Reproduce “GauHuman: Articulated Gaussian Splatting from Monocular Human Videos” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shoukang Hu, Ziwei Liu
GauHuman, developed by S-Lab at Nanyang Technological University, introduces an articulated 3D Gaussian Splatting framework to reconstruct and render high-fidelity human models from monocular videos. The system achieves training times of 1-2 minutes and real-time rendering at up to 189 FPS, while maintaining state-of-the-art visual quality across diverse datasets.
Key insights
- Integrating established human priors (like SMPL) through LBS transformation and initialization within 3D Gaussian Splatting drastically accelerates training and improves reconstruction quality for articulated subjects.
- Carefully designed adaptive control for 3D Gaussians, utilizing KL-divergence to guide splitting/cloning and a novel merging strategy, is crucial for achieving both high fidelity and computational efficiency without excessive redundancy.
- Refining initial pose and LBS weights with small neural networks, while maintaining a canonical Gaussian representation, enables the capture of fine-grained human details like cloth wrinkles efficiently.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.