TrafficSigns_YOLO-corrected

Traffic sign detection — leak-safe, standardized version of the "Traffic Signs Dataset for YOLOv8–YOLOv11 benchmarking". 交通标志检测 —— 「YOLOv8–YOLOv11 交通标志基准数据集」的防泄漏标准化版。

English · 中文

---

English

What is this?

A cleaned, leak-safe, standardized object-detection version of the Traffic Signs Dataset for YOLOv8–YOLOv11 benchmarking — six traffic-sign classes (left turn, straight ahead, right turn, dead end, no entry, stop) captured with an Intel RealSense D435i on an indoor FIRA 1:10-scale autonomous-vehicle track and annotated with CVAT.

Reference: Reveles-Martínez, R., Gamboa-Rosales, H., et al. "Benchmarking YOLOv8 to YOLOv11 Architectures for Real-Time Traffic Sign Recognition in Embedded 1:10 Scale Autonomous Vehicles." Technologies (MDPI), 2025. Upstream repository: rrevelesm/Articulos-Benchmarking-YOLOv8-to-YOLOv11.

On the surface the upstream release is clean (6,000 images, exactly one box each, six perfectly balanced classes). However, the 6,000 images are Roboflow-augmented copies derived from a much smaller set of base captures: 2,526 of them form near-identical clusters (identical 8×8-dHash and 32×32 pixel-MSE ≤ 1; the largest cluster has 25 images), and the upstream random split scattered these near-identical copies across train/valid/test — i.e. train/validation/test leakage.

This release groups each near-identical cluster entirely onto one side of a seeded, class-stratified split (replacing the upstream 4200 / 1200 / 600 with 4259 / 858 / 883), and ships ready-to-train YOLO TXT + Pascal VOC XML annotations with configuration files.

⚠️ Upstream license: none stated. The README says only "[LICENSE TO BE SPECIFIED - suggest CC BY 4.0]", and the Roboflow project metadata marks the licence as "Private". See License & attribution.

Corrections vs. the official release

#	Correction	Detail
1	Leak-safe re-split (main fix)	The 6,000 images are heavily augmented; 2,526 form near-identical clusters (identical 8×8-dHash and 32×32 pixel-MSE ≤ 1; largest = 25). Upstream scattered them across splits. Each cluster is now assigned entirely to one split, and a seeded, class-stratified split (4259 / 858 / 883) replaces the upstream 4200 / 1200 / 600.
2	Dual format + packaging	Upstream is YOLO TXT only. This release adds Pascal VOC XML from the same coordinates, plus classes.txt, data.yaml and per-split CSV tables.

No duplicates, out-of-range boxes, empty labels, malformed rows or EXIF anomalies were found — the only substantive issue is the split leakage.

Image pixels are never modified. All 6,000 images are byte-for-byte identical to upstream (640×640 RGB).

Scope — structural clean-up only. This is not a re-annotation.

Structural checks performed on the released set:

Check	Method	Result
Filename ↔ image ↔ label pairing	bidirectional	0 mismatches / orphans
Empty label files	file scan	0
Out-of-range / degenerate boxes	coordinate scan	0
Cross-split duplicate images	byte-identical md5 across splits	0
Near-identical leakage	cluster assignment	0 clusters straddle a split

Dataset at a glance

Property	Value
Images	6,000 (640 × 640 RGB .jpg)
Classes	6 — left_turn, straight_ahead, right_turn, dead_end, no_entry, stop
Bounding boxes	6,000 (exactly one per image)
Formats	Pascal VOC XML · YOLO TXT
Split	train 4,259 / val 858 / test 883 (seeded, class-stratified)

Boxes per class: left_turn 1,000 · straight_ahead 1,000 · right_turn 1,000 · dead_end 1,000 · no_entry 1,000 · stop 1,000

