Splat-LOAM: Gaussian Splatting LiDAR Odometry and Mapping

Emanuele Giacomini · Luca Di Giammarino · Lorenzo De Rebotti · Giorgio Grisetti · Martin R. Oswald

Accepted ICCV 2025

Quickstart

Installation

Prerequisites

First, download this repository and all its submodules:

git clone --recursive https://github.com/rvp-group/Splat-LOAM.git

To run Splat-LOAM, you strictly need an NVIDIA GPU with CUDA capabilities. The implementation has been tested with CUDA versions 11.8 and 12.6, however, other versions should be compatible

Moreover, depending on how you intend to run the implementation, you might need:

• Docker and the NVIDIA Container Toolkit
• Pixi
• Anaconda

1. Recommended (Docker)

The Docker configuration can be used both for execution and for development and its definitely the less-painful way to setup your machine.

To build the Docker image, simply run

./docker/build.sh

Once installed, we can run the container:

./docker/run.sh . <YOUR_DATA_FOLDER>

The first argument should point to the root project directory while the second argument should be the absolute path to your data.

While in the container, everything you need will be found in the /workspace folder. Specifically:

• /workspace/repository/ is a shared folder that points to the host repository (changes here are preserved)
• /workspace/data/ is a shared folder that points to<YOUR_DATA_FOLDER>

For who's interested in development: you can change code in the host repository without restarting the container which is quite handy. We still need a way to make LSP work, if you have any tips, get in contact with us!

2. Pixi

Note

Pixi configuration uses CUDA 12.6 while the Docker configuration uses CUDA 11.8. It shouldn't create any issues but its worth noticing this.

Assuming that you already installed Pixi, simply run

pixi install

Once the environment is setup, we suggest starting a shell in the Pixi environment to continue forward:

pixi shell

3. Anaconda

Thanks to @KwanWaiPang for the contribution! Create a new environment from the environment file and activate it:

conda env create -f environment.yaml
conda activate Splat-LOAM

Run Splat-LOAM

The main entrypoint for Splat-LOAM is the run.py script. It provides the following commands:

• slam: Run SLAM mode from a dataset
• mesh: Extract a mesh from the SLAM output
• eval_odom: Evaluate the RPE of an estimated trajectory against the reference.
• eval_recon: Evaluate the 3D Reconstruction metrics (Accuracy, Completeness, Charmfer-L1, F1-score) of one or more estimated mesh against the reference point cloud.
• crop_recon: Crop the reference point cloud given a set of estimated meshes. Useful for multi-approach evaluation
• generate_dummy_cfg: Generate a configuration file in yaml format with default parameters setup

To run the slam application, we first need a configuration file!

1. Configuration

Tip

TL;DR

The configs/ folder contains some configuration files that you can use to start.

We are still working on porting the configurations to the repository. Odometry configurations in car-like scenarios are still not working correctly!

Just remember to change the data paths!

[Details (click to expand)]

A configuration is a YAML file that contains several info required for the SLAM application to run correctly. From a high level perspective, it is composed of the following components:

Configuration:
  inherit_from : str  # base configuration path (default to null)
  data: DatasetConfig # Input data, format, location etc
  preprocessing: PreprocessingConfig # From PointCloud to Images
  output: OutputConfig # Output data, format, location etc
  logging: LoggingConfig # Logging backends
  mapping: MappingConfig # How Gaussians are updated
  tracking: TrackingConfig # How new poses are estimated
  opt: OptimizationConfig # GS learning rates

You don't need to write the full configuration unless you want to change everything. If you wish, you can also start from another configuration by filling the inherit_from parameter:

inherit_from: configs/a_cool_config.cfg # paths can be relative from project root directory

If you're reading this, it's likely that you want to pass your data to Splat-LOAM so let's see how you can do it.

First, in our context, an input sequence should provide:

1. Point Clouds (timestamped)
2. (Optionally) Poses (also timestamped)

To maximize compatibility with different datasets, we rely on three main abstraction entities to parse input data:

• PointCloudReader : Given data in any format, provides <Point Cloud, timestamp> elements
• TrajectoryReader : Given a trajectory in any format, provide f(Point Cloud, timestamp) -> pose
• DatasetReader : Given data in any format, provides iterable for <Point Cloud, timestamp, pose>

Since publicly available datasets follows more or less similar patterns and formats, once the DatasetReader object is informed of the dataset type, it will setup both PointCloudReader and TrajectoryReader to handle the underlying data.

