wwtylo/DDOS

---
license: cc-by-nc-4.0
pretty_name: DDOS
task_categories:
- image-segmentation
- depth-estimation
task_ids:
- semantic-segmentation
tags:
- drones
- uav
- aerial
- vision
- wires
- cables
- outdoor
- segmentation
- semantic segmentation
- depth
- weather
- flying
- computer vision
- cv
size_categories:
- 10K<n<100K
---
# DDOS: The Drone Depth and Obstacle Segmentation Dataset

The Drone Depth and Obstacle Segmentation (DDOS) dataset comprises synthetic aerial images captured by drones, along with corresponding depth maps and pixel-wise semantic segmentation masks. DDOS is purpose-built to support research and development in computer vision, focusing on tasks such as depth estimation and obstacle segmentation from aerial imagery. Emphasizing the detection of thin structures like wires and effective navigation in diverse weather conditions, DDOS serves as a valuable resource for advancing algorithms in autonomous drone technology.

- **Paper:** [DDOS: The Drone Depth and Obstacle Segmentation Dataset](https://arxiv.org/abs/2312.12494)

---------

## Data Structure

DDOS is organised as follows:

- Data Splits:
- Train: Contains 300 flights with a total of 30k images for training.
- Validation: Contains 20 flights with a total of 2k images for validation during model development.
- Test: Contains 20 flights with a total of 2k images for the final evaluation of the trained model.

- Environments:
- Neighbourhood: Contains data captured in urban and residential environments.
- Park: Contains data captured in park and natural environments.

- Flights:
- Each flight is represented by a unique flight ID and is contained within the corresponding environment directory.

- Data for Each Flight:
- Image: Contains RGB images captured by the drone camera.
- Depth: Contains depth maps representing the distance of objects from the camera. These maps are saved as uint16 PNG images, where pixel values range from 0 to 65535, representing distances from 0 to 100 meters linearly.
- Segmentation: Contains pixel-wise segmentation masks for semantic segmentation. Classes, as well as their corresponding mappings, are mentioned below.
- Flow: Contains optical flow data representing the apparent motion of objects between consecutive frames.
- Surface Normal: Contains surface normal maps representing the orientation of object surfaces.

Overview of file structure:
```
data/
├── train/
│ ├── neighbourhood/
│ │ ├── 0/
│ │ │ ├── depth/
│ │ │ │ ├── 0.png
│ │ │ │ ├── ...
│ │ │ │ └── 99.png
│ │ │ ├── flow/
│ │ │ │ ├── 0.png
│ │ │ │ ├── ...
│ │ │ │ └── 99.png
│ │ │ ├── image/
│ │ │ │ ├── 0.png
│ │ │ │ ├── ...
│ │ │ │ └── 99.png
│ │ │ ├── segmentation/
│ │ │ │ ├── 0.png
│ │ │ │ ├── ...
│ │ │ │ └── 99.png
│ │ │ ├── surfacenormals/
│ │ │ │ ├── 0.png
│ │ │ │ ├── ...
│ │ │ │ └── 99.png
│ │ │ ├── metadata.csv
│ │ │ └── weather.csv
│ │ ├── ...
│ │ └── 249/
│ │ └── ...
│ └── park/
│ ├── 0/
│ │ ├── depth/
│ │ │ └── ...
│ │ ├── flow/
│ │ │ └── ...
│ │ ├── image/
│ │ │ └── ...
│ │ ├── segmentation/
│ │ │ └── ...
│ │ ├── surfacenormals/
│ │ │ └── ...
│ │ ├── metadata.csv
│ │ └── weather.csv
│ ├── ...
│ └── 49/
│ └── ...
├── validation/
│ └── ...
└── test/
└── ...
```

---------
## Additional Information

**Class Mapping:** The segmentation masks use the following class labels for obstacle segmentation:
```python
CLASS_MAPPING = {
'ultra_thin': 255,
'thin_structures': 240,
'small_mesh': 220,
'large_mesh': 200,
'trees': 180,
'buildings': 160,
'vehicles': 140,
'animals': 100,
'other': 80
}
```
**Metadata:** The dataset contains metadata, such as coordinates, pose, acceleration, weather conditions and camera parameters, which provide valuable contextual information about each flight.

---------
## Dataset Usage

- **Data Loading:**
To load and use the DDOS dataset in your projects, you can refer to the official PyTorch data loading tutorial: [PyTorch Data Loading Tutorial](https://pytorch.org/tutorials/beginner/basics/data_tutorial.html).
This tutorial will guide you through the process of loading data, creating data loaders, and preparing the dataset for training or evaluation using PyTorch.

- **Respect the Data Splits:**
Please ensure that the testing data is not used for validation. Mixing these datasets could lead to inaccurate assessments of model performance. Maintaining separate datasets for testing and validation helps ensure reliable evaluation and accurate reporting of results.

---------
## License
DDOS is openly licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)

---------
## Citation
If you use DDOS in your research or projects, please cite our paper:

```
@article{kolbeinsson2023ddos,
title={{DDOS}: The Drone Depth and Obstacle Segmentation Dataset},
author={Benedikt Kolbeinsson and Krystian Mikolajczyk},
journal={arXiv preprint arXiv:2312.12494},
year={2023}
}
```