Pose estimation for objects with rotational symmetry

Official implementation of the code for the paper published in IROS18. We generate a synthetic database of thousands of different objects for pose estimation. These follow realistic scenarios with several objects on a table, with objects that contain several levels of symmetries. We propose a model that is able to predict objects symmetries and poses. More information about the project in our project page or the paper

REQUIREMENTS

Generating the dataset and running experiments requires PyTorch 0.4. The CAD models are available for download here. They are based by a realistic database of textures which can be download from here

RUN CODE

To generate a synthetic dataset from the object CAD models please see the dataset generation folder

The training code is based on PyTorch and more documented in the training folder

CITATION

If you use this code or ideas from the paper in your research, please cite our paper:

@inproceedings{corona2018pose,
  title={Pose Estimation for Objects with Rotational Symmetry},
  author={Corona, Enric and Kundu, Kaustav and Fidler, Sanja},
  booktitle={2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={7215--7222},
  year={2018},
  organization={IEEE}
}