ELite: Ephemerality meets LiDAR-based Lifelong Mapping

Hyeonjae Gil* · Dongjae Lee* · Giseop Kim · Ayoung Kim

(* Equal Contribution)

Arxiv | Video

About ELite

ELite is a LiDAR-based lifelong mapping framework that integrates ephemerality—the likelihood of a point being transient or persistent—into the entire pipeline. Unlike binary classifications of map elements (static vs. dynamic), ELite introduces a two-stage ephemerality concept to represent changes across different time scales:

• Local ephemerality: Captures short-term variations (e.g., moving cars vs. parked cars).
• Global ephemerality: Represents long-term changes (e.g., parked cars vs. new buildings)

By leveraging ephemerality, ELite seamlessly aligns multiple sessions, removes dynamic objects, and matains accurate maps.

Installation

1. Clone the repository

git clone git@github.com:dongjae0107/ELite.git

2. Set up conda environment

We tested with python 3.10, but any python version will work fine.

conda create -n elite python=3.10
conda activate elite

3. Install requirements

cd ELite
pip install -e .

Optional: For CUDA-accelerated matching, we recommend installing pygicp and using the PyGICPScanMatcher. Follow the instructions below or refer to the fast_gicp repository for more details.

git clone git@github.com:koide3/fast_gicp.git
cd fast_gicp
python3 setup.py install --user

Run

Example - Parkinglot dataset

Download the dataset from this link or use the script below to download example sequences (01 and 02):

bash scripts/download_parkinglot.sh

To download additional sequences, simply modify the script as needed.

Then, initialize the ELite with first session:

python3 run_elite.py ./config/parkinglot_first.yaml

For the other sessions, run:

python3 run_elite.py ./config/parkinglot.yaml

Run with Your Own Dataset

Data preparation

You need to provide a Session as input. Each session must include:

• A Scans/ folder containing point cloud files (in .pcd format), starting from 000000.pcd
• A poses.txt file containing LiDAR poses in KITTI format

The expected directory structure is:

Session/
├── poses.txt
└── Scans/
    ├── 000000.pcd
    ├── 000001.pcd
    ├── ...
    └── 00000N.pcd

You can easily generate session data using the saver tool provided with SC-LIO-SAM.
Alternatively, SC-A-LOAM and FAST_LIO_SLAM are also supported.

Initial transform between sessions

To run the alignment module, you need to provide an initial transformation for the new session. Initial transformation is for coarse aligning the first few scans of new session to the previous session's coordinate frame. The simplest method is to use ICP in CloudCompare and use the transformation matrix. In near future, we are planning to automate this process by introducing global localization-based version (e.g. Scan Context) as written in the paper.

Run ELite

Once your dataset is prepared, update the parameters in config/sample.yaml. Make sure to set or remove the prev_outputs_dir field depending on whether this is the first session or a subsequent session.

To run ELite:

python3 run_elite.py ./config/sample.yaml

Citation

If you use ELite for any academic work, please cite our paper.

@INPROCEEDINGS { hjgil-2025-icra,
    author={Hyeonjae Gil and Dongjae Lee and Giseop Kim and Ayoung Kim},
    title={Ephemerality meets LiDAR-based Lifelong Mapping},
    booktitle={Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
    year={2025},
    month={May.},
    address={Atlanta},
}

Contact

If you have any questions, please contact:

• Hyeonjae Gil (now9728@gmail.com)
• Dongjae Lee (dongjae0107@gmail.com)

Acknowledgement

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)(No. RS-2024-00461409), and in part by the Robotics and AI (RAI) Institute.