Top-down and bottom-up attentional multiple instance learning for still image action recognition
作者：
Highlights：
• An end-to-end approach that discovers action-related image parts.
• Creates pixelimage-level attention masks.
• Image-level fine-grained action maps are used to localize the action related cues.
• Extensive evaluations on four different still image action recognition datasets.
• Spatial and top-down attention can help identify action-related regions and pixels.
摘要
•An end-to-end approach that discovers action-related image parts.•Creates pixelimage-level attention masks.•Image-level fine-grained action maps are used to localize the action related cues.•Extensive evaluations on four different still image action recognition datasets.•Spatial and top-down attention can help identify action-related regions and pixels.
论文关键词：Still image action recognition,Multiple instance learning,Attention
论文评审过程：Received 11 November 2020, Revised 25 January 2022, Accepted 11 February 2022, Available online 22 February 2022, Version of Record 17 March 2022.
论文官网地址：https://doi.org/10.1016/j.image.2022.116664
