TLC-Calib Processed Dataset

This repository provides the official processed dataset release for Targetless LiDAR-Camera Calibration with Neural Gaussian Splatting (RA-L 2026).

The release contains five KITTI-360 sequences and three FAST-LIVO2 scenes, organized in a consistent, calibration-ready format. Each released scene includes synchronized camera images, LiDAR poses, calibration parameters, per-frame point clouds, and a scene-level LiDAR map.

For the Waymo Open Dataset, we provide the exact scene and frame configurations used in our experiments, together with preprocessing instructions, rather than redistributing the processed data.

The five-scene KITTI-360 evaluation protocol introduced in TLC-Calib covers straight, rotational, and zigzag motion patterns with varying levels of difficulty. In particular, the small_rotation and large_zigzag sequences broaden the evaluation to include additional rotational and combined-motion conditions.

The released data and preprocessing specifications can be used both to reproduce the TLC-Calib experiments and to evaluate other LiDAR–camera calibration methods under consistent input conditions.

• Paper: https://doi.org/10.1109/LRA.2026.3665066
• Code: https://github.com/SNU-VGILab/TLC-Calib
• Waymo preprocessing: https://github.com/zang09/TLC-Calib_preprocessing

The data files are distributed as scene-level ZIP archives. The Hugging Face Dataset Viewer shows only the scene manifest in metadata/scenes.csv; download the archives as files to run TLC-Calib.

Availability

Source dataset	Released content	Number of scenes	Notes
KITTI-360	Processed scene archives	5	Subject to the original KITTI-360 terms; see LICENSE and THIRD_PARTY_NOTICES.md.
FAST-LIVO2	Processed scene archives	3	Publish only after confirming a dataset redistribution license or written permission from the source-data owner.
Waymo Open Dataset	Configuration and reproduction instructions only	3	No Waymo-derived data is redistributed here.

Repository Layout

TLC-Calib/
├── FAST-LIVO2/
│   ├── README.md
│   ├── Building.zip
│   ├── Landmark.zip
│   └── Sculpture.zip
├── KITTI-360/
│   ├── README.md
│   ├── large_rotation.zip
│   ├── large_zigzag.zip
│   ├── small_rotation.zip
│   ├── small_zigzag.zip
│   └── straight.zip
├── Waymo/
│   └── README.md
├── metadata/
│   └── scenes.csv
├── CONFIG.md
├── LICENSE
├── THIRD_PARTY_NOTICES.md
└── README.md

Download

Install the current Hugging Face Hub client and download the repository:

python -m pip install -U huggingface_hub
hf download b1o1o1m/TLC-Calib \
  --repo-type dataset \
  --local-dir ./TLC-Calib

To download one archive only:

hf download b1o1o1m/TLC-Calib \
  KITTI-360/large_rotation.zip \
  --repo-type dataset \
  --local-dir ./TLC-Calib

Extract the Released Archives

Run the following command from the repository root:

find FAST-LIVO2 KITTI-360 \
  -type f -name '*.zip' \
  -execdir unzip -n '{}' \;

Each archive should contain exactly one top-level scene directory. For example:

KITTI-360/large_rotation.zip
└── large_rotation/
    ├── images/
    ├── lidar/
    ├── params/
    ├── pcds/
    ├── README.md
    └── valid_frame.txt

After extraction, a TLC-Calib scene follows this layout:

<scene_name>/
├── images/
│   ├── image_00/
│   ├── image_01/
│   ├── ...
│   └── image_XX/
├── lidar/
│   ├── map.ply
│   └── rgb_map.ply
├── params/
├── pcds/
├── README.md
└── valid_frame.txt

image_00 to image_XX are placeholders. The number of camera folders depends on the source dataset.

Contents

• images/: synchronized camera images indexed by zero-based local frame ID.
• lidar/: aggregated scene-level LiDAR maps, including rgb_map.ply (colorized) and map.ply (without color).
• params/: camera intrinsics, per-frame poses, camera-to-LiDAR extrinsics, ground-truth and initialization extrinsics, and optional LiDAR timestamps.
• pcds/: per-frame LiDAR point clouds in binary PCD format.
• README.md: scene-specific source and frame-range information.
• valid_frame.txt: mapping from each local frame index to the corresponding original frame ID.

Conventions

• All modalities use the same zero-based local frame index.
• images/image_XX/000123.png and pcds/000123.pcd refer to the same sample.
• Line i in valid_frame.txt, params/lidars.txt, and params/cam*.txt corresponds to local index i.
• If present, line i in params/timestamps.txt is the LiDAR timestamp for local index i.
• PCD files contain x y z intensity fields and are stored in binary format.

Parameter Files

• cam0.txt ... cam*.txt: per-frame camera poses. Each line is a flattened row-major 4 x 4 matrix (1 x 16).
• cam0_to_lidar.txt ... cam*_to_lidar.txt: per-camera transforms. Each file contains one 1 x 17 row: [camera_id, flattened row-major 4 x 4 matrix].
• cams_to_lidar_gt.txt: ground-truth camera-to-LiDAR transforms, one 1 x 17 row per camera.
• cams_to_lidar_init.txt: initialization transforms, one 1 x 17 row per camera.
• intrinsics.txt: camera intrinsics as 3 x 3 matrices grouped by camera.
• lidars.txt: per-frame LiDAR poses, one flattened row-major 4 x 4 matrix per line.
• timestamps.txt: optional LiDAR timestamps, one scalar per local frame.

Scene selections and frame ranges are listed in CONFIG.md and metadata/scenes.csv.

License and Upstream Terms

The repository metadata uses license: other because the repository combines assets derived from multiple sources with different or separately specified terms. Read both of the following before use:

• LICENSE
• THIRD_PARTY_NOTICES.md

The repository-level notice does not replace or override any upstream dataset license, registration requirement, attribution requirement, or use restriction.

Citation

If you use this dataset or its format, cite TLC-Calib:

@article{jung2026targetless,
  title   = {{Targetless LiDAR-Camera Calibration with Neural Gaussian Splatting}},
  author  = {Jung, Haebeom and Kim, Namtae and Kim, Jungwoo and Park, Jaesik},
  journal = {IEEE Robotics and Automation Letters},
  volume  = {11},
  number  = {4},
  pages   = {4777--4784},
  year    = {2026},
  doi     = {10.1109/LRA.2026.3665066}
}

Also cite every upstream dataset whose scenes you use. Required upstream references are listed in THIRD_PARTY_NOTICES.md.