KeenForgeAI/GC10-DET-corrected

GC10-DET-corrected


热轧钢板表面缺陷检测数据集 · 清理修正版 A cleaned and corrected release of the GC10-DET metallic surface defect dataset

English · 中文

English

Overview

GC10-DET-corrected is a cleaned copy of the public GC10-DET dataset (surface defects on hot-rolled steel strip, 10 classes, Pascal VOC annotations).

Image pixels and every bounding box coordinate are unchanged. Only the following were fixed:

1. 1.duplicate images removed
2. 2.images without annotations removed
3. 3.annotation class-name errors fixed

Dataset at a glance

Classes

The original dataset ships class names in Chinese pinyin (as above). Many third-party re-exports translate them to English (punching_hole, welding_line, crescent_gap, water_spot, oil_spot, silk_spot, inclusion, rolled_pit, crease, waist folding). This release keeps the original pinyin labels so it stays a drop-in replacement.

What was corrected

Starting point: the widely distributed archive of 2,306 images / 2,280 XML annotations (the paper claims 3,570 images, but that number has never matched the released data — see issue #2).

1. Duplicate images removed — 13

In the original data the same physical image appears more than once under different class folders (e.g. the identical file exists as both rolled_pit_img_02_425392000_00984 and inclusion_img_02_425392000_00984). This is a known defect of the per-class-folder layout and also causes train/test leakage if splits are drawn randomly.

Verified duplicates (pixel-identical, or identical after 90°/180°/270° rotation or mirroring). All 13 were un-annotated copies, so no annotation was lost by removing them:

Note: the duplicate pairs above are un-annotated on both sides, which is why all 13 fall out automatically in step 2 below. The corrected set therefore contains zero duplicates (verified by an exhaustive pixel-level comparison under all 8 axis-aligned orientations).

2. Images without annotations removed — 26

Result: 2,306 − 26 = 2,280 images, and 2,280 annotations — every image has exactly one XML, and every XML has exactly one image (0 orphans in either direction).

3. Class-name errors fixed — 132 annotation files

The original annotations contain 12 distinct class names instead of 10 (issue #6):

• —`10_yaozhed` → `10_yaozhe`: 131 boxes were labelled with a trailing d. Merging them with the 11 correctly-spelled boxes gives the class its true count of 142. This typo is also the cause of the common IndexError: index 10 is out of bounds when training YOLO on the raw data.
• —`d` → `1_chongkong`: the box (1997, 257)–(2048, 310) was inspected visually — it contains a real dark oval defect on the base steel, spatially separate from the 2_hanfeng weld band that spans the top of the same image. It is therefore a genuine defect instance and was re-assigned to the punching-hole class rather than dropped. (Some community forks simply delete this box; this release keeps it, so the box count stays at the original 3,542.)

After the fix the dataset contains exactly the 10 official classes and 3,542 boxes.

Directory structure

GC10-DET-corrected/
├── JPEGImages/     2,280 × .jpg   (2048×1000 grayscale)
├── Annotations/    2,280 × .xml   (Pascal VOC, class names fixed)
├── ImageSets/
│   ├── train.txt   1,599 stems
│   └── test.txt      681 stems
├── classes.txt     10 class names, one per line
├── LICENSE
└── README.md

ImageSets/ is copied verbatim from the upstream archive. Its train.txt + test.txt (1,599 + 681 = 2,280) map exactly onto the corrected image set, so the original split is preserved with no adjustment needed.

Provenance

Known upstream issues (for reference)

1. 1.Paper-vs-release size mismatch — paper says 3,570 images; released data has ~2,294–2,312 files (issue #2).
2. 2.Extra class names 10_yaozhed and d (issue #6).
3. 3.Same image duplicated across class folders → train/test leakage risk.
4. 4.Un-annotated images present.
5. 5.No author-documented train/test split (the ImageSets/ lists are distributed but never described in the paper).
6. 6.Annotation quality — a peer-reviewed study (Rattanaphan & Briassouli, Processes 12(3):456, 2024, doi:10.3390/pr12030456) reports "misclassified or incorrect labels, missing labels, and inconsistent labeling" and reduced the set to 3,243 instances after manual correction.

