MMIK-400K-IAC
收藏资源简介:
MMIK-400K-IAC 是一个面向工业异常认知的 32 类工业异常认知子集,用于该项目的问答(QA)与异常检测实验。它是 MMIK-400K 的子集,而不是论文中描述的包含 461,388 张图像的完整 MMIK-400K 集合。README 标注其语言为英语,任务类别包括图像分类和视觉问答,标签包括工业检测、异常检测、异常认知和开放问答。重要使用声明要求:在将本数据集用于发表论文或商业用途前,必须联系 2673679261@qq.com 并签署相关开源声明;未经许可使用者,维护方将追究盗用责任。 数据集包含 59,740 张数据图像和 5,970 个异常掩码,按物理图像文件夹、ClosedQA、OpenQA 和历史异常检测划分组织。物理图像文件夹中,训练集有 34,565 张正常图像(其中 15,360 张为增强正常图像),测试集有 25,175 张正常/异常图像。ClosedQA 训练集有 19,205 张图像,测试集有 25,175 张图像,每张图像 9 个问题;训练问答对为 172,845,测试问答对为 226,575,覆盖 9 种任务类型。OpenQA 训练集有 19,205 张图像,测试集有 25,175 张图像,覆盖 3 种任务类型;训练问答对为 57,615,测试问答对为 75,525,包含 normal_rule、deviation 和 evidence_decision。历史异常检测划分包含训练集 31,744 张图像和测试集 8,791 张图像,该划分与 QA 划分分离。 ClosedQA 文件保留九类标注问题:Anomaly Detection、Defect Classification、Defect Localization、Defect Description、Defect Analysis、Object Classification、Object Structure、Object Details 和 Object Analysis。OpenQA 包含 normal_rule、deviation 和 evidence_decision。因此数据集可用于图像分类、视觉问答、工业异常检测与异常认知,以及开放问答等任务。 文件组织方面,archives/category_01.tar 至 archives/category_32.tar 分别对应一个类别的图像与掩码文件夹;archives/manifest.json 将归档编号映射到类别名称。解压后恢复为 <category>/train/good/、<category>/test/<type>/ 和 <category>/ground_truth/<type>/。ClosedQA/Closed_train.json 与 ClosedQA/Closed_test.json 为划分后的 QA 标注,ClosedQA/Closed.json 为完整并集,不应作为训练划分使用。OpenQA/openqa_train.jsonl 与 OpenQA/openqa_test.jsonl 为划分后的开放问答数据,OpenQA/openqa_all.jsonl 为完整并集。AnomalyDetection/train.json、AnomalyDetection/test.json 和 AnomalyDetection/meta.json 构成历史异常检测比较划分。meta.json 是完整物理图像清单,与历史异常检测划分不同;domain_knowledge.json 包含类别知识记录。QA 图像路径带有 MMIK-400K-IAC/ 前缀,因此需将仓库下载到父文件夹下精确命名为 MMIK-400K-IAC 的子目录,并将类别归档解压到该目录后再训练或推理。异常检测 JSON 中的路径相对于 AnomalyDetection/ 文件夹,正常图像的掩码路径为空。 来源与使用方面,MMIK-400K 聚合来自多个工业数据源的图像,底层图像可能适用各来源的特定条款;README 未对第三方内容声明统一的一揽子许可。下游再分发前应查询原始来源与维护者。增强正常图像由原始训练图像派生,且没有合成 QA 记录。若使用该子集,应引用项目论文:Towards Industrial Anomaly Cognition: A Large-Scale Knowledge Benchmark, Challenges, and Baselines,作者 Zhang, Zekai 等,2026。
MMIK-400K-IAC is a 32-category industrial anomaly cognition subset for industrial anomaly cognition, used for question answering (QA) and anomaly detection experiments in this project. It is a subset of MMIK-400K, not the full MMIK-400K collection containing 461,388 images described in the paper. The README lists its language as English, task categories as image classification and visual question answering, and tags as industrial inspection, anomaly detection, anomaly cognition, and open QA. Important usage statement: before using this dataset for paper publication or commercial purposes, users must contact 2673679261@qq.com and sign the relevant open-source statement; unauthorized users will be pursued for misappropriation by the maintainers. The dataset contains 59,740 images and 5,970 anomaly masks, organized into a physical image folder, ClosedQA, OpenQA, and a historical anomaly detection split. In the physical image folder, the training set has 34,565 normal images (including 15,360 augmented normal images), and the test set has 25,175 normal/anomalous images. The ClosedQA training set has 19,205 images and the test set has 25,175 images, with 9 questions per image; there are 172,845 training QA pairs and 226,575 test QA pairs, covering 9 task types. The OpenQA training set has 19,205 images and the test set has 25,175 images, covering 3 task types; there are 57,615 training QA pairs and 75,525 test QA pairs, including normal_rule, deviation, and evidence_decision. The historical anomaly detection split contains 31,744 training images and 8,791 test images, and is separate from the QA splits. ClosedQA files retain nine annotated question types: Anomaly Detection, Defect Classification, Defect Localization, Defect Description, Defect Analysis, Object Classification, Object Structure, Object Details, and Object Analysis. OpenQA includes normal_rule, deviation, and evidence_decision. Therefore, the dataset can be used for tasks such as image classification, visual question answering, industrial anomaly detection and anomaly cognition, and open-ended QA. For file organization, archives/category_01.tar through archives/category_32.tar correspond to one categorys image and mask folders respectively; archives/manifest.json maps archive numbers to category names. After extraction, the structure is restored as <category>/train/good/, <category>/test/<type>/, and <category>/ground_truth/<type>/. ClosedQA/Closed_train.json and ClosedQA/Closed_test.json are the split QA annotations, while ClosedQA/Closed.json is the complete union and should not be used as a training split. OpenQA/openqa_train.jsonl and OpenQA/openqa_test.jsonl are the split open-QA data, while OpenQA/openqa_all.jsonl is the complete union. AnomalyDetection/train.json, AnomalyDetection/test.json, and AnomalyDetection/meta.json form the historical anomaly detection comparison split. meta.json is the complete physical image list and differs from the historical anomaly detection split; domain_knowledge.json contains category knowledge records. QA image paths have the MMIK-400K-IAC/ prefix, so the repository must be downloaded into a subdirectory under the parent folder exactly named MMIK-400K-IAC, and the category archives must be extracted into that directory before training or inference. Paths in the anomaly detection JSON are relative to the AnomalyDetection/ folder, and mask paths for normal images are empty. Regarding sources and usage, MMIK-400K aggregates images from multiple industrial data sources, and the underlying images may be subject to the specific terms of each source; the README does not declare a unified blanket license for third-party content. Original sources and maintainers should be checked before downstream redistribution. Augmented normal images are derived from original training images and have no synthetic QA records. If using this subset, the project paper should be cited: Towards Industrial Anomaly Cognition: A Large-Scale Knowledge Benchmark, Challenges, and Baselines, by Zhang, Zekai et al., 2026.
