JOintGS / README.md
nielsr's picture
nielsr HF Staff
Add model card for JOintGS
ce7afc4 verified
|
raw
history blame
2 kB
metadata
pipeline_tag: image-to-3d
license: cc-by-nc-sa-4.0

JOintGS: Joint Optimization of Cameras, Bodies and 3D Gaussians for In-the-Wild Monocular Reconstruction

This repository contains the pre-trained checkpoints for JOintGS, a unified framework that jointly optimizes camera extrinsics, human poses, and 3D Gaussian representations for robust, animatable 3D human avatar reconstruction from monocular video with coarse initialization.

Introduction

Reconstructing high-fidelity animatable 3D human avatars from monocular RGB videos remains challenging, particularly in unconstrained in-the-wild scenarios where camera parameters and human poses from off-the-shelf methods (e.g., COLMAP, HMR2.0) are often inaccurate.

JOintGS enables a synergistic refinement mechanism where explicit foreground-background disentanglement allows mutual reinforcement: static background Gaussians anchor camera estimation via multi-view consistency; refined cameras improve human body alignment through accurate temporal correspondence; and optimized human poses enhance scene reconstruction by removing dynamic artifacts from static constraints.

License

  • Source Code: The software in the associated GitHub repository is licensed under the MIT License.
  • Model Weights: The pre-trained checkpoints in this repository are released under the CC BY-NC-SA 4.0 License.

Citation

@article{jointgs2026,
  title={JOintGS: Joint Optimization of Cameras, Bodies and 3D Gaussians for In-the-Wild Monocular Reconstruction},
  author={Jiuhai Chen and others},
  journal={arXiv preprint arXiv:2602.04317},
  year={2026}
}