Kaputt
收藏资源简介:
Kaputt数据集是一个大规模的视觉缺陷检测数据集,由亚马逊和牛津大学应用人工智能实验室共同创建。该数据集包含超过238,421张零售物品的顶部RGB图像,以及相应的类别标签和分割掩模。数据集分为查询集和参考集,其中查询集包含100,267张图像,包括29,316个缺陷实例。每个查询图像都与最多三个未标记的参考图像相关联,这些图像展示了物品在“正常”条件下的外观。数据集涵盖了七个不同的缺陷类型,并提供了高分辨率的图像,能够捕捉到明显的和微妙的缺陷。该数据集旨在帮助研究人员开发更鲁棒和通用的视觉缺陷检测模型,以应对零售物流应用中的复杂性和细微之处。
The Kaputt Dataset is a large-scale visual defect detection dataset co-created by Amazon and the University of Oxford's Applied Artificial Intelligence Laboratory. It contains over 238,421 top-down RGB images of retail items, along with corresponding category labels and segmentation masks. The dataset is split into a query set and a reference set. The query set includes 100,267 images, encompassing 29,316 defect instances. Each query image is associated with up to three unlabeled reference images that demonstrate the item's appearance under "normal" conditions. The dataset covers seven distinct defect types and provides high-resolution images capable of capturing both obvious and subtle defects. This dataset aims to assist researchers in developing more robust and generalizable visual defect detection models to address the complexities and nuances of retail logistics applications.
Kaputt: A Large-Scale Dataset for Visual Defect Detection
数据集概述
- 专为物流场景中的缺陷检测设计的大规模数据集
- 包含超过230,000张图像,其中29,000多个缺陷实例
- 包含超过48,000个不同的物体
- 规模是MVTec数据集的40倍
研究背景
- 现有工业异常检测基准(如MVTec-AD和VisA)已趋于饱和,最优方法达到99.9% AUROC
- 零售物流中的异常检测面临新挑战:物体姿态和外观的多样性和可变性
- 现有领先的异常检测方法在此新场景下表现不佳
技术特点
- 在重度姿态和外观变化下,现有方法难以有效利用正常样本
- 在数据集上的最佳AUROC得分仅为56.96%
- 为零售物流异常检测设立了新的基准
访问信息
- 数据集下载地址:https://www.kaputt-dataset.com
- 需要通过在线表单申请访问权限
- 提交申请后会收到包含详细下载说明的电子邮件
使用条款
- 采用Creative Commons (CC BY-NC-ND 4.0)许可证
- 仅限用于计算机视觉领域的教育和科学研究目的
- 由亚马逊免费提供,"按原样"提供,不提供任何担保
引用信息
bibtex @inproceedings{kaputt2025, title = {Kaputt: A Large-Scale Dataset for Visual Defect Detection}, author = {H{"o}fer, Sebastian and Henning, Dorian and Amiranashvili, Artemij and Morrison, Douglas and Tzes, Mariliza and Posner, Ingmar and Matvienko, Marc and Rennola, Alessandro and Milan, Anton}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)}, month = {October}, year = {2025}, address = {Honolulu, Hawaii, USA}, publisher = {IEEE} }




