LiFlow: Flow Matching for 3D LiDAR Scene Completion

Dependencies

Installing python and pre-requisites packages with Anaconda:

conda create -n liflow python=3.9.21

conda activate liflow

conda install pytorch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 pytorch-cuda=11.8 -c pytorch -c nvidia

conda install openblas-devel -c anaconda

conda install cuda-toolkit -c nvidia/label/cuda-11.8.0

conda install cudatoolkit==11.8 -c pytorch

conda install anaconda::cython

Installing MinkowskiEngine:

export CUDA_HOME=$CONDA_PREFIX

git clone https://github.com/NVIDIA/MinkowskiEngine.git
cd MinkowskiEngine
python setup.py install --blas_include_dirs=${CONDA_PREFIX}/include --blas=openblas

Installing pytorch3D:

conda install -c iopath iopath

pip install "git+https://github.com/facebookresearch/pytorch3d.git"

Installing other dependencies on the code main directory:

cd LiFlow
pip install -r requirements.txt

To setup the code run the following command on the code main directory:

cd LiFlow
pip install -U -e .

The SemanticKITTI Dataset

The SemanticKITTI dataset has to be download from site and extracted in the following structure:

./liflow/
└── Datasets/
    └── SemanticKITTI
        └── dataset
          └── sequences
            ├── 00/
            │   ├── velodyne/
            |   |       ├── 000000.bin
            |   |       ├── 000001.bin
            |   |       └── ...
            │   └── labels/
            |       ├── 000000.label
            |       ├── 000001.label
            |       └── ...
            ├── 08/ # for validation
            ├── 11/ # 11-21 for testing
            └── 21/
                └── ...

The Apollo Dataset

The Apollo dataset can be downloaded from site and extracted in the following structure:

./liflow/
└── Datasets/
    └── LiDAR-MOS
          └── sequences
            ├── 00/	# for validation
            │   ├── velodyne/
            |   |       ├── 000000.bin
            |   |       ├── 000001.bin
            |   |       └── ...
            │   └── labels/
            |       ├── 000000.label
            |       ├── 000001.label
            |       └── ...
            └──  04/
                 └── ...

Ground truth generation

To generate the ground complete scenes you can run the map_from_scans.py script. This will use the dataset scans and poses to generate the sequence map to be used as ground truth during training:

python utils/map_from_scans.py --path ./Datasets/SemanticKITTI/dataset/sequences

Once the sequences map is generated you can then train the model

Training the LiFlow model

For training the LiFlow model, the configurations are defined in config/config_flow.yaml, and the training can be started with:

python train.py

Evaluate Flow Scene Completion

For running the scene completion evaluation:

python utils/eval_path.py --path path/to/data --flow flow_ckpt --refine refine_ckpt

For generating the scene completion point clouds:

python utils/flow_completion_pipeline.py --path path/to/data --flow flow_ckpt --refine refine_ckpt

Pre-trained weights

Pre-trained weights are avilable in site

References

The refinement network is provided from site

Citation

@inproceedings{matteazzi2026liflow,
  title={Liflow: Flow matching for 3d lidar scene completion},
  author={Matteazzi, Andrea and Tutsch, Dietmar},
  booktitle={European Conference on Computer Vision},
  pages={130--144},
  year={2026},
  organization={Springer}
}