Data-driven approach for body motion retargeting

Abstract

Motion retargeting(MR) deals with the problem of adapting the motion of one articulated animated character to another, potentially with different skeletal structure, like different bone lengths, or a completely different topology. Clearly, the animation and video-game industry would be highly benefited by the existence of such an automated tool. In this thesis, we introduce a novel neural architecture using transformers as the building block. To the best of our knowledge, this is the first work that employs transformers for MR. Similar work uses either RNN, CNN or temporal CNN. Additionally, by manipulating the skeletal graph, we propose a simple technique for seamless retargeting between skeletons with different number of joints, thus handling intra and cross-structure retargeting in a unified way. The system produces excellent results for intra-structure retargeting, and decent results for cross-structure retargeting, while succeeding in maintaining the original motion content. Finally, the approach is extensible to any number of datasets and any number of character joints.