Datasets:

KeenForgeAI
/

KolektorSDD2-corrected

⚠️ NON-COMMERCIAL — SHARE-ALIKE — 仅限非商业 · 相同方式共享

The upstream KolektorSDD2 is licensed CC BY-NC-SA 4.0. This release inherits that licence: no commercial use, credit the authors, share derivatives under the same licence. For commercial use contact the original author (Danijel Skočaj, Vicos).

Surface-defect detection/segmentation — cleaned, split and standardized version of KolektorSDD2. 表面缺陷检测/分割 —— KolektorSDD2 的清理、划分与规整版。

English · 中文

English

What is this?

A cleaned, split and standardized version of KolektorSDD2 (Kolektor Surface-Defect Dataset 2) — surface defects (scratches, minor spots, surface imperfections) on Kolektor commutators, captured as ~230×630 px images. Upstream: Vicos, University of Ljubljana — https://www.vicos.si/resources/kolektorsdd2/.

Reference: Božic, J., Tabernik, D. & Skočaj, D. "Mixed supervision for surface-defect detection: from weakly to fully supervised learning." Computers in Industry, 2021.

This release ships the dataset in three aligned views:

image classification — defect / normal (in labels.csv),
pixel segmentation — the ground-truth masks (masks/<split>/),
object detection — a single-class defect bounding box derived from each mask (Pascal VOC Annotations/ + YOLO labels/<split>/).

⚠️ Note: the boxes are derived from the masks (axis-aligned extent of the defect region), provided for convenience / detection pipelines. The masks remain the authoritative annotation. No pixel, mask or box coordinate was manually re-annotated.

Corrections vs. the official release

#	Correction	Detail
1	Removed 1 duplicate sample	`10301 (copy).png` + `10301_GT (copy).png` were byte-identical copies of `10301.png` / `10301_GT.png`; removed.
2	Benchmark-preserving split	Kept the official train/test split and carved a validation set (15% of train, stratified by defect/normal) from train: train 1,981 / val 350 / test 1,004. The official test set is untouched.
3	Three aligned views	Added a classification table (`labels.csv`), kept the pixel masks (`masks/`), and added derived single-class detection boxes (VOC + YOLO).
4	Dropped unreproducible splits	The upstream `split_weakly_*.pyb` pickle files are not reproducible and were excluded.
5	Packaging	`classes.txt` (`defect`), `data.yaml`, per-split CSVs; normalised filenames.

Image pixels and masks are never modified.

Structural checks performed on the released set:

Check	Method	Result
Image ↔ mask pairing	bidirectional	0 mismatches / orphans
Mask size vs. image	pixel dimensions	0 mismatches
Degenerate boxes derived from masks	mask extent	0
Out-of-range boxes	coordinate scan	0
Cross-split duplicate images	byte-identical md5	0
Exact / near duplicates	md5 + (dHash ≤ 2 & MSE ≤ 1)	0 after removing the `(copy)` pair

Dataset at a glance

Property	Value
Images	3,335 (~230×630 px, RGB `.png`)
Labels	`defect` 356 · `normal` 2,979
Masks	pixel-precise binary ground truth (per image)
Derived boxes	356 single-class `defect` boxes
Split	train 1,981 (defect 209 / normal 1,772) · val 350 (37 / 313) · test 1,004 (110 / 894)

Class imbalance: only ~10.7% of images contain a defect (356 / 3,335); defect pixels are ~0.3% of all pixels on average. Use stratified sampling / class weighting.

Structure

KolektorSDD2-corrected/
├── images/{train,val,test}/      # RGB .png (original)
├── masks/{train,val,test}/       # binary ground-truth masks (same filenames)
├── labels/{train,val,test}/      # YOLO (1 class = defect; empty file for normal)
├── Annotations/                  # Pascal VOC XML (0 objects for normal images)
├── JPEGImages/                   # copy of the images (VOC layout)
├── csv/{train,val,test}_labels.csv   # defect boxes only
├── labels.csv                    # ALL images: filename,split,label,defect,mask,bbox,W,H
├── classes.txt                   # "defect"
├── data.yaml                     # Ultralytics detection config (nc=1)
├── LICENSE                       # CC BY-NC-SA 4.0 + notice of changes
└── README.md

Quick start

# detection (derived boxes)
yolo detect train data=data.yaml model=yolov8n.pt epochs=100 imgsz=640

# classification / segmentation from the tables
import csv, pandas as pd
df = pd.read_csv("labels.csv")          # filename, split, label, defect, mask, bbox_xyxy, ...

Citation

1. The original dataset / paper — please always cite.

@article{bozic2021mixed,
  title   = {Mixed supervision for surface-defect detection: from weakly to fully
             supervised learning},
  author  = {Bo{\v{z}}i{\v{c}}, Jakob and Tabernik, Domen and Sko{\v{c}}aj, Danijel},
  journal = {Computers in Industry},
  year    = {2021}
}

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

@misc{kolektorsdd2_corrected,
  author    = {KeenForgeAI},
  title     = {KolektorSDD2-corrected: a cleaned, split and standardized release of KolektorSDD2},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/KolektorSDD2-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from KolektorSDD2 (Bo\v{z}i\v{c}
               et al., Computers in Industry 2021). CC BY-NC-SA 4.0.}
}

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.}
}

License & attribution

CC BY-NC-SA 4.0 (Attribution-NonCommercial-ShareAlike 4.0 International). The upstream KolektorSDD2 is licensed under CC BY-NC-SA 4.0; this corrected release is distributed under the same licence. Non-commercial use only; credit the authors, indicate changes, and license derivatives under CC BY-NC-SA 4.0. See LICENSE.

