CRACK500 with noisy annotation masks

This is the repository to host the dataset to train noisy labelled crack segmentation algorithm proposed by Zhang, et al. (upcoming) in the article "SelectSeg: Uncertainty-based selective training and prediction for accurate crack segmentation under limited data and noisy annotations".

In the zip folder, you can find "train_crop_mask_(10,20,50,100)pct", corresponding to the case where 10%, 20%, 50%, and 100% masks have been replaced by their noisy version.

For other files, including training images, and validation/test dataset, we refer the interested readers to https://github.com/fyangneil/pavement-crack-detection for the original development.

本仓库用于托管Zhang等人即将发表的论文《SelectSeg：有限数据与带噪标注下面向精准裂缝分割的基于不确定性的选择性训练与预测》中提出的带噪标注裂缝分割算法的训练数据集。在该压缩包中，您可找到名为`train_crop_mask_(10,20,50,100)pct`的文件，其分别对应10%、20%、50%及100%掩码被替换为带噪版本的实验场景。其余文件（包括训练图像、验证/测试数据集）的相关信息，敬请有兴趣的读者前往原始开发仓库查阅：https://github.com/fyangneil/pavement-crack-detection

搜集汇总

背景与挑战

背景概述

该数据集是用于训练带有噪声标注的裂缝分割算法的资源，包含不同比例（10%、20%、50%、100%）的噪声标注掩码版本，旨在支持在有限数据和噪声标注下进行准确的裂缝分割研究。数据集属于计算机视觉和结构工程领域，由香港科技大学发布，适用于图像分割任务。