The code of paper Robust Self-supervised LiDAR Odometry via Representative Structure Discovery and 3D Inherent Error Modeling accepted by IEEE Robotics and Automation Letters (RA-L), 2021.
The self-supervised two-frame odometry network contains three main modules including the Geometric Unit Feature Encoding module, Geometric Unit Transformation Estimation module and the Ego-motion voting module.
The comparison (on estimated trajectories) of our method with other competitive baselines (left). The visualization of our estimated point covariances (right).
As the dependencies is complex, a dockerfile has been provide. You need to install docker and nvidia-docker2 first and then set up the docker image and start up a container with the following commands:
cd RSLO
sudo docker build -t rslo .
sudo docker run -it --runtime=nvidia --ipc=host --volume="HOST_VOLUME_YOU_WANT_TO_MAP:DOCKER_VOLUME" -e DISPLAY=$DISPLAY -e QT_X11_NO_MITSHM=1 rslo bash
You need to download the KITTI odometry dataset and unzip them into the below directory structures.
./kitti/dataset
|──sequences
| ├── 00/
| | ├── calib.txt
| │ ├── velodyne/
| | | ├── 000000.bin
| | | ├── 000001.bin
| | | └── ...
| ├── 01/
| | ...
| └── 21/
└──poses
|──00.txt
|──01.txt
| ...
└──10.txt
Then, create hdf5 data with
python script create_hdf5.py ./kitti/dataset ./kitti/dataset/all.h5
The trained models on the KITTI dataset have been uploaded to the OneDrive. You can download them and put them into the directory "weights" for testing.
export PYTHONPATH="$PROJECT_ROOT_PATH:$PYTHONPATH"
export PYTHONPATH="$PROJECT_ROOT_PATH/rslo:$PYTHONPATH"
python -u $PROJECT_ROOT_PATH/evaluate.py multi_proc_eval \
--config_path $PROJECT_ROOT_PATH/config/kitti_eval_ours.prototxt \
--model_dir ./outputs/ \
--use_dist True \
--gpus_per_node 1 \
--use_apex True \
--world_size 1 \
--dist_port 20000 \
--pretrained_path $PROJECT_ROOT_PATH/weights/ours.tckpt \
--refine False \
Note that you need to specify the PROJECT_ROOT_PATH, i.e. the absolute directory of the project folder "RSLO" and modify the path to the created data, i.e. all.h5, in the configuration file kitti_eval_ours.prototxt before running the above commands. A bash script "script/eval_ours.sh" is provided for reference.
A basic training script demo is shown as below. You can increase the GPU number, i.e. the variable "GPUs", according to your available resources. Generally, larger batch sizes produce stabler training procedures and better final performances.
export PYTHONPATH="$PROJECT_ROOT_PATH:$PYTHONPATH"
export PYTHONPATH="$PROJECT_ROOT_PATH/rslo:$PYTHONPATH"
GPUs=1 # the number of gpus you use
python -u $PROJECT_ROOT_PATH/train_hdf5.py multi_proc_train \
--config_path $PROJECT_ROOT_PATH/config/kitti_train_ours.prototxt \
--model_dir ./outputs/ \
--use_dist True \
--gpus_per_node $GPUs \
--use_apex True \
--world_size $GPUs \
--dist_port 20000 \
--refine False \
We thank for the open-sourced codebases spconv and second
To cite our paper
@article{xu2022robust,
title={Robust Self-supervised LiDAR Odometry via Representative Structure Discovery and 3D Inherent Error Modeling},
author={Xu, Yan and Lin, Junyi and Shi, Jianping and Zhang, Guofeng and Wang, Xiaogang and Li, Hongsheng},
journal={IEEE Robotics and Automation Letters},
year={2021},
publisher={IEEE}
}
@inproceedings{xu2020selfvoxelo,
title = {SelfVoxeLO: Self-supervised LiDAR Odometry with Voxel-based Deep Neural Networks},
author = {Yan Xu and Zhaoyang Huang and Kwan{-}Yee Lin and Xinge Zhu and Jianping Shi and Hujun Bao and Guofeng Zhang and Hongsheng Li},
booktitle = {4th Conference on Robot Learning, CoRL 2020, 16-18 November 2020, Virtual Event / Cambridge, MA, {USA}},
volume = {155},
pages = {115--125},
publisher = {{PMLR}},
year = {2020},
}
- Inference code and pretrained models (9/10/2022)
- Training code (10/12/2022)
- Code cleaning and refactor
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent