Knowledge-Driven Compositional Action Recognition

Yang Liu, Fang Liu, Licheng Jiao, Qianyue Bao, Shuo Li, Lingling Li, Xu Liu

Abstract: Highlights•We utilize grammar and commonsense knowledge for compositional action recognition.•Our method builds unseen “action-objects” compositions to improve generalization.•We inject structured commonsense into label to verify the rationality.•Our method achieves excellent results in both compositional and few-shot settings.