How to construct low-altitude aerial image datasets for deep learning.

Area of Science:

• Computer Vision
• Robotics
• Data Science

Background:

• Unmanned Aerial Vehicle (UAV) technology combined with computer vision is driving increased adoption of UAV applications.
• Deep learning-based computer vision requires extensive, task-specific datasets for effective algorithm training.
• Existing aerial image datasets often lack the specificity needed for specialized UAV tasks.

Purpose of the Study:

• To propose a framework for constructing custom datasets for specific computer vision tasks in low-altitude aerial imagery.
• To address the need for large-scale, specialized aerial image datasets to improve UAV capabilities.

Main Methods:

• Analysis of existing low-altitude aerial image datasets and their characteristics.
• Recommendations for data collection strategies tailored to low-altitude aerial imagery.
• Guidance on image annotation tools and crowdsourcing platforms for generating labeled data.
• Introduction to data augmentation techniques, including traditional methods, oversampling, and generative adversarial networks (GANs).

Main Results:

• A comprehensive framework for building task-specific aerial image datasets is presented.
• Insights into the unique characteristics of low-altitude aerial images are provided.
• Effective methods for data annotation and augmentation are discussed to overcome data scarcity.

Conclusions:

• The proposed framework facilitates the creation of specialized datasets essential for advancing UAV computer vision.
• Data collection, annotation, and augmentation strategies are vital for improving the performance of UAVs in specific applications like object detection.