MVTec-FS

该数据集是一个少样本缺陷分类数据集，包含了来自MVTec AD数据集的1228张缺陷图像，这些图像带有实例级别的遮罩标签，覆盖了46种缺陷类型。在数据划分上，50%的缺陷图像被用作训练集，剩余的50%则构成测试集。为了创建该数据集，研究者们使用连通组件算法将图像级别的遮罩转换为实例级别的遮罩。该数据集的规模为1228张图像，包含46种缺陷类型，每种类型有9到58个实例。所面临的任务是少样本缺陷分类。

This is a few-shot defect classification dataset. It contains 1228 defect images sourced from the MVTec AD dataset, each with instance-level mask labels and covering 46 defect types. For data partitioning, 50% of the defect images are allocated to the training set, while the remaining 50% form the test set. To construct this dataset, researchers used the connected component algorithm to convert image-level masks into instance-level masks. The dataset has a total of 1228 images across 46 defect types, with each type containing 9 to 58 instances. The target task of this dataset is few-shot defect classification.

搜集汇总

背景与挑战

背景概述

MVTec-FS数据集是基于MVTec AD数据集改进的版本，专门为少样本学习研究设计，特别是针对缺陷分类任务。它提供了实例级注释，通过连接组件算法和手动细化生成，覆盖14个子数据集，支持1-shot、3-shot和5-shot学习范式，适用于缺陷分类、目标检测和多模态分类等应用。