Scope of this release: issues 2, 3 and 4 are fixed here. Issues 1, 5 and 6 are not addressed — in particular this release does not re-verify the correctness of individual boxes. It is a structural clean-up, not a re-annotation.

License

Upstream license is unstated by the original authors. The GC10-DET GitHub repository ships no LICENSE file, and the question has been asked in issue #4 without a reply. "CC BY 4.0" is widely repeated by downstream mirrors (Kaggle, Roboflow, Dataset Ninja), but it is not an author statement. The paper is CC BY 4.0 — that covers the article, not necessarily the dataset.

Accordingly:

• —The modifications in this release (de-duplication, removal of un-annotated images, class-name corrections, this README) are released under CC BY 4.0.
• —The underlying GC10-DET images and annotations remain under whatever terms the original authors intend. Verify the upstream license before any commercial or redistributive use.

See `LICENSE`.

How to cite

1. The original dataset — please always cite this.

GC10-DET is the primary source: all credit for the images and for the original annotations belongs to its authors.

@Article{lv2020deep,
  author    = {Lv, Xiaoming and Duan, Fajie and Jiang, Jia-jia and Fu, Xiao and Gan, Lin},
  title     = {Deep Metallic Surface Defect Detection: The New Benchmark and Detection Network},
  journal   = {Sensors},
  volume    = {20},
  number    = {6},
  pages     = {1562},
  year      = {2020},
  doi       = {10.3390/s20061562},
  publisher = {MDPI},
  issn      = {1424-8220}
}

2. This corrected release — please cite it as well.

This release is not identical to the upstream dataset. Compared with the archive it was derived from, 13 duplicate images and 26 un-annotated images were removed (2,306 → 2,280 images) and 132 annotation files had their class names corrected. A citation to the original alone therefore does not let a reader reproduce results computed on this version.

@misc{gc10det_corrected,
  author    = {KeenForgeAI},
  title     = {GC10-DET-corrected: a cleaned release of the GC10-DET steel surface defect dataset},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  doi       = {10.57967/hf/10526},
url       = {https://huggingface.co/datasets/KeenForgeAI/GC10-DET-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from Lv et al. (2020),
               doi:10.3390/s20061562. Upstream licence unstated.}
}

3. The annotation tool (optional).

@software{keenforge,
  author    = {KeenForgeAI},
  title     = {KeenForge: a local-first, offline image annotation and model-training desktop tool},
  year      = {2026},
  publisher = {KeenForgeAI},
  url       = {https://github.com/KeenForgeAI/KeenForge},
  note      = {MIT licensed. Developed by Lu Gan and Sam Li.}
}

中文

简介

GC10-DET-corrected 是公开数据集 GC10-DET（热轧带钢表面缺陷，10 类，Pascal VOC 标注） 的清理修正版。

图片像素与所有标注框坐标均未改动，只修正了以下三点：

1. 1.删除重复图片
2. 2.删除没有标注的图片
3. 3.修正标注中的类别名错误

数据概览

类别与数量

原始数据集的类别名就是中文拼音（如上）。很多第三方镜像把它翻译成英文 （punching_hole / welding_line / …）。本版本保留原始拼音标签，以便直接替换使用。

具体修正了什么

起点是广为流传的那份 2,306 张图片 / 2,280 个 XML 标注（论文声称 3,570 张， 但这个数字从未与实际发布的数据对上，见 issue #2）。

1. 删除重复图片 —— 13 张

原始数据里同一张物理图片会在不同类别目录下重复出现（例如完全相同的文件同时存在为 rolled_pit_img_02_425392000_00984 和 inclusion_img_02_425392000_00984）。 这是「按类别分目录」这一组织方式的已知缺陷，若随机划分训练/测试集还会造成数据泄漏。

经逐像素验证（含 90°/180°/270° 旋转与镜像），共 13 张重复。这 13 张全部是无标注的副本， 因此删除它们不损失任何标注。完整配对表见上方英文部分。

