LiCo-Film: Standardized Film-Based LiDAR Cover Contamination Dataset

The release includes the approved BSD-3-Clause license documentation in LICENCE.txt.

LiCo-Film is a dataset and benchmark for LiDAR cover contamination under dynamic driving conditions. It uses standardized transparent contamination films to create reproducible dry and wet proxy-contamination states and separate winter-road recordings to evaluate transfer to naturally accumulated dry real contamination.

The dataset supports contamination detection, contamination-type classification, fine-grained standardized-film state classification, and held-out binary transfer evaluation on dry real-contamination recordings.

Dataset Summary

LiDAR cover contamination can reduce valid returns, distort reflectivity, and affect cleaning decisions before downstream perception fully fails. LiCo-Film addresses this by combining controlled contamination films with dynamic driving recordings and fixed benchmark splits.

The standardized-film part contains 18 contamination states plus a clean reference:

• 12 dry films: opacity groups 40%, 70%, and 100%, each with four coverage levels.
• 6 wet films: three droplet-size groups, each with two coverage levels.
• 1 clean reference class.

The real-contamination part contains 15 dry real-contamination films collected during winter-road test drives in Finland, Sweden, and Norway. These real recordings are reserved for held-out binary transfer evaluation and do not have fine-grained physical severity labels in the current release.

Data Collection

All recordings use an Innoviz Two dTOF LiDAR sensor with 905 nm wavelength, 240 m range, 120 x 28.8 degree horizontal/vertical field of view, 20 Hz frame rate, and 4.3 million points per second. The sensor was mounted on the vehicle roof at approximately 2 m height.

The standardized training recordings were collected on routes in and around Wolfsburg, Germany, covering nine traffic environments such as urban roads, rural roads, highways, industrial areas, bridges, parking lots, and construction zones. The held-out standardized and real-contamination test recordings were collected on a spatially separate route with industrial, rural, and residential environments.

Training V1 and Training V2 also differ in the LiDAR processing chain. Training V1 was recorded with an ADAS blade perception-module setup, whereas Training V2 and the test recordings used the Innoviz iPEM. This affects the encoding of confidence-zero returns: Training V1 contains more automatically generated all-zero rows, while Training V2 may retain coordinates for points with confidence value 0. The official LOEO benchmark is fixed to Training V2, and the baseline scripts compute valid-point features after confidence filtering.

Dataset Contents

The binary public release contains:

• 447 scenes
• 439,967 binary LiDAR frames
• 689.2 GiB uncompressed binary frame data
• 11 compressed tar.zst data shards, approximately 344.2 GiB total
• per-scene scene.json metadata
• release-level metadata tables
• official LOEO split definitions and an additional exploratory all-release environment split
• checksum manifests
• scripts for reading frames, exporting frames to CSV, sample subset creation, feature extraction, and kNN/SVM baseline runs

Compressed data shards are stored as:

data/lico_film_data_001.tar.zst
...
data/lico_film_data_011.tar.zst
data/archive_manifest.csv
data/archive_plan.csv
data/archive_plan_scenes.csv

After extraction, the expected release root is:

LiCo-Film/
  data/
    01_ML_Training_V1_B2p0_ADAS_Pos1_Stand_Film_bin/
    02_ML_Training_V2_B2p0_iPEM_P2_Stand_Film_bin/
    03_ML_Test_Wob_B2p0_iPEM_P2_Stand_Film_bin/
    04_ML_Test_Wob_B2p0_iPEM_P2_Real_Film_bin/
  metadata/
  splits/
    environment_split_all_released_scenes.json
    LOEO_PROTOCOL.md
    loeo_protocol_folds.json
  logs/
  scripts/
  docs/
    RELEASE_README.md
  manifest.csv
  README.md
  LICENCE.txt
  artifact_metadata.json
  requirements.txt
  TEST_COMMANDS.txt
  UNPACK_AND_TEST.txt

The four folders below data/ are created automatically when the archive shards are extracted. Their names are also listed in metadata/scenes.csv under the dataset_folder column and in the split metadata where applicable.

Each scene is stored as:

data/<dataset_folder>/<scene_id>/
  frames/
    frame_000001.bin
    frame_000002.bin
    ...
  scene.json

Binary Frame Format

Each .bin frame stores raw little-endian fixed-width point records. There is no file header.

Each point record uses 8 bytes:

Field	Type	Scale
x_cm	signed int16	meters = x_cm / 100
y_cm	signed int16	meters = y_cm / 100
z_cm	signed int16	meters = z_cm / 100
reflectivity	unsigned int8	original integer value
confidence	unsigned int8	original integer value

Equivalent Python struct format:

POINT = struct.Struct("<hhhBB")

The original CSV Distance column is not stored and can be reconstructed from X, Y, and Z.

Unpacking

Extract all data shards into the same target folder. Do not extract each shard into a separate subfolder.

