Neural style transfer for character motion synthesis

Abstract

In the field of animation, three-dimensional characters are assigned skeletons, or character rigs, to sim ulate the naturalistic movements of their real-world counterparts. Animators manipulate the character rig to allow for smooth, naturalistic movement. This research proposes using a combination of deep learning techniques to learn the manifold of human character motion, trained on motion capture data. The motion manifold is represented by the latent space of the convolutional auto-encoder. This repre sentation is subsequently transferred to any given locomotion, providing a high-quality motion sequence with a different style. This research builds upon the recent work of neural style transfer for images, but transports it to a new domain of human character motion. A qualitative study has been conducted on a group of participants to determine the success of the framework’s ability in transferring style to a newly synthesised locomotion. Participants were able to recognise stylised aspects within the synthesised loco motions generated from the deep learning framework. An investigation into the relationship between the amount of style and content on the style transfer constraint is presented. Further, the effect of varying gaits on the style transfer process between actors running and walking in a normal and zombie style are investigated. One should view style and content as an inter-related concept. This research provides an in terval of acceptable amounts of style. After a point, the more style placed on the style transfer constraint does not yield major changes on the newly synthesised locomotion. The zombie stylise motion manifold took longer to transfer when the actor was running. This research demonstrates that style transfer for human locomotion is capable