Docs Datasets for Road Damage Detection

Datasets for Road Damage Detection

← Back to AI Applications

Overview

Training an accurate road damage detection model requires large, diverse, and well-annotated datasets. This page documents the dataset used in this project and provides information on public datasets for extending the system to detect additional damage types.

Current Project Dataset

Pothole Detection Dataset (Roboflow)

Source: Roboflow Universe - Pothole Detection

Dataset Statistics:

• Total Images: 4,510
• Train Set: 3,993 images (88.5%)
• Validation Set: 352 images (7.8%)
• Test Set: 165 images (3.7%)

Classes: 1 class - pothole (bowl-shaped depressions and holes in road surface)

Note: The source Roboflow project contains 8 classes (pothole, curb, dash, distressed, grate, manhole, marking, utility), but this dataset version (v7) is filtered to only include pothole annotations for focused single-class detection.

Preprocessing Applied

The dataset has been preprocessed with the following operations:

Operation	Description
Auto-Orient	Corrects image orientation based on EXIF data
Resize	Stretch to 840×840 pixels
Auto-Adjust Contrast	Histogram equalization for improved visibility
Filter Null	Requires ≥20% of images to contain annotations

Augmentations Applied

To increase dataset diversity and model robustness:

Augmentation	Details
Outputs per training example	3× (triples the training data)
Grayscale	Applied to 23% of images
Blur	Up to 2.5px blur radius
Noise	Up to 1.76% of pixels affected

Format: YOLO format (text files with normalized bounding box coordinates)

Access: The dataset is managed through Roboflow API. Configure your .env file with appropriate credentials to automatically download during training (see AI Applications Setup).

Using the Project Dataset

Quick Start with Roboflow

The training script (ai-model/train/train.py) automatically downloads the dataset from Roboflow. Configure your environment:

# 1. Copy environment template
cd ai-model
cp .env.example .env

# 2. Edit .env with your Roboflow credentials
ROBOFLOW_API_KEY=your_api_key_here
ROBOFLOW_WORKSPACE=jerry-cooper-tlzkx
ROBOFLOW_PROJECT=pothole_detection-hfnqo
ROBOFLOW_PROJECT_NAME=pothole_detection
ROBOFLOW_PROJECT_VERSION=7

# 3. Run training (dataset downloads automatically)
cd train
python train.py

Dataset Structure

After download, the dataset follows YOLO format:

dataset/
├── train/
│   ├── images/
│   │   ├── image_0001.jpg
│   │   └── ...
│   └── labels/
│       ├── image_0001.txt  # YOLO annotations
│       └── ...
├── valid/
│   ├── images/
│   └── labels/
├── test/
│   ├── images/
│   └── labels/
└── data.yaml  # Dataset configuration

YOLO Label Format (image_0001.txt):

# class_id center_x center_y width height (all normalized 0-1)
0 0.512 0.634 0.124 0.089

Where class_id is 0 for pothole (single class detection).

Extending to Additional Damage Types

To train the model on additional road damage types (cracks, rutting, etc.), you would need to collect and annotate a new dataset with those damage classes. The system architecture supports this through:

1. Data Collection: Capture images containing the desired damage types
2. Annotation: Use tools like Roboflow, LabelImg, or CVAT to annotate bounding boxes
3. Training: Run train.py with the new dataset configuration
4. Export: Convert the trained model to TFLite for edge deployment

The current Roboflow dataset management system makes it easy to extend the model’s capabilities as new annotated data becomes available.

Next Steps

1. Setup AI Environment - Configure training environment
2. Train Pothole Detection Model - Start training
3. Validate Performance - Evaluate model

---

← Back to AI Applications