修正后的数据集经全部 8 种轴向变换的逐像素比对验证，重复数为 0。

2. 删除没有标注的图片 —— 26 张

完整清单见上方英文部分。结果：2,306 − 26 = 2,280 张图片、2,280 个标注， 每个图片恰好对应一个 XML，每个 XML 恰好对应一个图片（双向零孤儿）。

3. 修正类别名错误 —— 132 个标注文件

原始标注里存在 12 个类别名，而官方只有 10 类（见 issue #6）：

• —`10_yaozhed` → `10_yaozhe`：131 个框多写了一个 d，与 11 个正确拼写的框合并后， 该类别真实数量为 142。这个错拼也是直接用原始数据训练 YOLO 时报 IndexError: index 10 is out of bounds 的原因。
• —`d` → `1_chongkong`：该框位于 (1997, 257)–(2048, 310)，经目视检查确认框内存在真实缺陷 —— 一个位于基础钢板上的深色椭圆斑块，与横贯图像顶部的 2_hanfeng 焊缝带在空间上分离。 因此它是真实缺陷实例，改判到冲孔类而不是丢弃。（部分社区分支选择直接删除此框； 本版本保留，故总框数维持原始的 3,542。）

修正后数据集包含正好 10 个官方类别、3,542 个框。

目录结构

GC10-DET-corrected/
├── JPEGImages/     2,280 个 .jpg   (2048×1000 灰度)
├── Annotations/    2,280 个 .xml   (Pascal VOC，类别名已修正)
├── ImageSets/
│   ├── train.txt   1,599 行
│   └── test.txt      681 行
├── classes.txt     10 个类别名，每行一个
├── LICENSE
└── README.md

ImageSets/ 从上游原样复制。train.txt + test.txt（1,599 + 681 = 2,280）与修正后的 图片集完全一一对应，因此原始划分得以原样保留，无需任何调整。

数据来源

上游已知问题（供参考）

1. 1.论文与实际发布的数据量不一致 —— 论文称 3,570 张，实际发布约 2,294~2,312 个文件（issue #2）。
2. 2.多出两个类别名 10_yaozhed 与 d（issue #6）。
3. 3.同一张图在不同类别目录下重复 —— 存在训练/测试集泄漏风险。
4. 4.存在没有标注的图片。
5. 5.作者没有文档化的训练/测试划分（ImageSets/ 随包发布，但论文中从未说明）。
6. 6.标注质量 —— 一篇同行评议研究（Rattanaphan & Briassouli, Processes 12(3):456, 2024, doi:10.3390/pr12030456）指出原始数据存在 「标签错误、标签缺失、标注不一致」，经人工修正后仅保留 3,243 个实例。

本版本的范围：修复了上述第 2、3、4 项。第 1、5、6 项未处理 —— 特别说明， 本版本没有重新核对每个标注框的正确性，它是一次结构性清理，而非重新标注。

许可证

上游许可证从未由原作者声明。 GC10-DET 的 GitHub 仓库没有 LICENSE 文件， 有人专门开了 issue #4 询问许可证，至今无人回复。「CC BY 4.0」是被 Kaggle / Roboflow / Dataset Ninja 等下游镜像 广泛沿用的说法，但并非作者声明。论文本身是 CC BY 4.0 —— 那覆盖的是文章，不一定是数据集。

因此：

• —本版本的修改部分（去重、删除无标注图片、类别名修正、本 README）以 CC BY 4.0 发布。
• —底层的 GC10-DET 图片与标注仍归原作者所声明的条款（目前未声明）。 在用于商业用途或再分发之前，请自行确认上游许可证。

详见 `LICENSE`。

如何引用

1. 原始数据集（请务必引用） —— GC10-DET 是数据的根本来源，图片与原始标注的 一切功劳归其作者。BibTeX 见上方英文部分（lv2020deep）。

2. 本修正版（请一并引用） —— 本版本与原数据集并不相同：相比上游归档，删除了 13 张重复图片与 26 张无标注图片（2,306 → 2,280 张），并修正了 132 个标注文件的类别名。 因此只引用原始论文无法让他人复现基于本版本得到的结果。BibTeX 见上方（gc10det_corrected）。

3. 标注工具（可选） —— 见上方 keenforge。