# Road Damage Detection and Classification with Detectron2 and Faster R-CNN

**Authors:** Vung Pham, Chau Pham, Tommy Dang  
**arXiv:** [2010.15021](https://arxiv.org/abs/2010.15021)  
**Date:** 2020-10-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2010.15021v1/

## Key points

- Importance of road maintenance for human safety and timely repair
- Evaluation of Detectron2's implementation of Faster R-CNN using various base models and configurations
- Experiments conducted using the Global Road Damage Detection Challenge 2020 dataset
- X101-FPN base model with default configurations is efficient and adaptable
- Achieved F1 scores of 51.0% and 51.4% for test1 and test2 sets, respectively
- Comparison with existing annotations to address discrepancies in prediction results
- Proposed strategies to enhance labeling process for the dataset
- Significance of road damage detection and classification for human safety on roads
- Valuable insights into the effectiveness of Detectron2's implementation in detecting and classifying road damages
- Potential improvements in labeling processes to enhance accuracy in future studies related to road maintenance and safety

## Summary

In their study, Vung Pham, Chau Pham, and Tommy Dang focus on the importance of road maintenance for human safety and the need to quickly detect and classify road damages for timely repair. They evaluate the effectiveness of Detectron2's implementation of Faster R-CNN using various base models and configurations. The researchers conduct experiments using the Global Road Damage Detection Challenge 2020 dataset from the IEEE Big Data Cup Challenge. The results of their experiments demonstrate that the X101-FPN base model for Faster R-CNN, along with Detectron2's default configurations, is efficient and adaptable enough to be applied in different countries participating in the challenge. The approach achieves F1 scores of 51.0% and 51.4% for the test1 and test2 sets of the challenge, respectively. However, despite visually promising prediction results, the F1 scores obtained are relatively low. To address this issue, the researchers compare their prediction results with existing annotations and identify discrepancies. Consequently, they propose strategies to enhance the labeling process for this dataset. Overall, this study highlights the significance of road damage detection and classification for ensuring human safety on roads. The evaluation of Detectron2's implementation of Faster R-CNN provides valuable insights into its effectiveness in detecting and classifying road damages. The findings also shed light on potential improvements in labeling processes to further enhance accuracy in future studies related to road maintenance and safety.