Weakly-Supervised Motion In-Betweening Learning via Pose Priors

Human motion in-betweening plays an essential role in the fields of film, modern games, and animation creation. Existing motion in-betweening approaches generally require the entire ground-Truth sequence for supervision. However, acquiring a training sequence that is both long-Term and accurate in the real world can be challenging. Because we cannot realistically expect the dataset to always cover motions of sufficient length, in this study, we propose a weakly supervised motion in-betweening method that aims to smoothly connect two sequences. We propose a stochastic human motion in-betweening learning framework to enable generation with a high sample diversity. Specifically, we leverage two pose level prior terms to enforce weak supervision of the generated human motion during training. We also perform a Discrete Cosine Transform to pursue motion smoothness and employ graph convolutional networks to extract temporal dependencies from skeleton-based human data. We conduct experiments on the Human3.6M dataset and demonstrate the effectiveness of our method in generating realistic yet smooth human motion with only weak supervision.