raccoon-corrected

Raccoon Detection Dataset — Corrected & Verified

A fully human-reviewed, re-annotated version of the classic Raccoon object detection dataset. Every one of the 193 images was manually reviewed; loose, missing, and incorrect bounding boxes were corrected.

中文摘要见文末 中文版修正报告。

Dataset at a glance

What was corrected (vs. the original annotations)

Every image was compared with the original VOC annotations using greedy IoU matching (matched ≥ 0.5, unchanged ≥ 0.95):

Key observations

• The original boxes were often loose (e.g., raccoon-1: box (81,88)-(522,408) → tightened to (80,105)-(529,405), IoU 0.91). The average IoU of matched boxes is 0.810, i.e. the corrected boxes bound the raccoon noticeably more tightly.
• Occluded / partially visible raccoons were kept with tight boxes; several missed instances were added.
• 7 exact/near duplicates were identified and removed.

Files

.
├── images/            # 193 images (jpg), unchanged from the original dataset
├── labels/            # 193 YOLO labels (one .txt per image)
├── splits/
│   ├── train.txt      # 153 images (same split as the original dataset)
│   └── test.txt       # 40 images
├── data.yaml          # Ultralytics dataset config
└── LICENSE

Label format

Standard YOLO: one line per object, normalized to [0, 1]:

class_id  x_center  y_center  width  height

class_id = 0 → raccoon

Quick start (Ultralytics)

from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.train(data="data.yaml", epochs=100, imgsz=640)

metrics = model.val()   # evaluate on splits/test.txt

Source & license

• Original dataset: experiencor/raccoon_dataset (fork of datitran/raccoon_dataset) — MIT License.
• This corrected version: images unchanged; annotations fully re-drawn and verified. Released under the same MIT License, with attribution to the original authors.
• If you use this dataset, please also credit the original Raccoon dataset.

How this version was made

• Tool: KeenForge — an open-source, local-first auto-labeling and training desktop tool (YOLO training loop with human-in-the-loop review).
• Process: every image opened and reviewed manually; each raccoon re-boxed tightly; missed animals added; non-raccoon objects and duplicates removed.
• Comparison statistics computed by greedy IoU matching between the original VOC boxes and the corrected YOLO boxes.

中文版修正报告

这是一个经过完整人工复核、重新标注的经典 Raccoon 检测数据集修正版。

修正内容（与原始标注对比，贪心 IoU 匹配：≥0.5 视为同一目标，≥0.95 视为未修改）：

• 框收紧/调整：182 个（平均 IoU 0.810 —— 原始框普遍偏松，修正后紧贴目标）
• 未修改：22 个
• 删除错误框：6 个
• 补充漏标：7 个（原标注遗漏的浣熊）
• 移除重复图片：7 张（raccoon-45/50/74/83/85/98/116）

标注工具：KeenForge（本地化标注 + 训练闭环工具）

许可：原始数据集为 MIT 许可；本修正版沿用 MIT，图片未改动，标注全部人工复核重画，请同时注明原始数据集来源。