Concrete Building Surface Defects Semantic Segmentation Dataset

This repository provides a publicly available subset of the concrete surface defect segmentation dataset used in our paper “A Mix Transformer Enhanced UNet for Concrete Building Surface Defects Semantic Segmentation”.

  This study utilizes several publicly available concrete defect image datasets, including CODEBRIM (Concrete Defect Bridge Image Dataset), MRBD (Multi-classifier for RC Bridge Defects), and CCCD (Concrete Crack Conglomerate Dataset). These datasets provide a rich source of samples, particularly in the areas of concrete bridge and conglomerate crack defects. To ensure the comprehensiveness of the study, we supplemented the datasets with additional samples collected from the field and online public resources, addressing the underrepresented defect types. We sincerely thank the authors and communities who have contributed and shared these publicly available datasets, as their valuable resources have greatly advanced research in concrete building surface defects semantic segmentation.

Mundt, M., Majumder, S., Murali, S., Panetsos, P., & Ramesh, V. (2019). CODEBRIM: COncrete DEfect BRidge IMage Dataset (Version 1.0). Zenodo. https://doi.org/10.5281/zenodo.2620293

Hüthwohl, P., Lu, R., & Brilakis, I. (2019). Multi-classifier for reinforced concrete bridge defects. Automation in Construction, 105, 102824. https://doi.org/10.1016/j.autcon.2019.04.019

Bianchi, E; Hebdon, M. Concrete Crack Conglomerate Dataset. University Libraries, Virginia Tech. Dataset. https://doi.org/10.7294/16625056.v1.

  Due to data-sharing restrictions, the full dataset cannot be released. We therefore construct and release a subset consisting of 200 pixel-level annotated images. The subset is carefully selected to preserve the defect occurrence frequencies and the class distribution of the original dataset, maximizing consistency with the full dataset in terms of structural characteristics and class imbalance, while allowing for minimal deviations during the sampling process. 

  The dataset includes four defect categories—crack, weathering, spalling, and exposed reinforcement—with dense pixel-wise annotations. The relative proportions of each defect type in the subset are consistent with those in the original dataset, ensuring that the structural characteristics of the full dataset are faithfully retained.

  This subset is intended to support reproducibility, benchmarking, and comparative evaluation of segmentation models. It serves as a representative proxy for the full dataset rather than a replacement.