Improvements are obtained without backbone finetuning, learning only interactions.

摘要

•Video action detection is addressed using spatio-temporal graphs.•A single graph can handle spatial and temporal relationships.•Improvements over robust backbones and state-of-the-art results are presented.•Improvements are obtained without backbone finetuning, learning only interactions.

论文关键词：Video understanding,Action detection,Graph learning

论文评审过程：Received 29 June 2020, Revised 1 February 2021, Accepted 24 February 2021, Available online 27 February 2021, Version of Record 5 March 2021.

论文官网地址：https://doi.org/10.1016/j.cviu.2021.103187