3DAeroRelief: The First 3D Benchmark UAV Dataset for Post-Disaster Assessment

3DAeroRelief is a high-resolution 3D point cloud dataset designed to benchmark semantic segmentation algorithms in post-disaster scenarios.

Download

The 3D point cloud data (Areas 1-8) is available for download via Dropbox:

• Download 3DAeroRelief Dataset

Dataset Structure

The project is organized into two parts: the Code Repository (hosted here on GitHub) and the 3D Data (hosted externally).

1. Repository Contents (GitHub)

Files included in this repository:

• ColmapConfig/: Contains the specific COLMAP configuration files used to generate the reconstructions.
• code/: Python utility scripts for data handling.
• labels.txt: List of semantic class names and IDs.

2. External Dataset Contents (Dropbox)

Files included in the download link:

• Area_<n>/: Contains the 3D data for each location.
  • pp<n>.ply: The raw RGB 3D point cloud (Geometry + Texture).
  • segmentpp<n>.ply: The associated label.

(GitHub Repository)              (External Dropbox)
3DAeroRelief/                    3DAeroRelief_Dataset/
├── ColmapConfig/                ├── Area_1/
│   ├── extractor.ini            │   ├── pp1.ply
│   ├── mapper.ini               │   └── segmentpp1.ply
│   └── matcher.ini              ├── Area_2/ 
├── code/                        │   ├── pp1.ply
│   └── add_labels_and_viz.py    │   └── segmentpp1.ply
└── labels.txt                   └── Area_3/ ... Area_8/

Usage and Data Processing

In the raw dataset, geometry (coordinates/colors) and semantic labels are stored in separate files (pp<n>.ply and segmentpp<n>.ply). We provide a Python script to merge these into a single, training-ready PLY file containing a label field.

Prerequisites

Install the required dependencies:

pip install numpy scipy open3d 

Merging Labels and Geometry

Use the code/add_labels_and_viz.py script to align labels to the geometry using Nearest Neighbor matching and save the result.

Basic Command:

python code/add_labels_and_viz.py --input 3DAeroRelief --output processed_data

Options:

• --viz: Generates additional visualization files (original RGB and false-color labels).

• --areas <Area_Name>: Process specific areas (e.g., --areas Area_1 Area_2).

• --workers <int>: Specify the number of parallel CPU processes (default: uses all available cores).

Output Format

The script generates a binary PLY file with the following properties, suitable for loading into standard 3D learning frameworks:

• x, y, z: 3D Coordinates.
• red, green, blue: RGB Texture.
• label: Integer Class ID (0-4).

Classes

The dataset classifies points into distinct semantic categories. The merging script maps the raw label colors to the following integer IDs:

ID	Class Name	Color (RGB)
0	Background	(0, 0, 0)
1	Building-Damage	(230, 25, 75)
2	Building-No-Damage	(70, 240, 240)
3	Road	(255, 255, 25)
4	Tree	(0, 128, 0)

Refer to labels.txt for the source mapping configuration.

License

• Dataset: Licensed under CC0. You are free to share and adapt the data, provided you give appropriate credit.
• Code: The provided utility scripts are licensed under the MIT License.

Citation

If you use this dataset in your research, please cite our paper.

@misc{le20253daerorelief3dbenchmarkuav,
      title={3DAeroRelief: The first 3D Benchmark UAV Dataset for Post-Disaster Assessment}, 
      author={Nhut Le and Ehsan Karimi and Maryam Rahnemoonfar},
      year={2025},
      eprint={2509.11097},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.11097}, 
}

Contact

For questions regarding the dataset or the paper, please open an issue in this repository or contact the corresponding author at: maryam@lehigh.edu