Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Optimizing YOLO Architectures for Optimal Road Damage Detection and Classification: A Comparative Study from YOLOv7 to YOLOv10

Vung Pham, Lan Dong Thi Ngoc, Duy-Linh Bui

TL;DR

The paper tackles automated road-damage detection under hardware constraints and imbalanced data by evaluating multiple YOLO architectures, introducing Coordinate Attention, and assembling a hybrid ensemble of a custom YOLOv7 model with Coordinate Attention plus Tiny YOLOv7. It addresses data processing challenges (cropping large images, VOC-to-YOLO annotation conversion, and 90/10 train/validation splits) and mitigates pothole underrepresentation by integrating an external pothole dataset. Through extensive experiments, the authors show that the ensemble achieves an F1 of 0.7027 with an inference time of 0.0547 seconds per image, securing 4th place in ORDDC'2024, while noting that newer YOLO versions offer improvements in mAP but do not consistently improve F1. The work demonstrates practical, reusable steps for real-time road-damage detection and provides a public pipeline on GitHub to support reproducibility and future enhancements, including test-time augmentations and broader ensembling strategies.

Abstract

Maintaining roadway infrastructure is essential for ensuring a safe, efficient, and sustainable transportation system. However, manual data collection for detecting road damage is time-consuming, labor-intensive, and poses safety risks. Recent advancements in artificial intelligence, particularly deep learning, offer a promising solution for automating this process using road images. This paper presents a comprehensive workflow for road damage detection using deep learning models, focusing on optimizations for inference speed while preserving detection accuracy. Specifically, to accommodate hardware limitations, large images are cropped, and lightweight models are utilized. Additionally, an external pothole dataset is incorporated to enhance the detection of this underrepresented damage class. The proposed approach employs multiple model architectures, including a custom YOLOv7 model with Coordinate Attention layers and a Tiny YOLOv7 model, which are trained and combined to maximize detection performance. The models are further reparameterized to optimize inference efficiency. Experimental results demonstrate that the ensemble of the custom YOLOv7 model with three Coordinate Attention layers and the default Tiny YOLOv7 model achieves an F1 score of 0.7027 with an inference speed of 0.0547 seconds per image. The complete pipeline, including data preprocessing, model training, and inference scripts, is publicly available on the project's GitHub repository, enabling reproducibility and facilitating further research.

Optimizing YOLO Architectures for Optimal Road Damage Detection and Classification: A Comparative Study from YOLOv7 to YOLOv10

TL;DR

The paper tackles automated road-damage detection under hardware constraints and imbalanced data by evaluating multiple YOLO architectures, introducing Coordinate Attention, and assembling a hybrid ensemble of a custom YOLOv7 model with Coordinate Attention plus Tiny YOLOv7. It addresses data processing challenges (cropping large images, VOC-to-YOLO annotation conversion, and 90/10 train/validation splits) and mitigates pothole underrepresentation by integrating an external pothole dataset. Through extensive experiments, the authors show that the ensemble achieves an F1 of 0.7027 with an inference time of 0.0547 seconds per image, securing 4th place in ORDDC'2024, while noting that newer YOLO versions offer improvements in mAP but do not consistently improve F1. The work demonstrates practical, reusable steps for real-time road-damage detection and provides a public pipeline on GitHub to support reproducibility and future enhancements, including test-time augmentations and broader ensembling strategies.

Abstract

Maintaining roadway infrastructure is essential for ensuring a safe, efficient, and sustainable transportation system. However, manual data collection for detecting road damage is time-consuming, labor-intensive, and poses safety risks. Recent advancements in artificial intelligence, particularly deep learning, offer a promising solution for automating this process using road images. This paper presents a comprehensive workflow for road damage detection using deep learning models, focusing on optimizations for inference speed while preserving detection accuracy. Specifically, to accommodate hardware limitations, large images are cropped, and lightweight models are utilized. Additionally, an external pothole dataset is incorporated to enhance the detection of this underrepresented damage class. The proposed approach employs multiple model architectures, including a custom YOLOv7 model with Coordinate Attention layers and a Tiny YOLOv7 model, which are trained and combined to maximize detection performance. The models are further reparameterized to optimize inference efficiency. Experimental results demonstrate that the ensemble of the custom YOLOv7 model with three Coordinate Attention layers and the default Tiny YOLOv7 model achieves an F1 score of 0.7027 with an inference speed of 0.0547 seconds per image. The complete pipeline, including data preprocessing, model training, and inference scripts, is publicly available on the project's GitHub repository, enabling reproducibility and facilitating further research.

Paper Structure

This paper contains 30 sections, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Popular object detection techniques
  4. Deep learning techniques for road damage detection
  5. Datasets
  6. RDD2022 dataset
  7. Pothole dataset
  8. Experiments
  9. Data processing
  10. Cropping large images
  11. Annotation conversion and train/validation splitting
  12. Deep Learning Models
  13. Comparing different YOLO versions
  14. Incorporating Coordinate Attention layers
  15. Training hyperparameters
  16. ...and 15 more sections

Figures (2)

  • Figure 1: Distribution of damage types in the RDD2022 dataset.
  • Figure 2: An example image from the Norway folder (Norway_000003.jpg). This large image appears to be a composite of two smaller images tiled from left to right. The majority of the road section and visible damages are concentrated in the lower-left corner, highlighted by the red rectangle with dimensions of $1,824 \times 1,824$ pixels.