HumanRig – A 3D Humanoid Character Automatic Rigging Task Dataset Released by Alibaba Amap

What is HumanRig

HumanRig is a 3D humanoid character automatic rigging research project developed by Alibaba’s AMAP team. It addresses the limitations of existing rigging technologies caused by the lack of high-quality datasets. By providing a large-scale, high-quality dataset and an innovative automatic rigging framework, HumanRig advances the automation of 3D character animation production.​

The HumanRig dataset includes 11,434 T-pose meshes that adhere to a unified skeleton topology and exhibit diverse head-to-body ratios, filling the gaps in existing datasets regarding scale, diversity, and skeletal consistency. The automatic rigging framework utilizes a Prior-Guided Skeleton Estimator (PGSE) and a Mesh-Skeleton Mutual Attention Network (MSMAN) to achieve coarse-to-fine 3D skeleton joint regression and skinning weight estimation, generating characters suitable for animation production with performance surpassing existing methods.

​

Main Features of HumanRig

• Provision of a Large-Scale High-Quality Dataset: HumanRig is the first large-scale dataset specifically designed for the task of automatic rigging of 3D humanoid characters, containing 11,434 high-quality AI-generated humanoid meshes. All models are presented in T-pose and strictly follow industry-standard skeleton topology, making them directly applicable to mainstream animation engines. The dataset boasts significant advantages in scale, diversity, and skeletal consistency, covering a wide range of character types from realistic humans to cartoon characters and even anthropomorphic animals.​

• Prior-Guided Skeleton Estimator (PGSE): Based on projecting 2D prior information into 3D space, it initializes a coarse skeleton, significantly reducing the complexity of the rigging task.​

• U-shaped Point Transformer as Mesh Encoder: By discarding the reliance on edge information of 3D meshes, it enhances the robustness of rigging on complex meshes.​

• Mesh-Skeleton Mutual Attention Network (MSMAN): By integrating mesh and skeleton features in a high-level semantic space, it achieves joint optimization of skeleton construction and skinning.​

Technical Principles of HumanRig

• Construction of the HumanRig Dataset:

  • Generation of Diverse 2D Images: Utilizing AI-driven 2D image generation technology, diverse T-pose character images are generated from textual descriptions.​

  • Generation of High-Quality 3D Meshes: Tools such as InstantMesh and Unique3D are employed to convert 2D images into high-quality 3D meshes.​

  • Selection and Optimization: From 17,268 initial meshes, 14,662 high-quality models are selected, and semi-automated tools like Mixamo are used for skeletal rigging, ultimately forming 11,434 high-quality rigged models.​

• Automatic Rigging Process:

  • Skeleton Initialization: A coarse skeleton is generated through the PGSE module.​

  • Feature Extraction: Skeleton and mesh features are extracted using an MLP-based skeleton encoder and a U-shaped Point Transformer, respectively.​

  • Feature Fusion and Optimization: The MSMAN module fuses skeleton and mesh features, achieving coarse-to-fine skeleton joint regression and skinning weight estimation.​

  • Generation of Animated Characters: The optimized skeleton and skinning weights are combined to generate characters suitable for animation production.​

Project Address

• Project Website: https://c8241998.github.io/HumanRig/

• arXiv Technical Paper: https://arxiv.org/pdf/2412.02317

Application Scenarios of HumanRig

• Game Development: In game development, HumanRig’s automatic rigging technology can significantly reduce the time and cost of character animation production, performing excellently when handling complex character models (such as those with intricate clothing or accessories).​

• Film Production: In the film industry, HumanRig’s automatic rigging technology can quickly generate high-quality character rigging, significantly enhancing production efficiency.​

• Virtual Reality (VR) and Augmented Reality (AR): In VR and AR applications, real-time interactive character animation is key to enhancing immersion. HumanRig’s automatic rigging technology can provide real-time skeletal animation support for virtual characters, ensuring natural and smooth character movements.​

• 3D Digital Humans: Through automatic rigging technology, services like Amap can quickly generate personalized 3D digital humans, offering users more interactive and engaging navigation experiences.​