UAVLight: A Benchmark for Illumination-Robust 3D Reconstruction in UAV Scenes

UAVLight is a benchmark for evaluating illumination-robust 3D reconstruction and novel-view synthesis in outdoor UAV scenes. Unlike standard reconstruction datasets that are usually captured under relatively stable lighting, UAVLight focuses on challenging real-world UAV scenarios where the scene appearance changes due to sunlight direction, cast shadows, exposure variation, and outdoor illumination conditions.

The benchmark provides multi-view UAV images, sparse reconstruction files, predefined train/test splits, sun direction annotations, and optional geometry assets. It is designed for studying lighting-aware reconstruction, robust novel-view synthesis, relighting-aware evaluation, and outdoor Gaussian Splatting / NeRF-style scene modeling.

A video preview is available on Hugging Face:

Watch UAVLight video preview

Dataset Download

The UAVLight dataset is hosted on Hugging Face:

Dataset: https://huggingface.co/datasets/dukang92/UAVLight

You can download the full dataset using the Hugging Face CLI:

huggingface-cli download dukang92/UAVLight --repo-type dataset --local-dir UAVLight

Alternatively, you can download selected scene archives directly from the data/ folder on Hugging Face.

Dataset Structure

The Hugging Face dataset repository is organized as:

UAVLight/
  assets/
    UAVLight_teaser.png
    uavlight.mp4

  data/
    <scene_id>.zip
    <scene_id>.zip
    ...

  metadata/
    scenes.csv
    zip_sizes.csv
    file_list.txt
    zip_list.txt
    summary.txt

Each scene is released as a separate zip archive. After extraction, each scene follows the structure:

<scene_id>/
  images/
  sparse/
  dense_points.ply
  downsampled_points.ply
  mesh.ply
  split.csv
  sun_directions.txt
  train_list.txt
  test_list.txt

File Description

• images/: multi-view UAV RGB images.
• sparse/: sparse reconstruction files, such as camera poses and COLMAP-style outputs.
• split.csv: predefined train/test split information.
• train_list.txt: training image list.
• test_list.txt: testing image list for novel-view synthesis evaluation.
• sun_directions.txt: sun direction annotations for illumination-aware reconstruction and analysis.
• dense_points.ply: dense point cloud, provided as an optional geometry asset.
• downsampled_points.ply: downsampled point cloud for lightweight visualization.
• mesh.ply: reconstructed mesh, provided as an optional geometry asset.

Usage

A typical workflow is:

# 1. Download the dataset
huggingface-cli download dukang92/UAVLight --repo-type dataset --local-dir UAVLight

# 2. Extract one scene
unzip UAVLight/data/<scene_id>.zip -d UAVLight/scenes/

Then use:

train_list.txt       for reconstruction / training
test_list.txt        for novel-view synthesis evaluation
sparse/              for camera poses and sparse reconstruction files
sun_directions.txt   for illumination-aware analysis

Intended Use

UAVLight is intended for academic research on:

• illumination-robust 3D reconstruction
• novel-view synthesis under outdoor lighting variation
• UAV-based scene reconstruction
• lighting-aware Gaussian Splatting and NeRF-style methods
• relighting and lighting-transfer evaluation
• robustness analysis under sunlight, shadow, and exposure changes

Limitations

UAVLight focuses on outdoor UAV scenes and illumination robustness. It does not aim to cover all possible outdoor environments, weather conditions, or dynamic scene changes. Geometry assets such as point clouds and meshes are provided as auxiliary reconstruction outputs and should not be treated as perfect ground truth.

Citation

If you use UAVLight in your research, please cite:

@inproceedings{du2026uavlight,
  title     = {UAVLight: A Benchmark for Illumination-Robust 3D Reconstruction in Unmanned Aerial Vehicle (UAV) Scenes},
  author    = {Kang Du and Xue Liao and Junpeng Xia and Chaozheng Guo and Yi Gu and Yirui Guan and Duotun Wang and Sheng Huang and Zeyu Wang},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2026}
}

License

The dataset is released for non-commercial research use. Please refer to the license information on the Hugging Face dataset page.

Contact

For questions about the dataset, please contact:

Kang Du
Email: kdu800@connect.hkust-gz.edu.cn