README.md

---
license: cc-by-nc-sa-4.0
pretty_name: UAVLight
task_categories:
- image-to-image
- image-to-3d
tags:
- computer-vision
- novel-view-synthesis
- 3d-reconstruction
- gaussian-splatting
- nerf
- relighting
- illumination-robustness
- uav
- outdoor-scenes
- benchmark
size_categories:
- 1K<n<10K
---

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

![UAVLight teaser](assets/UAVLight_teaser.png)

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

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

A short video preview is also available:

[Watch UAVLight video preview](assets/uavlight.mp4)

This dataset accompanies the paper:

**UAVLight: A Benchmark for Illumination-Robust 3D Reconstruction in Unmanned Aerial Vehicle (UAV) Scenes**

## Overview

UAVLight is intended to evaluate whether a 3D reconstruction or novel-view synthesis method can maintain stable geometry and appearance quality under outdoor illumination variations. In UAV capture, the same scene may exhibit substantial appearance changes across different capture times, sun positions, shadow layouts, and camera trajectories. These changes can make reconstruction and view synthesis more difficult than in standard static-lighting benchmarks.

The benchmark is particularly useful for studying:

- illumination-robust 3D reconstruction
- novel-view synthesis for UAV-captured scenes
- outdoor scene reconstruction under changing sunlight and shadows
- lighting-aware Gaussian Splatting and NeRF-style methods
- cross-view reconstruction consistency under illumination variation
- relighting and lighting-transfer evaluation

Each scene contains multi-view images, sparse reconstruction outputs, predefined train/test splits, sun direction annotations, and optional geometry assets such as point clouds and meshes.

## Repository Structure

The Hugging Face repository is organized as follows:

```text
UAVLight/
  README.md

  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
```

The `data/` directory contains scene-level zip archives. Each zip file corresponds to one UAV scene. The `metadata/` directory provides summary files describing the released scenes, archive sizes, and file lists. The `assets/` directory contains visual materials used by this dataset card.

## Scene Archive Structure

After extracting a scene archive, the directory structure is:

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

For example:

```text
1121211223101030/
  images/
  sparse/
  dense_points.ply
  downsampled_points.ply
  mesh.ply
  split.csv
  sun_directions.txt
  train_list.txt
  test_list.txt
```

## File Descriptions

### `images/`

This folder contains the multi-view RGB images for the scene. These images are the main visual observations used for reconstruction, novel-view synthesis, and benchmark evaluation.

### `sparse/`

This folder contains sparse reconstruction files, such as camera poses and COLMAP-style sparse reconstruction outputs. These files can be used to initialize or evaluate reconstruction methods that rely on calibrated cameras.

### `split.csv`

This file records the predefined split information for the scene. It can be used to identify which images belong to training and testing subsets.

### `train_list.txt`

This file contains the list of training images used for scene reconstruction or model fitting.

### `test_list.txt`

This file contains the list of testing images used for novel-view synthesis and benchmark evaluation.

### `sun_directions.txt`

This file provides sun direction annotations associated with the scene/images. These annotations are useful for illumination-aware reconstruction, lighting transfer, relighting-related analysis, and evaluating robustness under outdoor lighting variation.

### `dense_points.ply`

A dense point cloud reconstructed for the scene. This is provided as an optional geometry asset and may be useful for visualization, geometry analysis, or method initialization.

### `downsampled_points.ply`

A downsampled version of the point cloud. This file is smaller and can be useful for quick visualization or lightweight processing.

### `mesh.ply`

A reconstructed mesh for the scene. This is provided as an optional geometry asset and may be useful for visualization or geometry-related analysis.

## Metadata Files

The `metadata/` directory contains several files to help users inspect and manage the dataset.

### `metadata/scenes.csv`

A scene-level summary file. Each row corresponds to one scene and records whether the expected files are available, including image folders, sparse reconstruction files, geometry assets, sun direction annotations, and train/test split files.

### `metadata/zip_sizes.csv`

A summary of all released scene archives and their file sizes.

### `metadata/file_list.txt`

A full file list generated from the original packed dataset directory.

### `metadata/zip_list.txt`

A list of all released scene-level zip archives.

### `metadata/summary.txt`

A compact summary of the release, including the number of scenes, number of zip archives, and total compressed size.

## Download

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

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

Alternatively, you can download individual scene archives from the `data/` folder.

For example, after downloading one scene archive:

```bash
unzip data/<scene_id>.zip -d UAVLight_scenes/
```

The extracted scene will follow the structure described above.

## Usage Example

A typical workflow is:

```text
1. Download the dataset or selected scene archives.
2. Extract the scene zip files.
3. Use train_list.txt for reconstruction or model training.
4. Use test_list.txt for novel-view synthesis evaluation.
5. Use sparse/ camera files for pose information.
6. Optionally use sun_directions.txt for illumination-aware analysis.
7. Optionally use dense_points.ply, downsampled_points.ply, or mesh.ply for geometry visualization or initialization.
```

## Intended Use

UAVLight is intended for academic research on robust 3D reconstruction and novel-view synthesis in outdoor UAV scenes. Potential use cases include:

- benchmarking illumination-robust reconstruction methods
- evaluating Gaussian Splatting and NeRF-based methods under outdoor lighting variation
- studying the effect of sunlight, shadows, and exposure variation on 3D reconstruction
- developing lighting-aware scene representations
- evaluating relighting or lighting-transfer consistency in reconstructed scenes

## Limitations

UAVLight focuses on outdoor UAV scenes and illumination robustness. The dataset is not intended to cover all possible outdoor environments, weather conditions, or dynamic scene changes. Users should also note that geometry assets such as point clouds and meshes are provided as auxiliary reconstruction outputs and may not be perfect ground truth.

## License

This dataset is released for non-commercial research use only under the license specified in this repository.

## Citation

If you use UAVLight in your research, please cite:

```bibtex
@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}
}
```

## Contact

For questions about the dataset, please contact:

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