Image Annotation QC

该 Skill 通过本地 Python 脚本读取图像与 COCO/YOLO/VOC/LabelMe 标注，检测漏标、错标、偏移、重复等问题，计算质量评分，并生成 TXT、JSON、Excel 报告及错误可视化图片，支持场景与采样配置。

mingo-318

热度 83.53

最近更新 2026-05-11 17:28:46

分类：数据依赖：代码执行 / 读写文件应用场景：数据处理 / 报告撰写 / 自动化脚本 / 工作流

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add image-annotation-qc

Metadata

name: image-annotation-qc
description: Image Annotation Quality Control Tool - Automatically detect quality issues in bounding box and polygon segmentation annotations, generate visual reports. Supports COCO/YOLO/VOC/LabelMe formats, suitable for autonomous driving, industrial inspection, security monitoring and other AI training data quality control scenarios.
metadata: {"openclaw":{"emoji":"🔍","requires":{"bins":["python3"],"pip":["Pillow","openpyxl"]}}}

Image Annotation QC Tool

Automatically detect quality issues in image annotations and generate detailed reports with visualization.

Features

✅ Multi-format Support: COCO / YOLO / VOC / LabelMe
✅ Auto-detection: Automatically identify format without specification
✅ Scene-specific QC: general / road / industrial / security
✅ Auto-save Reports: Save to qc_report/ directory automatically
✅ Visualization: Generate images with error annotations (red boxes)
✅ Multi-format Reports: TXT / JSON / Excel

Quick Start

# Basic usage (auto-detect format)
python3 scripts/qc_tool.py -i <image_dir> -a <annotation_dir>

# Specify format
python3 scripts/qc_tool.py -i ./images -a ./annotations -f labelme

# Industrial inspection scenario
python3 scripts/qc_tool.py -i ./images -a ./annotations -d industrial

# Sample 100 images
python3 scripts/qc_tool.py -i ./images -a ./annotations -s 100

# Specify output directory
python3 scripts/qc_tool.py -i ./images -a ./annotations -o ./my_report

# Generate Excel report
python3 scripts/qc_tool.py -i ./images -a ./annotations --formats txt json xlsx

Parameters

Parameter	Short	Description	Default
`--image-dir`	`-i`	Image directory	(required)
`--annotation`	`-a`	Annotation file or directory	(required)
`--format`	`-f`	Annotation format	`auto`
`--domain`	`-d`	Application scenario	`general`
`--sample`	`-s`	Sample size	all
`--output`	`-o`	Output directory	`./qc_report`
`--formats`	-	Report formats	`txt json`
`--no-visual`	-	Disable visualization	false

Scenarios

general: General purpose
road: Autonomous driving (small objects, occlusions)
industrial: Industrial inspection (micro-defects)
security: Security monitoring

Output Example

╔══════════════════════════════════════════╗
║           Image Annotation QC              ║
╠══════════════════════════════════════════╣
║  Images: 1000   Annotations: 4523         ║
║  Errors: 127    Quality Score: 85.5       ║
║  Accuracy: 97.2%                          ║
╚══════════════════════════════════════════╝

Output Files

After QC completes, the following files are generated in the annotation directory:

annotations/
├── qc_report/
│   ├── qc_report.txt      # Text report
│   ├── qc_report.json     # JSON detailed report
│   ├── qc_report.xlsx     # Excel report (optional)
│   └── visual/            # Visualization images
│       ├── frame_001_qc.png
│       └── frame_002_qc.png

Error Types

Error	Description	Weight
Missing	Obvious targets not labeled	1.0
Wrong	Annotation region doesn't match target	0.8
Offset	Box edge deviates >10px	0.5
Too Large	Box significantly larger than target	0.3
Too Small	Box doesn't fully contain target	0.5
Duplicate	Same object annotated multiple times	0.5
Label Error	Wrong class label	0.8

Quality Score

Score = 100 × (annotations - errors×weight) / annotations

Grading:
- 90+: Excellent
- 80-89: Good
- 70-79: Pass
- <70: Fail

Installation

pip install Pillow openpyxl

As Module

from qc_tool import AnnotationQC

qc = AnnotationQC('images/', 'annotations/', format='labelme')
data = qc.load_annotations()
result = qc.check_annotations(data)
qc.generate_report(result, formats=['txt', 'json'])
qc.visualize_errors(result)
print(f"Quality Score: {result.quality_score}")