Reproduce “RigNet: Neural Rigging for Articulated Characters” on OpenResearch

About this paper

Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, Karan Singh

RigNet, developed by researchers from the University of Massachusetts Amherst and University of Toronto, introduces an end-to-end deep learning pipeline for automatically generating animation rigs from 3D character meshes, encompassing both skeleton prediction and skin weight estimation. It achieves state-of-the-art results, with 61.6% IoU for skeleton prediction and an average L1-norm error of 0.39 for skinning weights, outperforming previous methods like Pinocchio and NeuroSkinning.

Key insights

• Neural mesh attention and differentiable clustering, modulated by learned attention, enable adaptive and precise skeletal joint localization on 3D meshes.
• A learned BoneNet and RootNet, combined with a Minimum Spanning Tree algorithm, accurately predict the hierarchical connectivity and root of the animation skeleton.
• Using volumetric geodesic distances within a GMEdgeNet provides a skeleton-aware mesh representation that significantly improves skin weight prediction, reflecting anatomical relationships.

Summary by alphaXiv.