The repository contains the source code and pre-trained models of our paper (published on ICCV 2021): HRegNet: A Hierarchical Network for Large-scale Outdoor LiDAR Point Cloud Registration.
The overall network architecture is shown below:
We have fixed some bugs in the code and updated the pretrained weights for both two datasets. The registration performance would be better than the reported performance in the paper. The current tested results are listed here:
| RTE (m) | RRE (deg) | Success rate |
|---|---|---|
| 0.0557+-0.0746 | 0.1780+-0.1959 | 99.77% |
| RTE (m) | RRE (deg) | Success rate |
|---|---|---|
| 0.1218+-0.1122 | 0.2734+-0.1970 | 100.0% |
The code mainly requires the following libraries and you can check requirements.txt for more environment requirements.
- PyTorch 1.7.0/1.7.1
- Cuda 11.0/11.1
- pytorch3d 0.3.0
- MinkowskiEngine 0.5
Please run the following commands to install point_utils
cd models/PointUtils
python setup.py install
Training device: NVIDIA RTX 3090
The point cloud pairs list and the ground truth relative transformation are stored in data/kitti_list and data/nuscenes_list.
The data of the two datasets should be organized as follows:
DATA_ROOT
├── 00
│ ├── velodyne
│ ├── calib.txt
├── 01
├── ...
DATA_ROOT
├── v1.0-trainval
│ ├── maps
│ ├── samples
│ │ ├──LIDAR_TOP
│ ├── sweeps
│ ├── v1.0-trainval
├── v1.0-test
│ ├── maps
│ ├── samples
│ │ ├──LIDAR_TOP
│ ├── sweeps
│ ├── v1.0-test
The training of the whole network is divided into two steps: we firstly train the feature extraction module and then train the network based on the pretrain features.
- Train keypoints detector by running
sh scripts/train_kitti_det.shorsh scripts/train_nusc_det.sh, please reminder to specify theGPU,DATA_ROOT,CKPT_DIR,RUNNAME,WANDB_DIRin the scripts. - Train descriptor by running
sh scripts/train_kitti_desc.shorsh scripts/train_nusc_desc.sh, please reminder to specify theGPU,DATA_ROOT,CKPT_DIR,RUNNAME,WANDB_DIRandPRETRAIN_DETECTORin the scripts.
Train the network by running sh scripts/train_kitti_reg.sh or sh scripts/train_nusc_reg.sh, please reminder to specify the GPU,DATA_ROOT,CKPT_DIR,RUNNAME,WANDB_DIR and PRETRAIN_FEATS in the scripts.
Update: Pretrained weights for detector and descriptor are provided in ckpt/pretrained. If you want to train descriptor, you can set PRETRAIN_DETECTOR to DATASET_keypoints.pth. If you want to train the whole network, you can set PRETRAIN_FEATS to DATASET_feats.pth.
We provide pretrain models in ckpt/pretrained, please run sh scripts/test_kitti.sh or sh scripts/test_nusc.sh, please reminder to specify GPU,DATA_ROOT,SAVE_DIR in the scripts. The test results will be saved in SAVE_DIR.
If you find this project useful for your work, please consider citing:
@InProceedings{Lu_2021_HRegNet,
author = {Lu, Fan and Chen, Guang and Liu, Yinlong and Zhang Lijun, Qu Sanqing, Liu Shu, Gu Rongqi},
title = {HRegNet: A Hierarchical Network for Large-scale Outdoor LiDAR Point Cloud Registration},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision},
year = {2021}
}
We want to thank all the ICCV reviewers and the following open-source projects for the help of the implementation:
- DGR(Point clouds preprocessing and evaluation)
- PointNet++(unofficial implementation, for Furthest Points Sampling)
