Knowledge memorization and generation for action recognition in still images
作者：
Highlights：
• We propose to transfer the knowledge learned from videos to images to improve the performance of single image based human action recognition.
• To guide the generation process and generate discriminative information, we introduce a reconstruction loss regularized with classification loss.
• We prove that the knowledge learned from both color and flow sequences can transfer to single images to improve the performance of human action recognition and the knowledge from flow sequences benefits the improvement significantly.
摘要
•We propose to transfer the knowledge learned from videos to images to improve the performance of single image based human action recognition.•To guide the generation process and generate discriminative information, we introduce a reconstruction loss regularized with classification loss.•We prove that the knowledge learned from both color and flow sequences can transfer to single images to improve the performance of human action recognition and the knowledge from flow sequences benefits the improvement significantly.
论文关键词：Action recognition,Deep learning,Knowledge-transfer
论文评审过程：Received 26 January 2020, Revised 10 February 2021, Accepted 14 March 2021, Available online 20 July 2021, Version of Record 25 July 2021.
论文官网地址：https://doi.org/10.1016/j.patcog.2021.108188