Structure

TrafficSigns_YOLO-corrected/
├── images/{train,val,test}/      # 640x640 RGB .jpg
├── labels/{train,val,test}/      # YOLO: cls cx cy w h (normalized)
├── Annotations/                  # Pascal VOC XML (flat, from the same coordinates)
├── JPEGImages/                   # copy of the images (VOC layout)
├── csv/                          # per-split label tables
├── classes.txt                   # 6 class names, one per line
├── data.yaml                     # Ultralytics YOLO configuration
├── LICENSE
└── README.md

Quick start (Ultralytics YOLO)

yolo detect train data=data.yaml model=yolov8n.pt epochs=100 imgsz=640

Citation

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

@article{reveles2025traffic,
  title   = {Benchmarking YOLOv8 to YOLOv11 Architectures for Real-Time Traffic Sign
             Recognition in Embedded 1:10 Scale Autonomous Vehicles},
  author  = {Reveles-Mart{\'i}nez, Rafael and Gamboa-Rosales, Hamurabi and others},
  journal = {Technologies},
  year    = {2025},
  publisher = {MDPI}
}

2. This corrected release — please cite it as well. This split differs from upstream (near-identical augmentations are no longer scattered across splits), and VOC annotations were added, so a citation to the original alone does not describe this version.

@misc{trafficsigns_yolo_corrected,
  author    = {KeenForgeAI},
  title     = {TrafficSigns_YOLO-corrected: a leak-safe release of the Traffic Signs Dataset for YOLOv8-YOLOv11 benchmarking},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/TrafficSigns_YOLO-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from the Traffic Signs Dataset
               (Reveles-Martinez et al., Technologies 2025). 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.}
}

License & attribution

The upstream dataset states no license. Its README says only *"[LICENSE TO BE SPECIFIED

• suggest CC BY 4.0]"* and the Roboflow project marks the licence as "Private"; the associated article is open access (MDPI). The copyright status of the original images and annotations is therefore undetermined — verify it before any commercial use. See LICENSE in this repository.

Our modifications (re-split, format conversion, documentation) are released under CC BY 4.0. Image pixels are unmodified and remain subject to the upstream terms.

Provenance

• Source: Traffic Signs Dataset (Roboflow workspace fira-gpeby, project signals-ocroa), via GitHub rrevelesm/Articulos-Benchmarking-YOLOv8-to-YOLOv11.
• Upstream: 6,000 images / 6,000 boxes / 6 classes ×1,000; split 4200 / 1200 / 600; 2,526 images in near-identical clusters.
• Verified: 6,000 images (6,000 boxes, byte-identical); leak-safe re-split 4259 / 858 / 883; 0 cross-split duplicates, 0 near-identical clusters straddling a split, 0 invalid boxes.

---

中文

这是什么？

面向 YOLOv8–YOLOv11 基准测试的交通标志数据集的防泄漏、标准化目标检测版——六类交通标志 （左转、直行、右转、尽头、禁止驶入、停车），使用 Intel RealSense D435i 在室内 FIRA 1:10 比例 自动驾驶赛道采集，CVAT 标注。

论文：Reveles-Martínez, R., Gamboa-Rosales, H. 等，《Benchmarking YOLOv8 to YOLOv11 Architectures for Real-Time Traffic Sign Recognition in Embedded 1:10 Scale Autonomous Vehicles》， Technologies（MDPI），2025。上游仓库： rrevelesm/Articulos-Benchmarking-YOLOv8-to-YOLOv11。

上游版本表面上很干净（6000 图、每图恰好 1 框、六类完全均衡）。但这 6000 张图是 Roboflow 从 少量底图增强而来：其中 2526 张构成近重复簇（8×8 dHash 完全相同且 32×32 像素 MSE ≤ 1； 最大簇 25 张），而上游的随机划分把这些近重复副本散布到了 train/valid/test 之间——即划分泄漏。

