PDD2026: A Pixel-Level Annotated Dataset for Road Pavement Deformation Segmentation

Description

The Pavement Deformation Dataset (PDD2026) was developed to support pavement deformation analysis, pavement condition assessment, and automated road inspection in real-world environments. The dataset is designed for pixel-level pavement deformation segmentation and contains five pavement deformation categories: Depression, Rutting, Upheaval, Pothole, and Patching. Polygon-based annotations enable accurate delineation of pavement deformation boundaries and support quantitative pavement analysis. RGB pavement images were collected from asphalt roads across Lampung Province, Indonesia, including Bandar Lampung City, South Lampung, East Lampung, Central Lampung, Metro City, and Pesawaran Regency. Data acquisition was conducted using vehicle-mounted Logitech C925e Webcam, Intel RealSense D455 RGB-D camera, GoPro HERO10, and a Samsung Galaxy A55 smartphone. Images were captured at 2-second intervals while the survey vehicle traveled at approximately 10–40 km/h. Additional still images and representative video frames were acquired using the GoPro HERO10 and Samsung Galaxy A55. Data collection was performed under diverse road, weather, and illumination conditions using oblique-view camera configurations (30°–45°) with approximately 4–8 m pavement coverage. The use of multiple devices, camera positions, acquisition distances, viewpoints, illumination conditions, vehicle speeds, and road environments introduced substantial visual and geometric variability, improving the representativeness of real-world pavement inspection scenarios. A total of 6,332 RGB pavement images were collected and manually screened to ensure image quality and annotation suitability. Only images containing visible pavement deformation were retained. Since a single image may contain multiple pavement deformation instances, polygon annotations were manually created using CVAT according to predefined class definitions. The retained images were standardized to a resolution of 1024 × 1024 pixels while preserving the original RGB information. No image enhancement or data augmentation was applied. The dataset follows the standard YOLO segmentation directory structure and comprises 4,450 training (70%), 1,282 validation (20%), and 600 test (10%) images. The dataset was partitioned using a random image-level split. Although some images may originate from the same acquisition route or recording session, variations in acquisition devices, viewpoints, illumination conditions, and road environments reduce the likelihood of near-duplicate images appearing across different subsets. The repository contains RGB images, polygon annotation files in YOLO segmentation format, and a data.yaml configuration file defining dataset paths and class names. PDD2026 provides a publicly available resource for benchmarking pavement deformation segmentation models and supporting pavement condition assessment, infrastructure monitoring, and related computer vision research.

Steps to reproduce

To reproduce the dataset preparation and annotation workflow, users may follow the steps below. (1) Download the PDD dataset repository and extract all files while preserving the original directory structure. The repository contains the images, labels, and data.yaml files required for semantic segmentation experiments. (2) Verify the dataset organization by ensuring that the train, validation, and test subsets are correctly located in the images and labels directories. The dataset follows the standard YOLO segmentation format and can be directly used by supported deep learning frameworks. (3)Load the dataset using a YOLO-compatible semantic segmentation framework by specifying the provided data.yaml configuration file. The dataset contains 4,450 training, 1,282 validation, and 600 test images with polygon annotations for five pavement deformation categories. (4) To reproduce the annotation process or create new annotations, import the RGB pavement images into CVAT and create a segmentation task using the predefined classes: Depression, Rutting, Upheaval, Pothole, and Patching. Polygon annotations should be manually delineated along the visible deformation boundaries. (5) After annotation, export the annotations from CVAT using the YOLO Segmentation format. Replace or merge the exported annotation files with the existing labels directory while maintaining identical filenames between images and annotation files. (6) Additional pavement deformation categories may be incorporated by defining new class identifiers in both the CVAT project and the data.yaml configuration file. Newly annotated images can then be added to the corresponding train, validation, or test directories while preserving the original dataset structure. (7) The dataset can subsequently be used for semantic segmentation benchmarking or reannotated to support other computer vision applications, provided that the directory organization and annotation format remain consistent.

Institutions

• University of Indonesia

  West Java, Depok

Categories

Licence