Road Damage Detection using YOLOv7

A deep learning-based road damage detection system using YOLOv7, trained on the multi-national RDD2022 dataset. This project achieves 63.85% mAP@0.5 and 31.65% mAP@0.5:0.95 in detecting 4 types of road damage across images from 7 countries.

Overview

This project implements an automated road damage detection system using YOLOv7 object detection architecture. The system can identify and classify 4 types of road damage:

Class	Description	Example Count
D00	Longitudinal Crack	1,201 labels
D10	Transverse Crack	611 labels
D20	Alligator Crack	529 labels
D40	Pothole	375 labels

The model is trained on 21,983 images from 7 countries and validated on 1,189 images from the RDD2022 dataset.

Features

Multi-country generalization - Trained on diverse road conditions from China, Czech Republic, India, Japan, Norway, and United States
Ensemble inference - Combines predictions from two models trained at different resolutions (640×640 and 512×512)
Interactive web interface - Gradio-based UI for easy testing and demonstration
Custom YOLOv7 architecture - Optimized for road damage detection with 37.2M parameters
Transfer learning - Initialized from pre-trained YOLOv7 weights for faster convergence
Comprehensive logging - TensorBoard integration for training monitoring

Dataset

Dataset: RDD2022 - The multi-national Road Damage Dataset

Dataset Statistics

Country	Training Images	Image Size	Notes
China (Drone)	482	512×512	Aerial imagery
China (MotorBike)	43	512×512	Ground-level
Czech Republic	1,757	600×600	-
India	4,483	720×720	Largest subset
Japan	2,606	600×600, 1080×1080	Mixed resolutions
Norway	5,247	3643-4040×2035-2044	High-resolution
United States	0	-	Test set only
Total	21,983	-	50,879 labels

Model Performance

Best Model: yolov7all6 (640×640 resolution)

Training: 100 epochs | Batch Size: 10 | GPU: NVIDIA RTX 4060 Laptop

Metric	Value
mAP@0.5	63.85%
mAP@0.5:0.95	31.65%
Precision	68.67%
Recall	57.49%
Parameters	37.2M
Training Time	~28 hours

Per-Class Performance

Class	Precision	Recall	mAP@0.5	mAP@0.5:0.95
D00	66.3%	63.5%	66.6%	36.0%
D10	68.6%	53.2%	61.3%	29.5%
D20	71.7%	61.2%	69.6%	36.7%
D40	68.1%	52.0%	58.0%	24.4%

Installation

Prerequisites

Python 3.7 - 3.12
CUDA 11.8+ (for GPU acceleration)
Conda (recommended)

Step 1: Clone YOLOv7 Repository

git clone https://github.com/WongKinYiu/yolov7.git
cd yolov7

Step 2: Replace Configuration Files

IMPORTANT: Replace the default YOLOv7 files with custom configurations from this project:

Copy DetectingRoadDamage/folders_to_replace/training → Replace yolov7/cfg/training
Copy DetectingRoadDamage/folders_to_replace/data → Replace yolov7/data

Step 3: Install Dependencies

pip install -r requirements.txt

Key Dependencies:

torch>=1.7.0
torchvision>=0.8.1
opencv-python>=4.1.1
numpy>=1.18.5,<1.24.0
matplotlib>=3.2.2
tensorboard>=2.4.1
gradio==5.29.0
pillow>=7.1.2
tqdm
scipy
seaborn

Step 4: Fix Environment Issues (Windows)

If you encounter KMP_DUPLICATE_LIB_OK errors on Windows:

# Option 1: Conda environment variable
conda env config vars set KMP_DUPLICATE_LIB_OK=TRUE
conda deactivate
conda activate base  # or your environment name

# Option 2: Set system environment variable
set KMP_DUPLICATE_LIB_OK=TRUE

Step 5: Download Dataset

Download RDD2022 dataset from Figshare
Extract to your preferred location (e.g., F:/RoadDamageDetection/Datasets/RDD2022/)
Update paths in data/crdd22.yaml:

train: /path/to/RDD2022/images/train
val: /path/to/RDD2022/images/val

nc: 4
names: ['D00', 'D10', 'D20', 'D40']

Step 6: Download Pre-trained Weights

Download YOLOv7 pre-trained weights:

wget https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7_training.pt

Dataset Preparation (REQUIRED Before Training)

CRITICAL: Before starting training, you MUST run these notebooks to prepare the dataset:

Step 1: Validate Dataset Image Sizes

Run CheckSizes.ipynb to verify image dimensions and XML annotation consistency:

jupyter notebook CheckSizes.ipynb

Purpose:

Validates image dimensions match XML annotations
Identifies square vs. non-square images
Detects resolution inconsistencies across countries
Ensures dataset integrity before training

Step 2: Convert Dataset to YOLOv7 Format

Run Local_YOLOv7_DataConversion.ipynb to convert RDD2022 annotations to YOLOv7 format:

jupyter notebook Local_YOLOv7_DataConversion.ipynb

Purpose:

Converts XML annotations to YOLO txt format
Normalizes bounding box coordinates
Creates proper train/val splits
Organizes images and labels in YOLOv7 structure

Expected Output:

RDD2022/
├── images/
│   ├── train/
│   └── val/
└── labels/
    ├── train/
    └── val/

WARNING: Skipping these steps will result in training failures or incorrect annotations!

Project Structure

DetectingRoadDamage/
├── yolov7/                          # YOLOv7 implementation (~7.5GB - NOT included)
├── folders_to_replace/               # Trained model weights
│   ├── training                      # Custom model configurations [REPLACE yolov7/cfg/training]
│   └── data                          # Dataset configurations [REPLACE yolov7/data]
├── test_images/                           # Sample test images
├── index.ipynb                            # Ensemble inference + Gradio demo
├── CheckSizes.ipynb                       # [REQUIRED] Dataset validation utilities
├── Local_YOLOv7_DataConversion.ipynb      # [REQUIRED] Dataset format conversion
├── Directory_Structure_CRDDC_RDD2022.txt  # Dataset structure reference
├── commands to run.txt                    # Training command reference
├── yolov7all6-log.txt                     # Complete training logs
├── Documentation/
│   ├── PathFinders-Final Report.pdf
│   └── results_yolov7all6.txt
└── images/                                # Visualization results

Usage

Training

Model 1: yolov7all6 (640×640, 100 epochs)

cd yolov7

python train.py --workers 8 --device 0 --batch-size 10 \
  --data data/crdd22.yaml \
  --img 640 640 \
  --cfg cfg/training/yolov7-custom.yaml \
  --weights yolov7_training.pt \
  --name yolov7all6 \
  --hyp data/hyp.scratch.p5.custom.yaml \
  --epoch 100

Model 2: yolov7all61 (512×512, 200 epochs)

python train.py --workers 8 --device 0 --batch-size 10 \
  --data data/crdd22.yaml \
  --img 512 512 \
  --cfg cfg/training/yolov7-custom.yaml \
  --weights yolov7_training.pt \
  --name yolov7all61 \
  --hyp data/hyp.scratch.p5.custom.yaml \
  --epoch 200

Monitor Training with TensorBoard

cd yolov7
python -m tensorboard.main --logdir runs/train

Then open your browser to http://localhost:6006

Resume Training (if interrupted)

python train.py --resume

Inference

Single Image/Folder Detection (Model 1)

python detect.py \
  --weights runs/train/yolov7all6/weights/best.pt \
  --conf 0.10 \
  --img 640 \
  --source /path/to/test_images \
  --save-conf

Single Image/Folder Detection (Model 2)

python detect.py \
  --weights runs/train/yolov7all61/weights/best.pt \
  --conf 0.25 \
  --img 512 \
  --source /path/to/test_images \
  --save-conf

Parameters:

--weights: Path to trained model weights
--conf: Confidence threshold (0.0-1.0)
--img: Input image size
--source: Path to images, videos, or folder
--save-conf: Save confidence scores

Web Demo (Gradio)

The project includes an interactive Gradio web interface for ensemble inference using both trained models.

Run the Demo

jupyter notebook index.ipynb

Or run directly in Google Colab:

Upload index.ipynb to Colab
Mount Google Drive or upload weights
Run all cells
Access the Gradio interface

Features:

Upload single or multiple images
Adjustable confidence threshold (0.1 - 0.9)
Ensemble prediction from both models
Visual bounding boxes with class labels
Gallery view for batch results

Ensemble Pipeline

The demo combines predictions from:

yolov7all6.pt (640×640 resolution)
yolov7all61.pt (512×512 resolution)

Predictions are merged using Non-Maximum Suppression (NMS) for improved accuracy.

