DTG: Learning A Dynamic Token Graph for 3D Pose Forecasting

3D pose forecasting aims to predict future pose sequences based on historical poses, which has a wide range of practical applications. Previous methods mainly focus on representing body joints as 3D coordinates, yet ignoring the dependency modeling between joints. Furthermore, human forecasting is unstable when variations in time and pose type are considered. Therefore, we propose a Dynamic Token Graph Network (DTG) for 3D pose forecasting. First, to model the dependency between the body joints effectively, we represent joints by the composition of discrete tokens (see Fig. 1(b)) to replace 3D coordinates. Second, we design a novel dynamic graph neural network architecture to characterize the correlations of joints as time and pose type changes (e.g. Sitting and Walking). Comprehensive experiments on Human 3.6M, AMASS, and 3DPW datasets confirm the superiority of our method, which is applicable to both angle-based and coordinate-based pose representations.