Diverse human motion prediction via sampling on Grassmann manifold

Existing researches capture the multimodal nature of human motion through likelihood-based sampling on latent space. However, in the actual motion process, the high diversity of human training data often leads to mode collapse. Additionally, the training process is complicated by a large number of parameters. In this paper, a novel yet effective sampling method on the Grassmann manifold is proposed to enhance the accuracy and diversity of prediction. A linear dynamical system is utilized to model the spatiotemporal dependence in human motion sequences. The corresponding orthogonal basis vectors are connected as residuals of the encoded motion sequences. The feature space is enhanced by these orthogonal basis vectors. Subsequently, a series of random orthogonal vectors are sampled on the Grassmann manifold using Poisson weights and a reparameterization trick. The method is compared with the latest methods on the Human3.6M dataset and the HumanEva-I dataset. The results show that the method effectively improves the diversity metric by 3% and 10% with a low average error, code and pre-trained models are available at: https://github.com/Hq-Tong/SOGM.