Abstract—We present a technique to learn large human motion data captured with optical motion capture system, represent it in a low dimensional latent space, so as to generate natural and various human motions from it. To extract human motion features we use a convolutional autoencoder, and to represent the extracted features as a probability density function in a latent space we use a variational autoencoder. Motion generator is modeled as a map from a latent variable sampled in the latent space to a motion capture data. We stack the convolutional decoder on top of the variational decoder, which can sample a latent variable and produce a motion. As a result, our system can generate natural and various human motions from a 32-dimensional latent space.
Index Terms—Character animation, convolutional autoencoder, motion generative model, variational autoencoder.
Yuichiro Motegi and Makoto Murakami are with Graduate School of Information Sciences and Arts, Toyo University, Saitama, Japan (e-mail: s3b101710012@toyo.jp, murakami_m@toyo.jp).
Yuma Hijioka is with Graduate School of Science and Engineering, Toyo University, Saitama, Japan (e-mail: yuma.hijioka@ieee.org).
[PDF]
Cite: Yuichiro Motegi, Yuma Hijioka, and Makoto Murakami, "Human Motion Generative Model Using Variational Autoencoder," International Journal of Modeling and Optimization vol. 8, no. 1, pp. 8-12, 2018.
