EngJamesO/pothole-detector

Pothole Detection Models

This repository contains trained YOLO models for detecting potholes in road images. The models were developed to assist with road infrastructure monitoring and maintenance planning.

Model Overview

Five YOLO architectures were trained and evaluated for pothole detection:

Model	Parameters	mAP50	mAP50-95	Latency (ms)	FPS
YOLOv9c	25.3M	0.792	0.485	36.3	27.6
YOLOv12n	2.56M	0.785	0.463	13.9	72.1
YOLOv8n	3.01M	0.782	0.465	6.9	144.7
YOLOv11	2.58M	0.779	0.460	9.3	107.0
YOLOv8-FPN	2.21M	0.707	0.396	6.4	155.8

Intended Uses

• Primary Use: Automated pothole detection in road images for infrastructure assessment
• Secondary Use: Research in object detection, computer vision applications for transportation
• Target Users: Road maintenance departments, urban planners, researchers

Model Selection Guide

• YOLOv9c: Best overall accuracy (mAP50: 0.792), recommended when accuracy is prioritized over speed
• YOLOv8n: Good balance of accuracy and speed, suitable for real-time applications
• YOLOv12n: Lightweight with competitive accuracy, ideal for edge deployment
• YOLOv11: Efficient architecture with modern optimizations
• YOLOv8-FPN: Fastest inference, suitable for high-throughput scenarios

Performance at Different Confidence Thresholds

Based on qualitative evaluation (40 test images, 86 ground truth objects):

YOLOv9c (Recommended)

• Conf=0.25: Precision: 0.804, Recall: 0.831, F1: 0.790
• Conf=0.50: Precision: 0.846, Recall: 0.723, F1: 0.753
• Conf=0.70: Precision: 0.700, Recall: 0.505, F1: 0.558

YOLOv8n

• Conf=0.25: Precision: 0.807, Recall: 0.838, F1: 0.791
• Conf=0.50: Precision: 0.840, Recall: 0.710, F1: 0.751
• Conf=0.70: Precision: 0.600, Recall: 0.437, F1: 0.482

Usage

Installation

pip install ultralytics opencv-python

Python inference

from ultralytics import YOLO

# Load model
model = YOLO("yolov9c.pt")

# Run inference
results = model("path/to/image.jpg")

# Visualize results
results[0].show()

Limitations

• Models trained on specific road conditions; performance may vary on different terrain or weather conditions
• Detection accuracy depends on image quality and resolution
• Models may have difficulty with very small or heavily obscured potholes
• Performance not validated on night-time or extreme weather images