PowerShell example:

mkdir D:\LiCo-Film

tar -xf lico_film_release_metadata_scripts_20260710.tar.zst -C D:\LiCo-Film

Get-ChildItem .\lico_film_data_*.tar.zst | Sort-Object Name | ForEach-Object {
    tar -xf $_.FullName -C D:\LiCo-Film
}

cd /d D:\LiCo-Film
py -3.11 scripts\run_smoke_test.py --release-root . --out outputs\smoke_test

Full checksum verification:

cd /d D:\LiCo-Film
py -3.11 scripts\verify_checksums.py --root .

Expected final output:

all checksums match

Tasks and Splits

The benchmark defines three standardized-film tasks:

• Binary: clean vs. contaminated.
• Type: clean vs. dry-contaminated vs. wet-contaminated.
• Levels: 19-class classification of clean and standardized contamination states.

The primary benchmark protocol is an environment-disjoint leave-one-environment-out (LOEO) evaluation. Each fold holds out one route environment for testing and one disjoint environment for validation. The remaining environments are used for training.

The official LOEO benchmark protocol is defined on 02_ML_Training_V2_B2p0_iPEM_P2_Stand_Film_bin only, because this recording variant provides the complete nine-environment standardized-film matrix used for the paper benchmark. 01_ML_Training_V1_B2p0_ADAS_Pos1_Stand_Film_bin is included as additional released data, but it is not part of the official LOEO benchmark split because its rural and industrial route parts were not released as separately defined V1 environments.

The held-out real-contamination evaluation is restricted to binary clean-versus-contaminated classification. No real-contamination frames are used for training, validation, or model selection.

Split definitions are provided in:

splits/loeo_protocol_folds.json
splits/LOEO_PROTOCOL.md

An additional exploratory split over all released scenes is provided in:

splits/environment_split_all_released_scenes.json

This file is intended for follow-up analyses and is not the official ETFA benchmark protocol.

Baseline Results

The accompanying paper reports lightweight frame-level baselines using a 15-dimensional feature representation and classical kNN/SVM classifiers.

Environment-disjoint standardized-film LOEO macro-F1:

Model	Binary	Type	Levels
kNN	0.826 +/- 0.184	0.686 +/- 0.144	0.274 +/- 0.046
SVM	0.842 +/- 0.202	0.710 +/- 0.161	0.375 +/- 0.071

Held-out binary real-contamination transfer:

Model	Pooled macro-F1	Clean recall	Contaminated recall
kNN	0.918	0.928	0.982
SVM	0.839	0.972	0.947

These baselines establish initial reference performance. They are not intended as upper bounds on the dataset.

Project Context

LiCo-Film was developed as a research dataset in the context of the first author's doctoral research and industrial automotive research activities at Volkswagen AG. The dataset is released to support transparency around the reported research results and to facilitate further research on LiDAR cover contamination, sensor monitoring, and robust perception.

Limitations

• The current real-contamination collection covers dry contamination only.
• Real-contamination recordings do not include calibrated physical severity labels.
• A clean carrier-film control is not included, so carrier-film effects cannot be separated from weak real-contamination effects in the current release.
• The fine-grained Levels task is defined for standardized films, not for real contamination.
• Frame counts describe processing scale. Adjacent frames in a scene are temporally correlated and should not be treated as independent recordings.

Research Artifact Notice

The accompanying code and scripts are provided as research utilities and proof-of-concept examples for reading the release, creating small subsets, extracting simple frame-level features, and running lightweight baseline experiments. They are not part of, and are not deployed in, any Volkswagen product.

Unless stated otherwise, this repository should be treated as a research artifact and not as a maintained software product. Users are responsible for validating the data, scripts, and results for their own research or application context. This notice is intended to set appropriate expectations for research code and does not replace the terms of the applicable license and notice files.

Versioning Note

The official ETFA benchmark split in this release is fixed to Training V2. The additional Training V1 data are included as curated supplementary standardized training data, including the pre-release V1 route split and completion patch. Any later changes to the released data should be published as a new dataset version rather than modifying this release in place.

Citation

The dataset accompanies the paper:

@inproceedings{grimminger2026licofilm,
  title = {LiCo-Film: A Standardized Film-Based Dataset and Benchmark for LiDAR Cover Contamination},
  author = {Grimminger, Jannis and Breitner, Felix and Doshi, Jai Sandeep and Windmann, Alexander and Niggemann, Oliver},
  booktitle = {Proceedings of the IEEE International Conference on Emerging Technologies and Factory Automation (ETFA)},
  year = {2026},
  note = {Accepted, final bibliographic details pending}
}

License

The release is provided under the BSD-3-Clause "New" or "Revised" License. See LICENCE.txt for the approved license documentation and third-party component notices.

Contact

For questions about the dataset and benchmark, contact the dataset maintainer listed in the accompanying paper or repository metadata.