Provenance

Source: KolektorSDD2, Vicos / University of Ljubljana — https://www.vicos.si/resources/kolektorsdd2/
Upstream: 3,337 images (train 246 pos / 2,085 neg, test 110 pos / 894 neg) + masks; no val; split_weakly_*.pyb.
Verified: 3,335 images (356 defect); 1 duplicate pair removed; train 1,981 / val 350 / test 1,004; 0 invalid boxes, 0 mask mismatch, 0 cross-split leakage.

中文

这是什么？

KolektorSDD2（Kolektor 表面缺陷数据集 2）的清理、划分与标准化版——Kolektor 换向器上的划痕、细小斑点、表面瑕疵等缺陷，图像约 230×630 px。上游：卢布尔雅那大学 Vicos 实验室 —— https://www.vicos.si/resources/kolektorsdd2/。

论文：Božic, J., Tabernik, D. & Skočaj, D.《Mixed supervision for surface-defect detection: from weakly to fully supervised learning》，Computers in Industry，2021。

本版以三种对齐视图提供：

图像分类 —— defect / normal（见 labels.csv）；
像素分割 —— 真值掩码（masks/<split>/）；
目标检测 —— 由每个掩码派生的单类 defect 边界框（Pascal VOC Annotations/ + YOLO labels/<split>/）。

⚠️ 说明： 边界框是由掩码派生（缺陷区域的轴对齐外接框），仅为方便检测流程； 掩码仍是权威标注。未对任何像素、掩码或框坐标进行人工重标。

相对官方版的修正

#	修正	说明
1	删除 1 个重复样本	`10301 (copy).png` + `10301_GT (copy).png` 与 `10301.png` / `10301_GT.png` 逐字节相同，已删除。
2	保基准的划分	保留官方 train/test 划分，并从 train 中按缺陷/正常分层切出 15% 验证集：train 1,981 / val 350 / test 1,004。官方 test 完全不动。
3	三种对齐视图	新增分类表（`labels.csv`）、保留像素掩码（`masks/`）、新增派生的单类检测框（VOC + YOLO）。
4	剔除不可复现划分	上游 `split_weakly_*.pyb` pickle 不可复现，已排除。
5	规范化打包	`classes.txt`（`defect`）、`data.yaml`、按划分 CSV；文件名规范化。

图像像素与掩码从不修改。

已执行的结构化检查：图↔掩码配对 0 错配 / 掩码尺寸 0 不符 / 派生框 0 退化 / 越界框 0 / 跨划分重复 0 / 去重后完全与近重复 0。

数据集概览

属性	值
图片	3,335（约 230×630 px，RGB `.png`）
标签	`defect` 356 · `normal` 2,979
掩码	逐像素二值真值（每图一张）
派生框	356 个单类 `defect` 框
划分	train 1,981（缺陷 209 / 正常 1,772）· val 350（37 / 313）· test 1,004（110 / 894）

类别不平衡： 仅约 10.7% 的图为缺陷（356 / 3,335）；缺陷像素平均仅占全部像素约 0.3%。建议使用分层采样 / 类别加权。

目录结构

KolektorSDD2-corrected/
├── images/{train,val,test}/      # RGB .png（原始）
├── masks/{train,val,test}/       # 二值真值掩码（同名）
├── labels/{train,val,test}/      # YOLO（1 类 = defect；正常图为空文件）
├── Annotations/                  # Pascal VOC XML（正常图 0 个 object）
├── JPEGImages/                   # 图片副本（VOC 布局）
├── csv/{train,val,test}_labels.csv   # 仅缺陷框
├── labels.csv                    # 全部图片：filename,split,label,defect,mask,bbox,W,H
├── classes.txt                   # "defect"
├── data.yaml                     # Ultralytics 检测配置（nc=1）
├── LICENSE                       # CC BY-NC-SA 4.0 + 修改说明
└── README.md

快速开始

# 检测（派生框）
yolo detect train data=data.yaml model=yolov8n.pt epochs=100 imgsz=640

# 分类 / 分割：读取标签表
import pandas as pd
df = pd.read_csv("labels.csv")      # filename, split, label, defect, mask, bbox_xyxy, ...

引用

请同时引用原始论文与本修正版：1. 原始论文见上方英文 bozic2021mixed； 2. 本修正版见上方英文 kolektorsdd2_corrected；3. 标注工具（可选）见 keenforge。

许可证与署名

CC BY-NC-SA 4.0（署名—非商业—相同方式共享 4.0 国际）。 上游 KolektorSDD2 采用 CC BY-NC-SA 4.0，本修正版以同一许可发布。仅限非商业用途；须署名原作者、注明修改，并以 CC BY-NC-SA 4.0 发布衍生作品。详见 LICENSE。

来源

来源：KolektorSDD2，Vicos / 卢布尔雅那大学 —— https://www.vicos.si/resources/kolektorsdd2/
上游：3,337 图（train 246 正 / 2085 负，test 110 正 / 894 负）+ 掩码；无 val；split_weakly_*.pyb
已校验：3,335 图（缺陷 356）；删除 1 个重复样本；train 1,981 / val 350 / test 1,004；0 非法框、0 掩码不符、0 跨划分泄漏

Downloads last month: 179