An Elastic Shape Variational Autoencoder for Skeleton Pose Trajectories
The Elastic Shape Variational Autoencoder (ES-VAE) is a geometry-aware generative model for skeletal trajectories. It embeds movement into a low-dimensional space using transported square-root velocity fields on Kendall's shape manifold, achieving complete invariance to changes in shape, rotation, and speed. This method outperforms standard VAEs and other sequence models in activity recognition and clinical prediction.