CorrI2P: Deep Image-to-Point Cloud Registration via Dense Correspondence [arxiv, TCSVT]

Accepted by IEEE TCSVT

News !!!

Our new work Self-supervised Learning of LiDAR 3D Point Clouds via 2D-3D Neural Calibration achieves a higher accuracy, and we refer the readers to follow! [Code]

Data

KITTI

Here we provide KITTI prepared.
You can download it here.
Unzip these files, and the directory is as follows:

kitti
-calib
--00
--01
...
-sequences
--00
--01
...

NuScenes

Here we provide nuScenes prepared.
You can download it here.
We also provide the script for preparing NuScenes dataset in nuScenes_script folder (reffer to DeepI2P). They can be used to generate nuscenes dataset.

Usage

Install required lib as SO-Net or DeepI2P.
Indexmax

Train

python train.py

Test

python eval_all.py
python cal_error_all.py
python analysis.py

Note: There would be lots of intermediate results, please leave enough storage space.

Citation

@article{ren2022corri2p,
  title={Corri2p: Deep image-to-point cloud registration via dense correspondence},
  author={Ren, Siyu and Zeng, Yiming and Hou, Junhui and Chen, Xiaodong},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  volume={33},
  number={3},
  pages={1198--1208},
  year={2022},
  publisher={IEEE}
}

Acknowledgement

We thank the authors of DeepI2P for their public code.

If you want to use our code, please cite our work.