unzip, then copy each file to:

May 22, 2026 · View on GitHub

ERASOR2

Video • Install • How to Run • Paper • Contact Us

ERASOR2 demo
ERASOR2 comparison

ROS-free, instance-aware static map building

:package: Installation

# 1. Build (one cmake call, no ROS/catkin).
cmake -B build -S . && cmake --build build -j

# 2. Conda env for the Python preprocessors + evaluator.
conda env create -f scripts/environment.yml   # creates env "erasor2"
conda activate erasor2

See USAGE.md for the full dependency list and per-distro notes.

SemanticKITTI Setup

Download SemanticKITTI so the sequence folders live under <kitti_dir>/dataset/sequences. For example, the benchmark tree should look like this:

<kitti_dir>/                         # e.g., /home/<user id>/datasets/kitti
└── dataset/
    ├── poses/
    └── sequences/
        ├── 00/
        │   ├── velodyne/
        │   ├── labels/
        │   ├── **poses_suma_optim.txt** (important)
        │   └── times.txt
        ├── 01/
        ├── 02/
        ├── ...
        └── 10/

ERASOR2 uses SuMa poses for evaluation. Download the pose archive and place each poses_suma_optim.txt inside its matching sequence directory:

wget -O suma_poses_for_erasor_eval.zip "https://www.dropbox.com/scl/fi/9q3b1b9npsst1zjawgou3/suma_poses_for_erasor_eval.zip?rlkey=vx4igm68iuo3eobpolgq4tblg&st=yt1ola9b&dl=0"
# unzip, then copy each file to:
# <kitti_dir>/dataset/sequences/<seq>/poses_suma_optim.txt

For each benchmark config in config/erasor2/seq_{00,01,02,05,07}.yaml, set dataloader.abs_data_dir to <kitti_dir>/dataset/sequences and dataloader.abs_save_dir to your ERASOR2 output directory.

:rocket: How to Run

# 3. Generate per-frame Patchwork ground + HDBSCAN instance labels
#    for seqs 00, 01, 02, 05, 07 in one shot.
scripts/generate_labels.sh /path/to/kitti

# 4. Edit config/erasor2/seq_{00,01,02,05,07}.yaml to point at your
#    kitti and output directories, then run the full benchmark.
python scripts/run_benchmark.py

scripts/run_benchmark.py invokes run_pipeline.py for each yaml (mapgen → run_erasor2 → evaluate.py), then prints a single consolidated PR / RR / F1 table. See USAGE.md for further explanation — per-step breakdown, path-editing conventions, visualizer, YAML reference, and HeLiPR / HeLiMOS setup.

:bar_chart: Headline numbers

Some reproduced numbers may differ slightly from the paper after the ROS-free refactor, but the overall performance remains consistent with the reported HDBSCAN-based results. Because this implementation uses HDBSCAN for instance segmentation, compare against the HDBSCAN rows in Table III of the paper.

Seq	Frames	PR [%] ( $\color{#c026d3}\textsf{paper}$ / $\color{#0969da}\textsf{ours}$ )	RR [%] ( $\color{#c026d3}\textsf{paper}$ / $\color{#0969da}\textsf{ours}$ )	F1 ( $\color{#c026d3}\textsf{paper}$ / $\color{#0969da}\textsf{ours}$ )
00	4390 – 4530	$\color{#c026d3}98.649$ / $\color{#0969da}\mathbf{98.654}$	$\color{#c026d3}\mathbf{98.582}$ / $\color{#0969da}98.454$	$\color{#c026d3}\mathbf{0.986}$ / $\color{#0969da}0.9855$
01	150 – 250	$\color{#c026d3}93.554$ / $\color{#0969da}\mathbf{95.743}$	$\color{#c026d3}\mathbf{94.951}$ / $\color{#0969da}94.027$	$\color{#c026d3}0.943$ / $\color{#0969da}\mathbf{0.9488}$
02	860 – 950	$\color{#c026d3}98.339$ / $\color{#0969da}\mathbf{99.196}$	$\color{#c026d3}99.709$ / $\color{#0969da}\mathbf{99.902}$	$\color{#c026d3}0.990$ / $\color{#0969da}\mathbf{0.9955}$
05	2350 – 2670	$\color{#c026d3}97.473$ / $\color{#0969da}\mathbf{97.670}$	$\color{#c026d3}\mathbf{99.113}$ / $\color{#0969da}98.412$	$\color{#c026d3}\mathbf{0.983}$ / $\color{#0969da}0.9804$
07	630 – 820	$\color{#c026d3}\mathbf{98.767}$ / $\color{#0969da}96.135$	$\color{#c026d3}98.800$ / $\color{#0969da}\mathbf{98.989}$	$\color{#c026d3}\mathbf{0.988}$ / $\color{#0969da}0.9754$

_{$\color{#c026d3}\textsf{Magenta}$ = paper (Table III, HDBSCAN row), $\color{#0969da}\textsf{blue}$ = our re-run. Bold marks the higher value per cell.}

ERASOR2 reproduces within run-to-run noise (mean |ΔF1| = 0.006). Higher is better on all three metrics:

PR (Preservation Rate) measures how much true static structure remains after dynamic-object removal.
RR (Rejection Rate) measures how much dynamic structure is correctly rejected from the static map.
F1 is the harmonic mean of PR and RR, giving one balanced score when preservation and rejection both matter.

:books: Citation

If you use this code in academic work, please cite the ERASOR / ERASOR2 papers.

@article{lim2025erasor2,
  title   = {{ERASOR2}: Instance-Aware Robust 3D Mapping of the Static World in Dynamic Scenes},
  author  = {Lim, Hyungtae and others},
  journal = {IEEE Robotics and Automation Letters},
  year    = {2025}
}

@article{lim2021erasor,
  title   = {{ERASOR}: Egocentric Ratio of Pseudo Occupancy-based Dynamic Object Removal for Static 3D Point Cloud Map Building},
  author  = {Lim, Hyungtae and Hwang, Sungwon and Myung, Hyun},
  journal = {IEEE Robotics and Automation Letters},
  volume  = {6},
  number  = {2},
  pages   = {2272--2279},
  year    = {2021}
}

@inproceedings{lim2024helimos,
  title     = {{HeLiMOS: A dataset for moving object segmentation in 3D point clouds from heterogeneous LiDAR sensors}},
  author    = {Lim, Hyungtae and Jang, Seoyeon and Mersch, Benedikt and Behley, Jens and Myung, Hyun and Stachniss, Cyrill},
  booktitle = {2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages     = {14087--14094},
  year      = {2024}
}