UnPose: Uncertainty-Guided Diffusion Priors for Zero-Shot Pose Estimation

Zhaodong Jiang$^{1,2}$, Ashish Sinha$^{1}$, Tongtong Cao$^{1}$, Yuan Ren$^{1}$, Bingbing Liu$^{1}$, Binbin Xu$^{1}$

$^{1}$ Huawei Noah's Ark Lab, Toronto $^{2}$ University of Toronto, Canada

This repository contains the code and video demos for our project website UnPose accepted to CORL 2025.

TL;Dr: A zero-shot, model-free 6D pose estimation and reconstruction framework that incrementally refines a 3D Gaussian Splatting model using diffusion priors and uncertainty-guided fusion from RGB-D inputs.

Core Idea

Estimate epistemic uncertainty from a pretrained 2D-to-3D diffusion model to continually refine a 3DGS-represented object for 6DOF pose estimation in a factor-graph optimzation framework.

BibTex

If you use this work in your research, please cite our paper:

@inproceedings{jiang2025unpose,
  title={UnPose: Uncertainty-Guided Diffusion Priors for Zero-Shot Pose Estimation},
  author={Jiang, Zhaodong and Sinha, Ashish and Cao, Tongtong and Ren, Yuan and Liu, Bingbing and Xu, Binbin},
  booktitle={Conference on Robot Learning (CoRL)},
  year={2025}
}

Acknowledgments

Parts of this project page were adopted from the Nerfies page.

Website License

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.