Ghost-FWL is released under CC BY-NC 4.0. By requesting access you agree to use the dataset for non-commercial research or educational purposes only, to not redistribute the raw data, and to cite the Ghost-FWL paper in any resulting publication.

Ghost-FWL

Full-waveform LiDAR voxel dataset for ghost point detection, distributed as WebDataset shards.

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},
}

Configs

`ghost` — annotated dataset for ghost detection

One sample per frame:

member	content
`<key>.voxel.b2`	voxel grid, Blosc2-packed `(400, 512, 700)` array
`<key>.annotation.b2`	annotation voxel (`annotation_v1`), same shape
`<key>.annotation_expand.b2`	expanded annotation (`annotation_v1_expand`), when available
`<key>.json`	`frame_id`, `scene_id`, `hist_id`, `annotation_version`, `has_annotation_expand`

`mae` — dataset for MAE pretraining

One sample per frame:

member	content
`<key>.voxel.b2`	voxel grid, Blosc2-packed `(400, 512, 700)` array
`<key>.peaks.npy`	per-pixel peak list, `np.load(..., allow_pickle=True)`, `(204800, 3)`
`<key>.json`	`frame_id`, `category` (`"ghost"` or `"normal"`), `session`

Loading

With the webdataset package (recommended for training):

import io
import json

import blosc2
import numpy as np
import webdataset as wds
from huggingface_hub import HfApi, get_token

repo = "ryhara/Ghost-FWL"
files = [f for f in HfApi().list_repo_files(repo, repo_type="dataset") if f.endswith(".tar")]
urls = [f"https://huggingface.co/datasets/{repo}/resolve/main/{f}" for f in files if f.startswith("ghost/")]

dataset = wds.WebDataset(urls, shardshuffle=True).shuffle(100)
for sample in dataset:
    metadata = json.loads(sample["json"])
    voxel = blosc2.unpack_array2(sample["voxel.b2"])         # (400, 512, 700)
    annotation = blosc2.unpack_array2(sample["annotation.b2"])
    break

With datasets:

from datasets import load_dataset

dataset = load_dataset("ryhara/Ghost-FWL", "ghost", streaming=True)

Only the ghost config is registered for datasets / the Dataset Viewer. The mae shards contain <key>.peaks.npy object arrays, which datasets refuses to decode (allow_pickle=False), so read mae/*.tar with webdataset directly (above) or with the training-code loader.

Each <config>/manifest.json records the sample count, shard list, and any files excluded during conversion. <config>/shard_index.json maps every shard to its per-group sample counts (scene001/hist003 for ghost, ghost/<session> for mae) so loaders can open only the shards they need.

The Ghost-FWL training code ships loaders that stream these shards with the same preprocessing as the original directory layout (src/data/dataset_fwl_wds.py, src/data/dataset_fwl_mae_wds.py); set wds_root: hf://ryhara/Ghost-FWL in configs/config_*_wds.yaml.

Downloads last month: 64

Total file size:

736 GB

Paper for ryhara/Ghost-FWL

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

Paper • 2603.28224 • Published Mar 30 • 5

Request access to Ghost-FWL