Deep Learning Based Vehicle Detection on Real and Synthetic Aerial Images: Training Data Composition and Statistical...

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Deep learning algorithm performance relies heavily on training data quantity and quality.
• Acquiring real-world airborne sensor data is challenging and costly.
• Synthetic data generation via simulation environments offers a viable alternative.

Purpose of the Study:

• To evaluate the complete process chain for using synthetic data in vehicle detection.
• To identify key factors influencing detection performance with synthetic data.
• To derive design guidelines for synthetic data generation and application.

Main Methods:

• Training deep learning models with various configurations of real and synthetic aerial image datasets.
• Employing a statistical evaluation procedure with image descriptors for factor identification.
• Iteratively refining synthetic data generation based on findings and assessing detection performance improvements.

Main Results:

• Demonstrated the impact of different training data compositions (real vs. synthetic) on vehicle detection accuracy.
• Identified critical influencing factors in the synthetic data generation pipeline.
• Showcased performance enhancements achievable through optimized synthetic data strategies.

Conclusions:

• Synthetic data is a valuable resource for enhancing deep learning-based vehicle detection, especially with airborne sensor data.
• The study provides actionable design guidelines for generating and utilizing synthetic data effectively.
• Optimized synthetic data generation can significantly improve detection performance and reduce data acquisition costs.