Source-linked AI summary
RDD2022: A multi-national image dataset for automatic Road Damage Detection
Deeksha Arya, Hiroya Maeda, Sanjay Kumar Ghosh, Durga Toshniwal, Yoshihide Sekimoto
TL;DR
Road-damage detection needs data spanning multiple countries and heterogeneous road conditions. RDD2022 addresses this by combining six-country imagery with standardized annotations and challenge-ready data organization. The dataset supports automatic detection, low-cost road inspection, and benchmarking of related image-based algorithms, while remaining incomplete across all climatic and geographical conditions.
Problem
Existing road-damage data had limited country coverage, motivating a dataset that supports detection across six countries and broader conditions.
Method
RDD2022 constructs a six-country road-image dataset with annotated damage instances, using LabelImg or CVAT and bounding-box labels stored in XML format.
Results
RDD2022 provides data for automatic road-damage detection, low-cost road-condition inspection, challenge evaluation, and benchmarking related image-based algorithms.
Takeaways & Limitations
Municipalities, road organizations, and machine-learning researchers can use RDD2022 to develop, evaluate, and benchmark road-damage detection systems.
Takeaways & Limitations
RDD2022 does not exhaustively represent all climatic and geographical conditions across the six countries, so additional images are needed for more heterogeneous scenarios.
Abstract
from arXiv · showhide
The data article describes the Road Damage Dataset, RDD2022, which comprises 47,420 road images from six countries, Japan, India, the Czech Republic, Norway, the United States, and China. The images have been annotated with more than 55,000 instances of road damage. Four types of road damage, namely longitudinal cracks, transverse cracks, alligator cracks, and potholes, are captured in the dataset. The annotated dataset is envisioned for developing deep learning-based methods to detect and classify road damage automatically. The dataset has been released as a part of the Crowd sensing-based Road Damage Detection Challenge (CRDDC2022). The challenge CRDDC2022 invites researchers from across the globe to propose solutions for automatic road damage detection in multiple countries. The municipalities and road agencies may utilize the RDD2022 dataset, and the models trained using RDD2022 for low-cost automatic monitoring of road conditions. Further, computer vision and machine learning researchers may use the dataset to benchmark the performance of different algorithms for other image-based applications of the same type (classification, object detection, etc.).
Background and Summary
RDD2022 extends earlier road-damage datasets to six countries, motivated by improved model generalizability and broader coverage. It provides four standardized damage categories and country-wise dataset statistics for challenge-based detection research.
- RDD2018 evolved through annotation correction, GAN-based augmentation in RDD2019, and added multi-country data in RDD2020.
- The multi-country setting required excluding road-marking categories whose assessment criteria vary across countries and retaining four damage categories.
- RDD2022 extends RDD2020 to target road-damage detection across India, Japan, the Czech Republic, Norway, China, and the United States.
- RDD2022 is motivated by evidence that adding countries improves model generalizability and by the success of the GRDDC’2020 challenge.
- RDD2022 is released with annotated training images and unannotated test images for the CRDDC’2022 challenge, with country-wise and category-wise distributions reported.
- The dataset documents heterogeneous coverage across countries, including varied municipalities, road classes, environments, and China’s drone and motorbike imagery.
Methods
RDD2022 was assembled through country-specific image acquisition and a standardized road-damage annotation process. Images were collected using smartphones, specialized cameras, Street View, motorbikes, and drones, then labeled with bounding boxes and class labels.
- Data acquisition and annotation: RDD2022 preparation comprised image acquisition and damage annotation across six countries.Road images were captured and labeled using software to mark road damage.
- Data acquisition and annotation: Smartphone-mounted vehicles collected images in India, Japan, and the Czech Republic.Japan and the Czech Republic used 600x600 images; India used a different stated resolution.
- Data acquisition and annotation: Norway used two Basler_Ace2040gc cameras in a ViaPPS vehicle to create stitched wide-view images typically sized 3650x2044.The cameras used CMOS sensors and were mounted inside the vehicle windshield.
- Data acquisition and annotation: The United States used 640x640 Google Street View images, while China used motorbikes and DJI M600 Pro UAVs for pavement collection.China_MotorBike captured images at an average speed of 30km/h; the UAV produced China_Drone data.
- Data acquisition and annotation: China_Drone was included only in training to increase heterogeneity, while the dataset retained a low-cost automatic road-damage-detection goal.The stated objective emphasizes feasible methods for public use.
- Data acquisition and annotation: LabelImg annotated most images and CVAT annotated Norway images; recognized damage was enclosed by bounding boxes and assigned class labels stored in PASCAL VOC-like XML.Bounding boxes used xmin, ymin, xmax, ymax coordinates.
Data Records
RDD2022 is organized into country- and acquisition-specific folders containing training images with XML annotations. Most folders also include unlabeled test images, while China_Drone has no test directory.
- Repository organization: The dataset repository is organized into seven folders with a directory structure shown in Figure 14.The folders separate the China_Drone, China_MotorBike, Czech, India, Japan, and other country-specific data.
- Repository organization: China_Drone and China_MotorBike contain China data collected by drones and motorbikes, respectively.The repository includes sample images for both acquisition types.
- Repository organization: Czech, India, and Japan folders contain data collected using vehicle-mounted smartphones.Sample images are provided for these country folders.
- Repository organization: Each folder contains a train subfolder with JPG images and XML annotations, while test images are provided in every folder except China_Drone.Test annotations are not released; model predictions can be assessed through the CRDDC’2022 leaderboards.
Usage Notes
RDD2022 supports direct use and augmentation for road-damage research, broader image-based applications, benchmarking, and low-cost road-condition inspection. Its vehicle-mounted imagery and annotations enable model development, challenges, and extensions across multiple use cases.
- RDD2022 follows the PASCAL VOC format, allowing existing image-processing methods to be applied.
- Municipalities and road organizations can use RDD2022 as a foundation for smartphone-based, low-cost road-condition inspection.
- The dataset supports new deep convolutional architectures, modified existing architectures, and benchmarking algorithms for classification and object detection.
- Researchers can train, validate, and test algorithms for automatically identifying road damage across six countries.
- Vehicle-mounted images support damage detection from moving vehicles and quick inspection of large areas.
- Users may extend RDD2022 with new images or annotations for additional countries, road types, damage categories, segmentation, road assets, and traffic applications.