Weakly-supervised Action Transition Learning for Stochastic Human Motion Prediction

Abstract

Figure 1. Action-driven stochastic human motion prediction. Given a past motion (pink) and a sequence of future action labels, our model generates action-specific future poses (yellow). We show two different futures generated for the same history and actions. Our model allows these predictions to have different lengths. The motions are down-sampled to the same frame rate for visualisation.