Data Augmentation for Human Motion Prediction

Takahiro Maeda, Norimichi Ukita

Abstract: Human motion prediction is seldom deployed to real-world tasks due to difficulty in collecting a huge amount of motion data. We propose two motion data augmentation approaches using Variational AutoEn-coder (VAE) and Inverse Kinematics (IK). Our VAE-based generative model with adversarial training and sampling near samples generates various motions even with insufficient original motion data. Our IK-based augmentation scheme allows us to semi-automatically generate a variety of motions. Furthermore, we correct unrealistic artifacts in the augmented motions. As a result, our method outperforms previous noise-based motion augmentation methods.