Automatic rigging of 3D models with stacked hourglass networks and descriptors

• Main Authors: Verma, Ritika, Mittal, Sarthak, Pawar, Siddharth, Sharma, Moolchand, Goel, Shalini, Albuquerque, Victor Hugo C. de
• Format: Conference Proceeding
• Language: English
• Subjects: Animation; Rigging; Three dimensional models
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0184393

We put forward an approach for automated skeleton rigging of 3D point cloud models of segmented characters. Unlike earlier systems that fit predetermined skeleton templates or forecast predetermined sets of joints, our approach generates an animation skeleton that is tuned to the structure and geometry of the input 3D model. Our architecture is built on a stack of hourglass models trained using a large dataset of 3D-rigged characters mined from the web. It works with a volumetric representation of the input 3D shapes enhanced with geometric shape elements that provide different indications for joint and bone positions. The proposed method also allows straightforward user customization of the output skeleton’s level of detail. Our study shows that, compared to many alternatives and baselines, our approach predicts animation skeletons that are significantly more comparable to those made by people.