Optimizing Road Damage Detection with YOLOv10: A Resource-Efficient Approach Utilizing Augmentation, Data Sampling, and Hyperparameter Tuning

Efficient road damage detection is key for timely maintenance and decreasing long-term costs, especially for large-scale road networks. While existing models achieve high accuracy, they often overlook resource optimization, making them impractical for widespread deployment. In this study, we propose an optimized approach using the YOLOv10 model, designed to balance detection accuracy and inference speed, enabling scalable road damage detection across diverse regions. Key to our success were several optimization techniques, including strategic data sampling from a large, diverse dataset of 47,420 images across six countries, automated hyperparameter tuning for 10,000 epochs, and advanced data augmentation methods such as Bounding-Box-Safe Random Crop and Optimized Test-Time Augmentation. These strategies enhanced the model’s robustness and generalization across varying road damages. Our model was developed in response to the Optimized Road Damage Detection Challenge (ORDDC) 2024, where we attained an F1 score of 0.7073 and an inference speed of 0.2886, securing 5th place.