AutoSkin: Skeleton-based Human Skinning with Deep Neural Networks

Abstract

Articulated character animation has received increasing attention in applications like game or film production. One of the main challenges is the task of skinning that animator should specify the association between skeleton structure and skin surface which involves large amounts of manual weight painting and deformation tuning. In this paper, we propose a deep learning based framework for skinning. By using our framework, the skinning is seen as a multi-regression task and the network can infer the skeleton-skin association and compute the accurate weights automatically while only requiring minimal manual tuning. One significant advantage of our framework is that the impressive performance can be also achieved while dealing with complicated 3D characters, even these characters exist over a hundred bones or dress delicately. Our experiments have indicated that our framework could generate competitive results compared with the commercial software and we aim to apply it in the industry such as games or video applications.