MMIOC-1M-Curated-334
收藏资源简介:
MMIOC-1M Curated 334 是 MMIOC-1M 工业检测基准的精选 active train/validation 子集,关联论文《Unification of Closed-Open Industrial Detection Scenarios: New Large-Scale Benchmarks, Challenges and Baselines》(arXiv:2606.07953)。该发布不是论文中完整的 1M+ 样本基准,而是面向目标检测任务的训练与验证数据,任务类别为 object-detection,标签涉及 image、industrial-inspection 和 anomaly-detection,数据语言为英文。数据集总计包含 382,643 张图像、942,306 个 COCO bounding boxes 和 334 个 active defect categories。具体划分为:Train 有 300,990 张图像、742,463 个 COCO boxes、300,990 个 YOLO TXT labels;Validation 有 81,653 张图像、199,843 个 COCO boxes、81,653 个 YOLO TXT labels;Total 为 382,643 张图像、942,306 个 COCO boxes、382,643 个 YOLO TXT labels。类别列表 leibie.txt 按 YOLO 顺序 0–333 列出 334 个 active classes;COCO 保留原始 category IDs,读取 COCO 标注时应使用每个 COCO JSON 中的 category table;三个没有 active annotations 的类别未出现在发布类别列表中。文件方面,图像及对应 TXT 标签按编号 tar shards 打包在 archives/ 下以提高下载效率,COCO JSON 文件和 leibie.txt 可单独获取,archives/manifest.json 列出每个 shard、split 和图像数量。解压全部 shards 后可重建 train/images/*.jpg、train/labels/*.txt、train/annotations/train.json、val/images/*.jpg、val/labels/*.txt、val/annotations/val.json 和 leibie.txt。COCO image record 中的 file_name 在对应 train/images/ 或 val/images/ 下解析,同名 TXT 标签位于匹配的 labels/ 文件夹。README 提供下载与提取示例,使用 huggingface_hub.snapshot_download 下载 repo_id 为 ZZk12138/MMIOC-1M-Curated-334 的 dataset,并解压 archives/*.tar。发布披露指出,维护者在发布前移除了被识别为可能具有 conflicts of interest 的数据,相关标注也一并移除;发布图像集由 active train/annotations/train.json 和 val/annotations/val.json 记录定义,本地源文件夹中仍存在但不在这些记录中的图像不属于本发布;由于可用清理记录还包含其他标注质量排除,未单独给出利益冲突移除数量。范围与引用方面,本发布只包含 detection train/validation 数据和 leibie.txt,不包含项目单独的 QA 数据、权重或代码;源项目为 hellozzk/MMIO,使用该子集时应引用 arXiv:2606.07953。组件图像可能适用源特定权利,本卡不授予第三方源材料的一揽子许可。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,维护者将追究盗用责任。
MMIOC-1M Curated 334 is a curated active train/validation subset of the MMIOC-1M industrial inspection benchmark, associated with the paper Unification of Closed-Open Industrial Detection Scenarios: New Large-Scale Benchmarks, Challenges and Baselines (arXiv:2606.07953). This release is not the complete 1M+ sample benchmark from the paper, but training and validation data for object detection. Its task category is object-detection, with tags including image, industrial-inspection, and anomaly-detection, and the data language is English. The dataset contains 382,643 images, 942,306 COCO bounding boxes, and 334 active defect categories. Splits are: Train has 300,990 images, 742,463 COCO boxes, and 300,990 YOLO TXT labels; Validation has 81,653 images, 199,843 COCO boxes, and 81,653 YOLO TXT labels; Total has 382,643 images, 942,306 COCO boxes, and 382,643 YOLO TXT labels. The class list leibie.txt lists 334 active classes in YOLO order 0–333; COCO retains original category IDs, and users should use the category table in each COCO JSON when reading COCO annotations; three categories without active annotations are not included in the released class list. Files: images and corresponding TXT labels are packaged as numbered tar shards under archives/ to improve download efficiency; COCO JSON files and leibie.txt are available separately; archives/manifest.json lists each shard, split, and image count. After extracting all shards, users can reconstruct train/images/*.jpg, train/labels/*.txt, train/annotations/train.json, val/images/*.jpg, val/labels/*.txt, val/annotations/val.json, and leibie.txt. The file_name in COCO image records resolves under the corresponding train/images/ or val/images/, and same-named TXT labels are in the matching labels/ folder. The README provides download and extraction examples using huggingface_hub.snapshot_download with repo_id ZZk12138/MMIOC-1M-Curated-334 and extracting archives/*.tar. The release disclosure states that the maintainers removed data identified as potentially having conflicts of interest before release, and related annotations were also removed; the released image set is defined by active train/annotations/train.json and val/annotations/val.json records, and images still present in local source folders but not in these records are not part of this release; because available cleaning records also include other annotation quality exclusions, the number of conflict-of-interest removals is not separately given. In terms of scope and citation, this release only includes detection train/validation data and leibie.txt, and does not include the projects separate QA data, weights, or code; the source project is hellozzk/MMIO, and this subset should be cited as arXiv:2606.07953. Component images may be subject to source-specific rights, and this card does not grant a blanket license to third-party source materials. The usage statement requires contacting 2673679261@qq.com and signing the relevant open-source statement before using this dataset for published papers or commercial purposes; unauthorized users will be pursued by the maintainers for misappropriation.
