pathikg/drone-detection-dataset

Dataset Card for Drone Detection Dataset

This dataset card describes the processed version of the Drone Detection Dataset, originally curated by Maciej Pawełczyk and Marek Wojtyra, adapted to a COCO-style format for efficient usage in modern deep learning pipelines

Dataset Details

Dataset Description

The Drone Detection Dataset is a real-world object detection dataset for UAV detection tasks. It includes RGB images annotated with bounding boxes in the COCO format. This dataset is ideal for training and evaluating object detection models like Faster R-CNN, YOLO, and DETR.

• Curated by: Maciej Pawełczyk, Marek Wojtyra
• Shared by: pathikg
• Language(s) (NLP): Not applicable
• License: MIT License

Dataset Sources

• Repository: https://github.com/Maciullo/DroneDetectionDataset
• Paper: https://ieeexplore.ieee.org/document/9205392

Uses

Direct Use

This dataset is suitable for:

• Training object detection models.
• Research on UAV detection and monitoring in various environments and lighting conditions.

Out-of-Scope Use

Out-of-Scope Use

• The dataset is not ideal for real-time UAV tracking.
• It may not generalize well to unseen drone types or environments without further fine-tuning.

Dataset Structure

The dataset is structured as a Hugging Face DatasetDict with train and test splits. Each split contains features:

DatasetDict({
    train: Dataset({
        features: ['width', 'height', 'objects', 'image', 'image_id'],
        num_rows: 51446
    }),
    test: Dataset({
        features: ['width', 'height', 'objects', 'image', 'image_id'],
        num_rows: 2625
    })
})

Example Datapoint

{
 'width': 640,
 'height': 480,
 'objects': {
    'bbox': [[281.0, 210.0, 25.0, 19.0]],  # COCO format: [x, y, width, height]
    'category': [0],  # Category index for the drone
    'area': [475.0],  # Area of the bounding box
    'id': [0]         # Object ID
 },
 'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=640x480>,
 'image_id': 2
}

Features

• width and height: Dimensions of the image.
• objects:
  • bbox: Bounding boxes in COCO format [x, y, width, height].
  • category: Class labels (0 for drones).
  • area: Area of each bounding box.
  • id: Unique ID for each object.
• image: The image as a PIL object.
• image_id: Unique identifier for the image.

Dataset Creation

Curation Rationale

The dataset was designed to facilitate UAV detection, addressing challenges like detecting small objects across varying scales and environments. It focuses on static object detection rather than tracking.

Source Data

Data Collection and Processing

• Source: Frames extracted from publicly available YouTube videos.
• Processing: Bounding boxes were annotated manually in COCO format for initial samples and later supplemented with automated annotations using trained models.

Who are the source data producers?

The source videos were created by various YouTube users. The dataset annotations were curated as part of the authors' research.

Bias, Risks, and Limitations

• Bias: The dataset includes drones in specific environments and lighting conditions, which may limit its generalizability.
• Limitations:
  • The dataset does not cover all drone types or edge cases.
  • It may not be suitable for real-time applications without modifications.

Citation

BibTeX:

@article{pawelczyk2020drone,
  title={Real World Object Detection Dataset for Quadcopter Unmanned Aerial Vehicle Detection},
  author={Pawełczyk, Maciej and Wojtyra, Marek},
  journal={IEEE Access},
  volume={8},
  pages={174394--174409},
  year={2020},
  publisher={IEEE},
  doi={10.1109/ACCESS.2020.3026192}
}

APA:

Pawełczyk, M., & Wojtyra, M. (2020). Real World Object Detection Dataset for Quadcopter Unmanned Aerial Vehicle Detection. IEEE Access, 8, 174394-174409. https://doi.org/10.1109/ACCESS.2020.3026192

More Information

For details, refer to the original GitHub repository.

Dataset Card Authors

• me, pathikg

Dataset Card Contact

For questions, contact pathikg