MMIK-400K-IAC 数据集总结
基本信息
- 数据集名称:MMIK-400K-IAC
- 语言:en
- 大小类别:
10K<n<100K - 任务类别:image-classification、visual-question-answering
- 标签:industrial-inspection、anomaly-detection、anomaly-cognition、openqa
- 数据集详情页地址:https://huggingface.co/datasets/ZZk12138/MMIK-400K-IAC
- 项目页:https://hellozzk.github.io/MMIK-400K/
- GitHub:https://github.com/hellozzk/MMIK-400K
- 论文 PDF:暂缓提供
使用声明
- 使用本数据集发表论文或商业用途前,必须联系 2673679261@qq.com 并签署相关开源声明;未经许可使用者,将追究盗用责任。
- IMPORTANT USAGE NOTICE: Before using this dataset in a research publication or for commercial purposes, you must contact 2673679261@qq.com and sign the relevant open-source declaration. Unauthorized use may result in action to pursue liability for misappropriation.
数据集概述
- MMIK-400K-IAC 是用于项目 QA 和异常检测实验的 32 类工业异常认知子集。
- 包含 59,740 数据图像和 5,970 异常掩码。
- 不是论文中描述的完整 461,388 图像 MMIK-400K 集合。
数据划分
| 组件 | 训练 | 测试 | 备注 |
|---|---|---|---|
| 物理文件夹数据图像 | 34,565 正常 | 25,175 正常/异常 | 训练包含 15,360 增强正常图像 |
| ClosedQA 图像 | 19,205 | 25,175 | 每图 9 个问题 |
| ClosedQA 问答对 | 172,845 | 226,575 | 9 种任务类型 |
| OpenQA 图像 | 19,205 | 25,175 | 3 种任务类型 |
| OpenQA 问答对 | 57,615 | 75,525 | 正常规则、偏差、证据决策 |
| 历史异常检测划分 | 31,744 图像 | 8,791 图像 | 与 QA 划分分开 |
QA 任务类型
- ClosedQA 保留九种标注问题类型:Anomaly Detection、Defect Classification、Defect Localization、Defect Description、Defect Analysis、Object Classification、Object Structure、Object Details、Object Analysis。
- OpenQA 包含:
normal_rule、deviation、evidence_decision。
文件与结构
- 类别归档:archives 目录下 category_01.tar 至 category_32.tar,每类一个 tar 归档,包含图像与掩码文件夹;archives 目录下 manifest.json 映射归档编号到类别名称。解压后恢复类别下的
train/good、test/<type>、ground_truth/<type>目录结构。 - ClosedQA:ClosedQA 目录下 Closed_train.json、Closed_test.json 为拆分 QA 标注;Closed.json 为全联合,不应用作训练划分。
- OpenQA:OpenQA 目录下 openqa_train.jsonl、openqa_test.jsonl 为拆分开放 QA;openqa_all.jsonl 为全联合。
- AnomalyDetection:AnomalyDetection 目录下 train.json、test.json、meta.json 为历史异常检测对比划分。
- 仓库根目录下 meta.json 为完整物理图像清单,与历史异常检测划分不同。
- 仓库根目录下 domain_knowledge.json 为类别知识记录。
- QA 图像路径包含数据集名称 MMIK-400K-IAC 的目录前缀。下载仓库到父文件夹中的同名子目录,然后解压类别归档。异常检测 JSON 路径相对于 AnomalyDetection 目录。正常图像掩码路径为空。
来源与引用
- MMIK-400K 聚合多个工业数据源。
- 源特定条款可能适用于底层图像;未对第三方内容声明单一通用许可。下游再分发前需咨询原始来源与维护者。
- 增强正常图像源自原始训练图像;没有合成 QA 记录。
- 若使用该子集,请引用项目论文:
bibtex @misc{zhang2026towards, title={Towards Industrial Anomaly Cognition: A Large-Scale Knowledge Benchmark, Challenges, and Baselines}, author={Zhang, Zekai and Chen, Qinghui and Chen, Da and Li, Gang and Bian, Changlin and Liu, Yunlong and Zhou, Tao and Liu, Cong and Zhang, Jinglin and Bai, Cong and Chen, Shengyong and Cohen, Laurent D.}, year={2026} }




