Motion In-Betweening With Spatial and Temporal Transformers

Motion in-betweening that aims to generate motion transitions between known keyframes plays a significant role in the 3D character animation industry. However, generating long-term transitions is highly challenging due to the non-stationary nature and considerable spatio-temporal uncertainty of motions. Leading transformer-based methods operate at a single temporal scale while they overlook the spatial interactions among joints and temporally hierarchical structure of motions, leading to the generation of over-smoothed and weak transitions. In this paper, we propose a novel spatio-temporal framework for the motion in-betweening task. First, a spatial transformer is introduced to capture the per-frame spatial dependencies among joints, enhancing the capability of the model to generalize across diverse action types. Furthermore, to alleviate over-smoothing, a multi-scale temporal transformer is designed to generate dynamic and realistic transitions by capturing the hierarchical structure of motions, which includes both global motion trends and local subtle variations. Extensive experiments on the LAFAN1 dataset demonstrate that our method achieves state-of-the-art performance compared to existing methods. In addition, the corresponding ablation studies and sensitivity analyses verify the effectiveness of the proposed spatio-temporal framework.