Trajectory-Aware Body Interaction Transformer for Multi-Person Pose Forecasting

Figure 1. (a) In complex crowd scenarios, different people may interact with one another at varying levels (low and high interactions) and at different positions (i.e., between near and far distances). (b) The illustration of our main idea on body part interactions. We divide the body joints into 5 parts, and the Intra-Individual branch is used to explore part relationships for each individual and the Inter-Individual branch aims to capture interaction dependencies of body parts between individuals. Our TBIFomer facilitates to model body part interactions for intra-and inter-individuals simultaneously.