Here we show the current set of supported dataset formats:

DatasetReader	Notes	Cloud Format	Trajectory Reader
vbr	Vision Benchmark in Rome	rosbag	tum
kitti		bin	kitti
ncd	Newer College Dataset	rosbag	tum
oxspires	Oxford Spires	rosbag	tum
oxpires-vilens	w/ VILENS trajectory	pcd	vilens
generic	Customizable	bin|ply|pcd|rosbag	kitti | tum | vilens

Suppose I want to set up a configuration to read the popular quad-easy sequence of NCD. I'd simply write over a new configuration file:

data:
  dataset_type: ncd
  cloud_reader:
    cloud_folder: /path/to/ncd.bag
  trajectory_reader:
    filename: /path/to/trajectory.tum

And it's done. As long as the data remains consistent, the DatasetReader will handle the extra parameters (extrinsics, rosbag topics, formats, etc).

GT trajectory is optional unless tracking.method=gt is set. If available, it's used for initial guess alignment only and for evaluation metadata retrival.

More details on what each DatasetReader does, can be found in scene.dataset_reader.py

More details will be provided in the documentation (WIP)

Running SLAM

To launch the SLAM application, run:

  python3 run.py slam <path/to/config.yaml>

If output.folder is not set, the experiment results will be placed in results/<date_of_the_experiment>/.

Tip

If you want to solve Mapping-only, provide a trajectory in data.trajectory_reader.filename, set tracking to use it with tracking.method=gt and enable skipping of clouds that have no associated pose with data.skip_clouds_wno_sync=true

[Details (click to expand)]

Logging

We use two logging systems within Splat-LOAM. For events, we use the logging module. If you want to enable debug logs, include the --verbose / -v argument. For data, we rely on rerun. Specifically, at each frame, we log rasterized depths and normals and depth_L1 error map. We additionally log the densification mask and the current Gaussian model state.

You can customize the behavior of rerun through the configuration file:

logging:
  enable: true/false # Enable data logging

  EITHER
  rerun_spawn: true/false/none # Spawn GUI and binds to it <use if monitor is available>
  OR
  rerun_serve_grpc: true/false/none # serve log-data over gRPC <use for remote connections>
  OR
  rerun_connect_grpc_url: <str> # serve log-data over gRPC to an already instantiated viewer.

The Gaussian model observed in rerun is not rendered with the 2DGS rasterizer. Don't worry if it looks different from the rasterized images.

Mesh generation

After running SLAM, you can generate a mesh representation by running:

python3 run.py mesh <path/to/result/folder>

Important

The default behavior of the mesh application is a trade-off between quality and speed. Checkout the documentation to tune the meshing process.

Evaluation

3D Reconstruction

To evaluate the computed mesh, run the following:

python3 run.py eval_recon <reference_pointcloud_file> <estimate_mesh> 

By default, the script will provide the results both on the terminal and on the file eval_recon_<date_of_exp> located in the working directory.

Important

If the evaluation is needed for comparison with other methods, generate the cropped reference point cloud to ensure a fair comparison. This is achieved by running

python3 run.py crop_recon <reference_pointcloud_file> <mesh_1> <mesh_2> ... <mesh_n>

Odometry

If ground truth trajectory was provided in the experiment's configuration file, then you can evaluate the odometry estimate by running:

python3 run.py eval_odom <path/to/result/folder>

If no trajectory was provided, then a few more arguments are required:

python3 run.py eval_odom <path/to/odom/estimate> \
                          --reference-filename <path/to/reference/trajectory> \
                          --reference-format <tum|kitti|vilens> \
                          --estimate-format <tum|kitti|vilens> \

Note

If reference is expressed in KITTI format, you also need to provide the argument:

--kitti-timestamps <path/to/kitti/times.txt>

Additionally, for KITTI sequences, we do not evaluate if number of poses between reference and estimate doesn't match.

Citation

If you use Splat-LOAM for your research, please cite (currently in preprint):

@misc{giacomini2025splatloam,
      title={Splat-LOAM: Gaussian Splatting LiDAR Odometry and Mapping}, 
      author={Emanuele Giacomini and Luca Di Giammarino and Lorenzo De Rebotti and Giorgio Grisetti and Martin R. Oswald},
      year={2025},
      eprint={2503.17491},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2503.17491}, 
}

Contacts

If you have questions or suggestions, we invite you to open an issue! Additionally, you can contact:

• Emanuele Giacomini : [email protected]