A Benchmark Image Dataset of Reinforced Concrete Beam–Column Joint Failures

Reinforced concrete beam–column joints are critical structural components whose damage patterns provide important evidence for diagnosing structural failure mechanisms. However, publicly available visual datasets dedicated to beam–column joint failures remain highly limited, especially those with expert-verified structural annotations. We introduce a curated beam–column joint image dataset with expert-verified annotations for structural damage analysis. Each sample is associated with categorical metadata describing joint configuration, failure mechanism, damage type, and graded damage severity, together with expert-written textual diagnostic descriptions. The dataset contains 572 beam–column joint images and 1,716 image–text diagnostic description pairs. To demonstrate the technical usability of the dataset, we provide baseline experiments for several representative tasks, including failure-mode classification, image-to-text diagnostic description generation, and text-conditioned image generation. Experimental results show that the dataset supports reproducible benchmarking across both vision and vision–language modeling tasks. Overall, the dataset provides a standardized benchmark for structural damage recognition, multimodal representation learning, and generative modeling in structural engineering.钢筋混凝土梁柱节点（Reinforced concrete beam–column joints）是关键结构构件，其破坏模式可为结构失效机制的诊断提供重要依据。然而，当前专门面向梁柱节点破坏的公开可视化数据集仍极度匮乏，尤其是带有经专家核验的结构标注的数据集。我们构建了一套经整理筛选的梁柱节点图像数据集，附带专家核验的结构标注以用于结构损伤分析。每个样本均关联有分类元数据，涵盖节点构造形式、失效机制、损伤类型以及损伤严重程度分级，同时附带专家撰写的文本诊断描述。该数据集包含572张梁柱节点图像与1716组图像-文本诊断描述配对。为验证该数据集的技术可用性，我们针对若干典型任务开展了基准实验，包括破坏模式分类、图像到文本的诊断描述生成以及文本条件图像生成。实验结果表明，该数据集可支撑计算机视觉与视觉语言建模任务的可复现基准测试。总体而言，本数据集为结构工程领域的结构损伤识别、多模态表征学习以及生成建模提供了标准化基准。