Novel Object 6D Pose Estimation with a Single Reference View

This is the PyTorch implementation of paper SinRef-6D developed by J. Liu, W. Sun, K. Zeng, J. Zheng, H. Yang, L. Wang, H. Rahmani, and A. Mian. SinRef-6D is a single reference view-based CAD model-free novel object 6D pose estimation method, which is simple yet effective and has strong scalability for practical applications.

Real-World Demo

SinRef-6D deployment in real-world robotic manipulation scenarios. Notably, the reference view is not carefully selected. We randomly select a normal view (free of occlusion, with minimal self-occlusion, and resembling a typical robotics manipulation view) using an Intel RealSense L515 RGB-D camera as the reference view.

To the best of our knowledge, we are the first to present a method for novel object 6D absolute pose estimation using only a single reference view in real-world robotic manipulation scenarios. This approach simultaneously eliminates the need for object CAD models, dense reference views, and model retraining, offering enhanced efficiency and scalability while demonstrating strong generalization to potential real-world robotic applications.

The complete code will be released after paper acceptance.

Citation

If you find our work useful, please consider citing:

@article{SinRef-6D,
  author={Liu, Jian and Sun, Wei and Zeng, Kai and Zheng, Jin and Yang, Hui and Wang, Lin and Rahmani, Hossein and Mian, Ajmal},
  title={Novel Object 6D Pose Estimation with a Single Reference View},
  journal={arXiv preprint arXiv:2503.05578},
  year={2025}
}

Licence

This project is licensed under the terms of the MIT license.