Learning character motion alignment with periodic autoencoders

Abstract

The present disclosure provides a periodic autoencoder that can be used to generate a general motion manifold structure using local periodicity of the movement whose parameters are composed of phase, frequency, and amplitude. The periodic autoencoder is a novel neural network architecture that can learn periodic features from large unstructured motion datasets in an unsupervised manner. The character movements can be decomposed into multiple latent channels that can capture the non-linear periodicity of different body segments during synchronous, asynchronous, and transition movements while progressing forward in time, such that it captures spatial data and temporal data associated with the movements.