A Curated RGB Dataset for Real-Time Road Damage Detection using YOLO

Description

This dataset consists of 1,075 high-resolution, annotated RGB images designed to support the development and benchmarking of real-time object detection models for road infrastructure maintenance. It is specifically curated to facilitate the detection of surface defects using the YOLO (You Only Look Once) architecture, contributing to automated pavement distress analysis and smart city monitoring systems. Dataset Composition The dataset is categorized into three primary classes of road distress, representing the most critical targets for structural assessment: ⦁ Crack: Longitudinal, transverse, or alligator cracking patterns in the asphalt surface. ⦁ Pothole: Significant structural cavities or depressions in the pavement. ⦁ Surface Erosion: General degradation, stripping, or localized wear of the top pavement layer. Technical Specifications ⦁ Total Images: 1,075 ⦁ Image Format: 3-channel RGB (Color) ⦁ Resolution: Harmonized to 640 X 640 pixels for optimal YOLO training efficiency. ⦁ Annotation Format: Provided in YOLO-compliant `.txt` format (normalized coordinates). Data Split and Preprocessing To ensure robust model evaluation and prevent overfitting, the dataset is partitioned into a training and validation framework: ⦁ Training Set (914 images): Used for supervised feature extraction and weight optimization. ⦁ Validation Set (161 images): An independent subset used for hyperparameter tuning and performance monitoring during training. ⦁ Preprocessing: Images have been resized and normalized to $640 \times 640$, maintaining aspect ratio through padding where necessary to ensure spatial consistency for the detector’s backbone. Applications This dataset is optimized for: ⦁ Real-time Object Detection: Development of lightweight YOLO models for deployment on edge devices or mobile inspection vehicles. ⦁ Pavement Management Systems (PMS): Automating the identification and quantification of road damage to prioritize repair schedules. ⦁ Edge Computing Benchmarking: Testing the inference speed and accuracy trade-offs of various YOLO versions (e.g., YOLOv8, YOLOv10, YOLOv12) on resource-constrained hardware.

Steps to reproduce

1. Download the dataset: Access the dataset files from this Mendeley Data repository and extract the contents into your working directory. 2. Dataset structure: Ensure the dataset is structured in a standard YOLO format, separating images and their corresponding `.txt` annotation files into the predefined train and validation splits: Road_Damage_Dataset/ ├── train/ │ ├── images/ │ │ ├── crack_001.jpg │ │ ├── ... │ ├── labels/ │ │ ├── crack_001.txt │ │ ├── ... ├── val/ │ ├── images/ │ │ ├── pothole_101.jpg │ │ ├── ... │ ├── labels/ │ │ ├── pothole_101.txt │ │ ├── ... └── dataset.yaml (Note: The `dataset.yaml` file should define the relative paths to the train/val image directories, the number of classes (3), and the class names: Crack, Pothole, Surface Erosion). 3. Data preprocessing: ⦁ Verify all images are resized to the harmonized 640x640 pixel dimension (maintaining aspect ratio via padding if required). ⦁ Ensure bounding box annotations in the `.txt` files are normalized (values between 0 and 1) in the standard YOLO format: `class_id center_x center_y width height`. ⦁ Utilize the predefined split of 914 images for training and 161 images for validation. 4. Model training example (optional): ⦁ Set up a YOLO training environment using a framework like PyTorch (via the Ultralytics library). ⦁ Apply spatial and color-level data augmentations (such as mosaic, scaling, horizontal flipping, and HSV adjustments) to improve model robustness to varying road and lighting conditions. ⦁ Initialize training for a real-time object detection architecture like YOLOv8, YOLOv10, or YOLOv12, pointing the training script to your `dataset.yaml` configuration. 5. Evaluation: ⦁ Evaluate the model using standard object detection metrics: Precision, Recall, and mean Average Precision (mAP@0.5 and mAP@0.5:0.95). ⦁ Run inference on the validation set to generate bounding box predictions overlaid on the road imagery. ⦁ Review the generated confusion matrix and precision-recall curves to analyze the model's accuracy across the specific distress classes and identify any misclassifications (e.g., confusing severe cracking with surface erosion).

Institutions

• Daffodil International University

  Dhaka Division, Dhaka

Categories

Licence