Road Surface Irregularities Dataset

The dataset consists of 18,542 manually annotated images of road surface defects and traffic-calming elements, including potholes, manholes, and speed-reducing bumps. Data were collected from multiple sources to ensure high visual diversity, including a subset of the Road Pothole Images Dataset (left-side traffic conditions) and a large portion recorded in Vilnius, Lithuania, under right-side traffic conditions. This combination improves robustness to traffic-side configuration and environmental variability.

Images were acquired using different monocular cameras, including an FHD dashcam, a high-resolution mobile phone camera (30 FPS), and a ZED2 stereo camera (single-lens images used for 2-D detection). The use of multiple devices introduces variations in resolution, field of view, exposure, and compression artifacts, enhancing detector generalization.

The dataset is split into 15,827 training, 1,358 validation, and 1,357 testing images. Three object classes are defined: potholes, manholes, and speed-reducing bumps. Bumps are not subclassified during detection; geometric differentiation is performed later based on estimated physical dimensions. All bounding boxes were manually annotated following a consistent labeling policy using Roboflow and Label Studio.

Subset of this dataset images is taken from (S. Nienaber et al. 2015) open access dataset of South African Potholes, but they were reannotated for 3 classes, therefore referencing it. Also, some images were taken from random Roboflow datasets. Please cite their research too if using this dataset:

[1] S. Nienaber, M.J. Booysen, R.S. Kroon, “Detecting potholes using simple image processing techniques and real-world footage”, SATC, July 2015, Pretoria, South Africa.
[2] S. Nienaber, R.S. Kroon, M.J. Booysen , “A Comparison of Low-Cost Monocular Vision Techniques for Pothole Distance Estimation”, IEEE CIVTS, December 2015, Cape Town, South Africa.

The publication about current dataset is not yet published and this description should be updated after the publicitation. Finally we will ask to cite our publication after it is published if using this dataset, or comparing your results to our results. There will be also Github repository available with training, detection, validation codes and actual application in irregularities detection, characterization and tracking, after publication the link will be added to this description