MAD-ICP

It Is All About Matching Data -- Robust and Informed LiDAR Odometry

Accepted RA-L 2024

Preprint

Install using pip

You can download/install MAD-ICP using pip

pip install mad-icp

Usage

We provide a Python launcher for Rosbag1, Rosbag2, and KITTI binary formats. The dataset configuration is important for the sensor characteristics and extrinsic information (typically, ground truths are not expressed in the LiDAR frame). In configurations/datasets/dataset_configurations.py we provide configurations for many datasets.

The internal parameters (used by default) are stored in configurations/mad_params.py. All the experiments have been run with this same set. You can specify a new set in configurations/mad_params.py and use it with the option --mad-icp-params.

Both the options --dataset-config and --mad-icp-params also accept .cfg files like those in configurations.

To run the pipeline, choose the appropriate dataset configuration (kitti for this example) and type:

mad_icp --data-path /input_dir/ \
        --estimate-path /output_dir/ \
        --dataset-config kitti

Our runner directly saves the odometry estimate file in KITTI format (homogenous matrix row-major 12 scalars); soon, we will provide more available formats like TUM.

Our pipeline is anytime realtime! You can play with parameters num_keyframes and num_cores and, if you have enough computation capacity, we suggest increasing these (we run demo/experiments with num_keyframes=16 and num_cores=16).

Data associtation and registration tools

If you want to use our MAD-tree to perform nearest neighbor or use MAD-ICP to perform registration between two point clouds, here few easy examples.

---

Building from source

Building is tested by our CI/CD pipeline for Ubuntu 20.04 and Ubuntu 22.04 (using g++).

The following external dependencies are required.

Dependency	Version(s) known to work
Eigen	3.3
OpenMP
pybind11
yaml (optional for C++ apps)

If your system lacks any dependency (except for OpenMP) we download local copies using FetchContent. If you want to build and install the package, assuming you're inside the repository, you can use pip as follows:

pip install .

Moreover, you can build the C++ library (along with the pybinds) by typing:

mkdir build && cd build && cmake ../mad_icp && make -j

Building and Running C++ Apps [Optional]

If you want to avoid Python, we provide the bin_runner C++ executable (located in mad_icp/apps/cpp_runners/bin_runner.cpp) that accepts binary cloud format (KITTI, Mulran, etc.). You can build the executable using

mkdir build && cd build && cmake -DCOMPILE_CPP_APPS=ON ../mad_icp && make -j

And run

cd build/apps/cpp_runners
./bin_runner -data_path /path_to_bag_folder/ \
             -estimate_path /path_to_estimate_folder/ \
             -dataset_config ../../../mad_icp/configurations/datasets/kitti.cfg \
             -mad_icp_config ../../../mad_icp/configurations/params.cfg 

Important

If running on the KITTI dataset, enable the flag -kitti for KITTI scan correction (not documented anywhere). We do not (currently) provide a viewer for this executable.

What is missing?

• ROS/ROS2 optional dependencies

Cite us

If you use any of this code, please cite our paper:

@article{ferrari2024mad,
  title={MAD-ICP: It Is All About Matching Data--Robust and Informed LiDAR Odometry},
  author={Ferrari, Simone and Di Giammarino, Luca and Brizi, Leonardo and Grisetti, Giorgio},
  journal={IEEE Robotics and Automation Letters},
  year={2024},
  doi={10.1109/LRA.2024.3456509}
}