Joint Learning with Group Relation and Individual Action

Chihiro Nakatani, Hiroaki Kawashima, Norimichi Ukita

Abstract: This paper proposes a method for group relation learning. Different from related work in which the manual annotation of group activities is required for supervised learning, we propose group relation learning without group activity annotation through recognition of individual action that can be more easily annotated than group activities defined with complex inter-people relationships. Our method extracts features informative for recognizing the action of each person by conditioning the group relation with the location of this person. A variety of experimental results demonstrate that our method outperforms SOTA methods quantitatively and qualitatively on two public datasets.