MAD-ICP

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

Preprint

You can download/install MAD-ICP using pip

pip install mad-icp

We provide a Python launcher for Rosbag1, Rosbag2, and KITTI binary formats. The dataset configuration file is important for the sensor characteristics and extrinsic information (typically, ground truths are not expressed in the LiDAR frame).

The internal parameters are in configurations/params.cfg (at the moment if you install this via pip you need to clone the repo). All the experiments have been run with this same set. To run the pipeline, choose the appropriate dataset configuration file (kitti.cfg for this example) and type:

cd mad-icp/
mad_icp --data-path /input_dir/ \
        --estimate-path /output_dir/ \
        --dataset-config mad-icp/configurations/datasets/kitti.cfg \
        --mad-icp-config mad-icp/configurations/params.cfg

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).

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

The following external dependencies are required.

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 -j8

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 

• ROS/ROS2 optional dependencies

Our work is currently under review; in the meantime, if you use any of this code, here is the preprint information:

@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={arXiv preprint arXiv:2405.05828},
  year={2024}
}