本版本把每个近重复簇整体分配到同一划分，并用固定种子、按类别分层的划分 （4259 / 858 / 883）替换上游的 4200 / 1200 / 600，同时提供开箱即用的 YOLO TXT + Pascal VOC XML 双格式标注及配置文件。

⚠️ 上游未声明许可证。 README 仅写 *"[LICENSE TO BE SPECIFIED - suggest CC BY 4.0]"*， Roboflow 项目元数据将许可标为 *"Private"*。详见 许可证与署名。

相对官方版的修正

#	修正	说明
1	防泄漏重划分（核心修正）	6000 张图为重度增强产物；其中 2526 张构成近重复簇（8×8 dHash 相同且 32×32 像素 MSE ≤ 1；最大 25 张）。上游将其散布于各划分。本版把每个簇整体分配到一个划分，并用固定种子、按类别分层的划分（4259 / 858 / 883）替换上游的 4200 / 1200 / 600。
2	双格式 + 规范化打包	上游仅有 YOLO TXT；本版由同一坐标生成 Pascal VOC XML，并补充 classes.txt、data.yaml 与按划分的 CSV 表格。

未发现重复、越界框、空标注、坏行或 EXIF 异常——唯一实质问题是划分泄漏。

图像像素从不修改——全部 6000 张图与上游文件逐字节一致（640×640 RGB）。

本版范围——仅结构化清理，不是重新标注。

已执行的结构化检查：0 配对错误 / 0 空标注 / 0 越界或退化框 / 0 跨划分重复 / 0 近重复簇跨越划分。

数据集概览

属性	值
图片	6,000（640 × 640 RGB .jpg）
类别	6 —— left_turn、straight_ahead、right_turn、dead_end、no_entry、stop
标注框	6,000（每图恰好 1 个）
格式	Pascal VOC XML · YOLO TXT
划分	train 4,259 / val 858 / test 883（固定种子、类别分层）

各类框数： left_turn 1,000 · straight_ahead 1,000 · right_turn 1,000 · dead_end 1,000 · no_entry 1,000 · stop 1,000

目录结构

TrafficSigns_YOLO-corrected/
├── images/{train,val,test}/      # 640x640 RGB .jpg
├── labels/{train,val,test}/      # YOLO：cls cx cy w h（归一化）
├── Annotations/                  # Pascal VOC XML（扁平，来自同一坐标）
├── JPEGImages/                   # 图片副本（VOC 布局）
├── csv/                          # 各划分标注表
├── classes.txt                   # 6 个类别名称，每行一个
├── data.yaml                     # Ultralytics YOLO 配置文件
├── LICENSE
└── README.md

快速开始 (Ultralytics YOLO)

yolo detect train data=data.yaml model=yolov8n.pt epochs=100 imgsz=640

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 reveles2025traffic； 2. 本修正版见上方英文 trafficsigns_yolo_corrected；3. 标注工具（可选）见 keenforge。

许可证与署名

上游数据集未声明许可证。 其 README 仅写 *"[LICENSE TO BE SPECIFIED - suggest CC BY 4.0]"*， Roboflow 项目将许可标为 *"Private"*；关联论文由 MDPI 开放获取。因此原始图像与标注的 版权状态未定——商用前请自行确认。详见本仓库 LICENSE。

我们对数据集的修改部分（重划分、格式转换、文档）以 CC BY 4.0 发布。 图像像素未修改，仍受上游条款约束。

来源

• 来源：Traffic Signs Dataset（Roboflow 工作区 fira-gpeby，项目 signals-ocroa），经 GitHub rrevelesm/Articulos-Benchmarking-YOLOv8-to-YOLOv11
• 上游：6,000 图 / 6,000 框 / 6 类各 1,000；划分 4200 / 1200 / 600；2526 张位于近重复簇
• 已校验：6,000 图（6,000 框，逐字节一致）；防泄漏重划分 4259 / 858 / 883；0 跨划分重复、0 近重复簇跨划分、0 非法框