Ghost-FWL: A Large-Scale Full-Waveform LiDAR Dataset for Ghost Detection and Removal

Kazuma Ikeda^1*   Ryosei Hara^1*  

Rokuto Nagata¹   Ozora Sako¹   Zihao Ding¹   Takahiro Kado²   Ibuki Fujioka²   Taro Beppu²  

Mariko Isogawa¹   Kentaro Yoshioka¹

¹ Keio University
² Sony Semiconductor Solution
^* Equal contribution

CVPR 2026

This is the official implementation of Ghost-FWL

Installation

• requirements
  • install uv
  • create Weights and Biases account
• clone the repositoryW

  git clone git@github.com:Keio-CSG/Ghost-FWL.git

• init (only after cloning the repository)

  uv sync
  
  # (optional)
  uv run pre-commit install
  uv run pre-commit autoupdate

Dataset

See README_dataset.md for more details.

Pretrain

uv run python scripts/run_train.py --config configs/config_pretrain.yaml

Train

uv run python scripts/run_train.py --config configs/config_train.yaml

Estimate

uv run python scripts/run_estimate.py --config configs/config_estimate.yaml

Test

Test Recall

uv run python scripts/run_test.py --config configs/config_test.yaml

Test Ghost Removal Rate

uv run python src/visualize/evaluate_pcd_batch.py --config src/visualize/configs/evaluate_pcd_batch.yaml

Visualize

• vis_pred.py
  • visualize prediction results, ground truth annotations, and temporal histogram at peak locations (matplotlib)

uv run python src/visualize/vis_pred.py --config configs/config_test.yaml

• vis_pcd.py
  • save .pcd file from estimated results or ground truth annotations

uv run python src/visualize/vis_pcd.py --config src/visualize/configs/vis_pcd.yaml

• vis_pcd_batch.py
  • batch save .pcd file from estimated results or ground truth annotations

uv run python src/visualize/vis_pcd_batch.py --config src/visualize/configs/vis_pcd_batch.yaml

• interactive_histogram_viewer.py
  • visualize intensity map and histogram at the clicked location

uv run python src/visualize/interactive_histogram_viewer.py /path/to/voxel.b2 /path/to/{prediction,annotation}.b2

Citation

@inproceedings{ikeda2026ghostfwl,
  title = {Ghost-FWL: A Large-Scale Full-Waveform LiDAR Dataset for Ghost Detection and Removal},
  author = {Ikeda, Kazuma and Hara, Ryosei and Nagata, Rokuto and Sako, Ozora and Ding, Zihao and Kado, Takahiro and Fujioka, Ibuki and Beppu, Taro and Isogawa, Mariko and Yoshioka, Kentaro},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2026},
}

Code Reference

The primary references used for the implementation are listed below. Please refer to the original papers for all citations.

• Lidar Waveforms are Worth 40x128x33 Words
• MARMOT: Masked Autoencoder for Modeling Transient Imaging
• VideoMAE

Contact

If you have any questions, please post an issue and mention @ike-kazu and @ryhara.