BFDD: A Pixel-Level Aligned RGB-IR Image Dataset for Building Façade Defect Segmentation

The Building Façade Defect Detection (BFDD) Dataset is a novel multimodal dataset specifically designed to semantic segmentation research in structural health monitoring and automated building inspection. This dataset comprises 788 strictly aligned pairs of visible (RGB) and infrared (IR) images covering 5 common façade defect categories: Cracks, Peeling, Hollow Areas, Stains, and Erosion. Each of the image pairs have been cropped to the same size: 640×512 pixels. The visible imagery captures rich surface textures and color variations, while the thermal modality reveals sub-surface delamination (hollows) and moisture-related anomalies often invisible in the RGB spectrum.

Since the raw data consist of image sequences acquired during inspection runs, where frames are captured in rapid succession along similar trajectories, in dataset partitioning, a SfM aware grouping strategy was employed (for details, please refer to the article). The image pairs were aligned using a robust feature-based registration method. This ensures strict pixel-level correspondence between the RGB textures and the IR thermal signatures, enabling rigorous training and evaluation of multimodal fusion networks. Each image in the dataset has been manually annotated in pixel level.

The dataset is publicly available to foster reproducible research in multimodal computer vision for civil infrastructure.

建筑外立面缺陷检测（Building Façade Defect Detection, BFDD）数据集是专为结构健康监测与自动化建筑巡检中的语义分割研究打造的新型多模态数据集。该数据集包含788组严格对齐的可见光（RGB）与红外（IR）图像对，涵盖5类常见外立面缺陷：裂缝、起皮、空鼓、污渍与侵蚀。所有图像对均被裁剪至统一尺寸：640×512像素。可见光图像可呈现丰富的表面纹理与色彩变化，而热成像模态则可展现RGB光谱下通常难以察觉的表层下分层（空鼓）与受潮相关异常。由于原始数据为巡检过程中采集的图像序列，各帧沿相似轨迹快速连续拍摄，因此在数据集划分阶段采用了运动恢复结构（Structure from Motion, SfM）感知分组策略（详细说明请参见相关论文）。图像对通过鲁棒的基于特征的配准方法完成对齐，确保RGB纹理与红外热特征间实现严格的像素级对应，可为多模态融合网络的严谨训练与评估提供支撑。数据集中的每幅图像均已完成像素级人工标注。本数据集已公开，旨在推动土木基础设施领域多模态计算机视觉的可复现研究。