Real-time 3D human motion generation with labeled motion capture data

Abstract

Recently, users encounter 3D human motion easily than before according to computer graphics development. Real-time motion generation is often required for interactivity in some motion generation fields such as games. Data-driven generative models for motion demand enough amount of motion clips as a training data set for each type of motion, and it is raised as a problem when the model should generate unusual motions whose motion clips are barely found. Therefore, in this research, we propose a model named Labeled Motion VAEs(L-MVAE). Existing MVAE, or Motion VAEs, uses the latent space of conditional variational autoencoder for reinforcement learning to synthesize desired 3D human motion and can generate motions in real-time. In L-MVAE, label information of motions is added to MVAE with a concept of label distribution. L-MVAE can synthesize motions in real-time likewise, and it produces high-quality motions even with a relatively small size of training data set. The performance of the proposed model is compared with the existing MVAE using the task in which a character moves to a destination.