BirDrone

Citation Author(s):
Fatin Najihah
Muhammad Zamri
IIUM
Teddy Surya
Gunawan
IIUM
Submitted by:
Teddy Surya Gunawan
Last updated:
Fri, 05/10/2024 - 09:26
DOI:
10.21227/ettb-0w28
Data Format:
License:
0
0 ratings - Please login to submit your rating.

Abstract 

The BirDrone dataset is compiled by aggregating images of small drones and birds sourced from various online datasets. It comprises 2970 high-resolution images (640x640 pixels), each featuring unique backdrops and lighting conditions. This dataset is designed to enhance machine learning models by simulating real-world scenarios.

 

Dataset Specifications:

  • Image Count: 2970 images, with 2617 drone images and 353 bird images.
  • Image Resolution: Each image is uniformly sized at 640x640 pixels.
  • Annotation Details: The dataset includes 6162 annotations with the smallest bounding box sized at 7x14 pixels and the largest at 65x182 pixels.
  • Classes: Two categories are represented—drones and birds.
  • Annotation Format: Annotations are formatted according to YOLOv8 specifications.

Pre-processing and Augmentation:

  • Pre-processing Techniques: Images have undergone auto-orientation, resizing, and auto-contrast adjustments to standardize and enhance visual clarity.
  • Augmentation Techniques: To increase variability and robustness, the images have been augmented with rotation and exposure adjustments, preparing the dataset for diverse environmental and lighting conditions.

Dataset Distribution:

  • Training Set: 80% (2376 images)
  • Validation Set: 20% (594 images)

File Size: 96.3 MB

Instructions: 

The RAR file contains the train and valid folder. In each folder, there are images and labels folders, with data.yaml compatible with Roboflow.