Manifold-Aware Self-Training for Unsupervised Domain Adaptation on Regressing 6D Object Pose

Installation

git clone [email protected]:Gorilla-Lab-SCUT/MAST.git
mkdir deps && cd deps
git clone https://github.com/ylabbe/bop_toolkit_cosypose.git
git clone https://github.com/ylabbe/bullet3.git
git clone https://github.com/ylabbe/job-runner
git clone https://github.com/ylabbe/bop_toolkit_challenge20
cd ..
conda env create -n MAST --file MAST/config_env/environment.yaml
conda activate MAST
python setup.py develop  # install locally
runjob-config MAST/config_env/job-runner-config.yaml  # config runjob

Downloading and preparing data

• create a folder local_data
• download bop_datasets
• download Linemod_preprocessed
• download OCCLUSION_LINEMOD
• download VOCdevkit
• go to CosyPose google drive download URDF files and put them in local_data/urdfs, download detector-bop-lmo-pbr--517542 and put it in local_data/experiments

Training

• step 1: To train on synthesis datasets, using scripts in train_src.sh
• step 2: To train on both real and synthesis datasets, using scripts in train_st.sh

Testing

Please see the scripts in test.sh

Acknowledgements

Our implementation leverages the code from CosyPose.

Citation

@inproceedings{ijcai2023p193,
  title     = {Manifold-Aware Self-Training for Unsupervised Domain Adaptation on Regressing 6D Object Pose},
  author    = {Zhang, Yichen and Lin, Jiehong and Chen, Ke and Xu, Zelin and Wang, Yaowei and Jia, Kui},
  booktitle = {Proceedings of the Thirty-Second International Joint Conference on
               Artificial Intelligence, {IJCAI-23}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {Edith Elkind},
  pages     = {1740--1748},
  year      = {2023},
  month     = {8},
  note      = {Main Track},
  doi       = {10.24963/ijcai.2023/193},
  url       = {https://doi.org/10.24963/ijcai.2023/193},
}