USD: Underwater Structure Defect Dataset with Suspended Impurities
收藏资源简介：
The USD dataset is designed for underwater structure surface defect detection under suspended-impurity interference. It consists of four subsets: USD-SSI, USD-RSI, USD-VDO, and USD-VDW. USD-SSI contains 16 sets of underwater structure videos with synthetically generated suspended impurities. Impurity samples of diverse sizes, shapes, and colors were extracted from real underwater footage and inserted into original videos using a custom random motion generation program. Corresponding impurity mask videos and manually labeled defect annotations are provided. USD-RSI comprises 14 sets of real-world underwater structural videos containing naturally occurring suspended impurities, providing a benchmark for evaluating detection algorithms under realistic underwater conditions. USD-VDO is an annotated underwater structure defect dataset containing 541 images without suspended impurities, along with manually labeled ground truth annotations of defects. USD-VDW consists of 70 frame sets randomly selected from USD-RSI. It includes corresponding E-procedure processed results and manually labeled defect ground truth, making it suitable for evaluating the complete E-D cascaded framework. The dataset can be used for suspended impurity localization, underwater structure defect detection, defect segmentation, and evaluation of robust visual perception methods under complex underwater environments. GitHub repository: https://github.com/your-username/USD-Dataset
USD数据集专为悬浮杂质（suspended impurities）干扰下的水下结构表面缺陷检测任务设计，共包含四个子集：USD-SSI、USD-RSI、USD-VDO与USD-VDW。 USD-SSI包含16组合成了悬浮杂质的水下结构视频。研究人员从真实水下影像中提取了尺寸、形状与颜色各异的杂质样本，并通过自研的随机运动生成程序将其嵌入原始视频中，同时提供了对应的杂质掩码（mask）视频以及人工标注的缺陷标注文件。 USD-RSI包含14组带有天然悬浮杂质的真实水下结构视频，可为在真实水下环境中评估检测算法提供基准测试数据集（benchmark）。 USD-VDO为带标注的水下结构缺陷数据集，包含541张无悬浮杂质的图像，以及人工标注的缺陷真值（ground truth）标注。 USD-VDW包含从USD-RSI中随机抽取的70帧序列（frame sets），附带对应的E-procedure处理结果与人工标注的缺陷真值，适用于评估完整的E-D级联框架（E-D cascaded framework）。 该数据集可应用于悬浮杂质定位、水下结构缺陷检测、缺陷分割，以及复杂水下环境下鲁棒视觉感知（robust visual perception）方法的性能评估。 GitHub仓库地址：https://github.com/your-username/USD-Dataset