MMIOC-1M Curated 334 数据集总结
基本信息
- 数据集页面:https://huggingface.co/datasets/ZZk12138/MMIOC-1M-Curated-334
- pretty_name:MMIOC-1M Curated 334
- 语言:en
- 规模类别:100K<n<1M
- 任务类别:object-detection
- 标签:image、industrial-inspection、anomaly-detection、arxiv:2606.07953
- 关联论文:Unification of Closed-Open Industrial Detection Scenarios: New Large-Scale Benchmarks, Challenges and Baselines
- 论文链接:https://huggingface.co/papers/2606.07953、https://arxiv.org/abs/2606.07953
发布内容
- 本仓库发布与上述论文相关的 MMIOC-1M 工业检测基准的 curated active train and validation 子集。
- 包含 382,643 张图像、942,306 个 COCO 边界框、334 个有效缺陷类别。
- 本发布不是论文中描述的完整 1M+ 样本基准。
数据统计
| Split | Images | COCO boxes | YOLO TXT labels |
|---|---|---|---|
| Train | 300,990 | 742,463 | 300,990 |
| Validation | 81,653 | 199,843 | 81,653 |
| 总计 | 382,643 | 942,306 | 382,643 |
类别与标注
- 类别列表文件:https://huggingface.co/datasets/ZZk12138/MMIOC-1M-Curated-334/blob/main/leibie.txt,列出 334 个有效类别,按 YOLO 顺序 0–333。
- COCO 保留原始类别 ID;读取 COCO 标注时应使用每个 COCO JSON 中的类别表。
- 三个无有效标注的类别未出现在发布类别列表中。
发布披露
- 维护者在发布前移除了被识别为可能具有利益冲突的数据,相关标注也随图像移除。
- 发布图像集由有效的训练与验证 COCO 标注记录定义:https://huggingface.co/datasets/ZZk12138/MMIOC-1M-Curated-334/blob/main/train/annotations/train.json、https://huggingface.co/datasets/ZZk12138/MMIOC-1M-Curated-334/blob/main/val/annotations/val.json。
- 本地源文件夹中仍存在但不在上述记录中的图像不属于本次发布。
- 没有专门归因于利益冲突移除的数量,因为可用清理记录还包括其他标注质量排除项。
文件组织与下载
- 图像和对应 TXT 标签按编号 tar 分片打包在数据集仓库的 archives 目录下:https://huggingface.co/datasets/ZZk12138/MMIOC-1M-Curated-334/tree/main/archives。
- COCO JSON 文件和类别列表可单独访问。
- 分片清单文件:https://huggingface.co/datasets/ZZk12138/MMIOC-1M-Curated-334/blob/main/archives/manifest.json,列出每个分片、split 和图像数量。
- 解压全部分片后,将形成训练/验证图像、YOLO TXT 标签、训练/验证 COCO JSON 标注以及类别列表。
- 下载可使用 huggingface_hub.snapshot_download,并解压 archives 下的 *.tar 分片。
- 每条 COCO 图像记录中的 file_name 解析到对应训练或验证图像目录;同名的 TXT 标签位于匹配的标签目录。
范围与引用
- 本发布仅包含检测训练/验证数据和类别列表;不包含项目单独的 QA 数据、权重或代码。
- 源项目:https://github.com/hellozzk/MMIO。
- 使用该子集时引用 arXiv:2606.07953。
- 组件图像可能适用源特定权利;本数据集卡片不向第三方源材料授予一揽子许可。
使用声明
- 使用本数据集发表论文或用于商业用途前,必须联系 2673679261@qq.com 并签署相关开源声明。
- 未经许可使用者,将被追究盗用责任。