Results

Training Progression (yolov7all6)

Epoch	Box Loss	Obj Loss	Cls Loss	mAP@0.5	mAP@0.5:0.95
0	0.0639	0.01289	0.01437	25.06%	9.10%
10	0.04682	0.0118	0.00574	46.34%	19.94%
50	0.04147	0.01102	0.00391	60.16%	28.92%
90	0.03702	0.01007	0.00250	63.25%	31.20%
99	0.03639	0.00985	0.00235	63.85%	31.65%

Sample Detections

Place your sample detection images in the images/ folder to showcase results.

Key Files

File	Description
`index.ipynb`	Ensemble inference pipeline + Gradio UI
`CheckSizes.ipynb`	[REQUIRED] Dataset validation and image size analysis
`Local_YOLOv7_DataConversion.ipynb`	[REQUIRED] Convert RDD2022 XML to YOLO format
`Directory_Structure_CRDDC_RDD2022.txt`	Expected dataset folder structure reference
`cfg/training/yolov7-custom.yaml`	Custom YOLOv7 architecture (4 classes)
`data/crdd22.yaml`	RDD2022 dataset configuration
`data/hyp.scratch.p5.custom.yaml`	Custom training hyperparameters
`yolov7all6-log.txt`	Complete training logs for Model 1 (100 epochs)
`yolov7all61-log.txt`	Complete training logs for Model 2 (200 epochs)
`commands to run.txt`	All training/inference commands used

Troubleshooting

1. CUDA Out of Memory

# Reduce batch size
python train.py --batch-size 4 ...  # Instead of 10

2. KMP Duplicate Library Error (Windows)

set KMP_DUPLICATE_LIB_OK=TRUE

3. Image Size Mismatch

Make sure to use the correct --img parameter:

Model 1 (yolov7all6): --img 640
Model 2 (yolov7all61): --img 512

4. Missing Weights

Download pre-trained weights:

wget https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7_training.pt

References

Dataset

@data{RDD2022,
  author = {Arya, Deeksha and Maeda, Hiroya and Ghosh, Sanjay Kumar and others},
  title = {RDD2022: The multi-national Road Damage Dataset},
  year = {2022},
  publisher = {figshare},
  doi = {10.6084/m9.figshare.21431547}
}

YOLOv7

@article{wang2022yolov7,
  title={{YOLOv7}: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors},
  author={Wang, Chien-Yao and Bochkovskiy, Alexey and Liao, Hong-Yuan Mark},
  journal={arXiv preprint arXiv:2207.02696},
  year={2022}
}

Original Project

Repository: mdptlab/roaddamagedetector2022
YOLOv7 Implementation: WongKinYiu/yolov7

License

This project is based on YOLOv7 (GPL-3.0 License) and the RDD2022 dataset. Please refer to the original repositories for licensing information.

Contributors

Team: PathFinders
Institution: Texas A&M University - Corpus Christi
Course: Capstone Project

Acknowledgments

YOLOv7 team for the excellent object detection framework
RDD2022 dataset creators for the comprehensive road damage dataset
Original roaddamagedetector2022 project for inspiration and guidance

Last Updated: November 2025

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
DetectingRoadDamage		DetectingRoadDamage
README.md		README.md

Folders and files

Latest commit

History

Repository files navigation

Road Damage Detection using YOLOv7

Table of Contents

Overview

Features

Dataset

Dataset Statistics

Model Performance

Best Model: yolov7all6 (640×640 resolution)

Per-Class Performance

Installation

Prerequisites

Step 1: Clone YOLOv7 Repository

Step 2: Replace Configuration Files

Step 3: Install Dependencies

Step 4: Fix Environment Issues (Windows)

Step 5: Download Dataset

Step 6: Download Pre-trained Weights

Dataset Preparation (REQUIRED Before Training)

Step 1: Validate Dataset Image Sizes

Step 2: Convert Dataset to YOLOv7 Format

Project Structure

Usage

Training

Model 1: yolov7all6 (640×640, 100 epochs)

Model 2: yolov7all61 (512×512, 200 epochs)

Monitor Training with TensorBoard

Resume Training (if interrupted)

Inference

Single Image/Folder Detection (Model 1)

Single Image/Folder Detection (Model 2)

Web Demo (Gradio)

Run the Demo

Features:

Ensemble Pipeline

Results

Training Progression (yolov7all6)

Sample Detections

Key Files

Troubleshooting

1. CUDA Out of Memory

2. KMP Duplicate Library Error (Windows)

3. Image Size Mismatch

4. Missing Weights

References

Dataset

YOLOv7

Original Project

License

Contributors

Acknowledgments

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages