{}
Dataset Card for RDD_2020
The RDD2020 dataset is a comprehensive collection of 26,336 road images from India, Japan, and the Czech Republic, annotated with over 31,000 instances of road damages. This dataset is designed to support the development and evaluation of machine learning models for automatic road damage detection, offering a valuable resource for municipalities and road agencies for efficient road condition monitoring.
Dataset Details
Dataset Description
Source: Mendeley Data - DOI: 10.17632/5ty2wb6gvg.1
Size: 1.13 GB
Format: Images (JPEG) and Annotations (XML in PASCAL VOC format)
Resolution:
- India: 720 × 720 pixels
- Japan and Czech: 600 × 600 pixels
Categories: Longitudinal Cracks (D00), Transverse Cracks (D10), Alligator Cracks (D20), Potholes (D40)
Funded by [optional]: [More Information Needed]
Shared by [optional]: [More Information Needed]
Dataset Sources [optional]
- Homepage https://data.mendeley.com/datasets/5ty2wb6gvg/1
- Data article: https://doi.org/10.1016/j.dib.2021.107133
Uses
Direct Use
RDD2020 dataset can be directly used for developing and benchmarking machine learning models aimed at automatic detection and classification of road damages. This includes developing new deep learning architectures or modifying existing ones to improve detection accuracy across different types of road damages
Dataset Structure
The data will follow the structure below: { "image_id": "Czech_000248", "country": "Czech", "type": "train", "image": "<PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=600x600>", "image_path": "train/Czech/images/Czech_000248.jpg", "crack_type": ["D20", "D20"], "crack_coordinates": { "x_min": [188, 3], "x_max": [309, 171], "y_min": [463, 438], "y_max": [509, 519] } }
Dataset Creation
Curation Rationale
The RDD2020 dataset was curated with the objective of facilitating the development, testing, and benchmarking of machine learning models for road damage detection, catering specifically to the needs of municipalities and road agencies. A significant aspect of the dataset's curation process was the conversion of images into the Python Imaging Library (PIL) format and the meticulous parsing of XML annotations to ensure a seamless integration between the image data and the associated labels. This conversion process was driven by the need to simplify the handling of image data for machine learning applications, as the PIL format is widely supported by data processing and model training frameworks commonly used in the field.
Additionally, the parsing of XML files to extract detailed annotations about the type and coordinates of road damages allows for precise labeling of the data. This approach ensures that each image is directly associated with its corresponding damage type and location. The dataset's diversity, with images sourced from three different countries, aims to enable the creation of robust models that are effective across various environmental conditions and road infrastructures, thereby broadening the applicability and relevance of the trained models.
Data Collection and Processing
Road images (.jpg) were collected using a vehicle-mounted smartphone, moving at an average speed of about 40Km/h. XML files were created using the LabelImg tool to annotate the road damages present in the images.
Who are the source data producers?
Deeksha Arya, Hiroya Maeda, Sanjay Kumar Ghosh, Durga Toshniwal, Hiroshi Omata, Takehiro Kashiyama, Toshikazu Seto, Alexander Mraz, Yoshihide Sekimot
Annotations [optional]
Annotation process
Each image in the dataset comes with corresponding XML files containing annotations in PASCAL VOC format. These annotations describe the location and type of road damages present in the images, categorized into four main types: Longitudinal Cracks (D00), Transverse Cracks (D10), Alligator Cracks (D20), and Potholes (D40).
Bias, Risks, and Limitations
The dataset primarily includes images from three countries (India, Japan, and the Czech Republic), which may not fully represent road conditions worldwide. Users should be cautious when generalizing models trained on this dataset to other regions.
Recommendations
Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.