Automatic skeleton generation using hierarchical mesh segmentation

Abstract

High fidelity and interactive 3D characters abound in virtual and augmented reality. However, making one usually requires manual skeleton extraction and specification on how the body reacts with the translation and rotation of the bones. While there are previous works that already solve this problem, most systems have limitations with the input or the output cannot be directly used for animation. With the goal of bringing any humanoid toys to life, a pipeline is presented to automatically generate the toy's skeletal structure and skinning weights. First, the mesh is segmented into several parts using normal characteristic value (NCV) and global point signatures (GPS) as candidate points for segmentation. Then, joint locations are generated based on the segmentation results. Further-more, the skinning weights are generated by solving the Laplace diffusion equation. Experimental results show that our pipeline is robust enough to extract skeletal structures from graphic artists' models as well as from scanned models. In addition, our pipeline is deformation-invariant as it can generate the same skeletal structure of a model having different poses. Finally, our pipeline achieves both appealing virtual realism and fast speed. The output can be directly used to setup skeleton-based animations in games as well as real-time virtual and augmented reality applications within minutes.