Interaction Transformer for Human Reaction Generation

We address the challenging task of human reaction generation, which aims to generate a corresponding based on an input action. Most existing works do not focus generating and predicting cannot motion when only action is given as input. To this limitation, we propose novel interaction Transformer (InterFormer) consisting network with both temporal spatial attention. Specifically, attention captures dependencies characters their interaction, while learns between different body parts each character those are part interaction. Moreover, using graphs increase performance via distance module that helps nearby joints from characters. Extensive experiments SBU K3HI, DuetDance datasets demonstrate effectiveness InterFormer. Our method general can be used more complex long-term interactions. also provide videos generated reactions code pre-trained models at https://github.com/CRISTAL-3DSAM/InterFormer