Multi-stream pose convolutional neural networks for human interaction recognition in images
作者：
Highlights：
• We explore the contribution of poses to human–human interaction recognition in images.
• Several pose-based representations are proposed.
• Multiple pose based CNN streams are utilized, where various fusion architectures for pose-based representations are explored.
• An extended benchmark dataset for human–human interaction recognition in images is collected.
• Experiments show that paying more attention on poses has a positive affect on recognizing interactions.
摘要
•We explore the contribution of poses to human–human interaction recognition in images.•Several pose-based representations are proposed.•Multiple pose based CNN streams are utilized, where various fusion architectures for pose-based representations are explored.•An extended benchmark dataset for human–human interaction recognition in images is collected.•Experiments show that paying more attention on poses has a positive affect on recognizing interactions.
论文关键词：Human–human interactions,Convolutional neural networks,Poses
论文评审过程：Received 13 October 2020, Revised 23 March 2021, Accepted 7 April 2021, Available online 18 April 2021, Version of Record 19 April 2021.
论文官网地址：https://doi.org/10.1016/j.image.2021.116265
