MatchNorm

This repository contains python scripts for training and testing of paper Learning-based Point Cloud Registration for 6D Object Pose Estimation in the Real World, ECCV2022.

Prerequisites

1. BOP Toolkit

2. TUD-L, LINEMOD and Occluded-LINEMOD

3. conda env create -f environment.yml

Command Line

Training

python scratch.py --mode='ori' --bop_dataset='tudl'

Testing

[mAP Metric]

python scratch.py --mode='ckpt' --bop_dataset='tudl' --exp_name='bpnet_tudl'

python scratch.py --mode='ckpt' --bop_dataset='lm' --exp_name='bpnet_lm'

python scratch.py --mode='ckpt' --bop_dataset='lmo' --exp_name='bpnet_lm'

[BOP Metric]

python eval_bop19.py --result_filenames='zheng-bpnet_tudl-test.csv'

python eval_bop19.py --result_filenames='zheng-bpnet_lm-test.csv'

python eval_bop19.py --result_filenames='zheng-bpnet_lmo-test.csv'

Pretrained Model

1. pretrained models are saved at ./ckpt

2. bop benchmark result files are saved at ./bop_res

Citation

@inproceedings{Zheng2022,
    title={Learning-based Point Cloud Registration for 6D Object Pose Estimation in the Real World},
    author={Dang, Zheng and Wang, Lizhou and Guo, Yu and Salzmann, Mathieu},
    booktitle = {European Conference on Computer Vision(ECCV) 2022},
    month = {October},
    year={2022}
}

License

MIT License