autoannotate-vision 0.1.0

SOTA unsupervised auto-annotation SDK for image classification

AutoAnnotate-Vision 🎯

State-of-the-art unsupervised auto-annotation SDK for image classification with GUI

AutoAnnotate-Vision automatically clusters and organizes unlabeled image datasets using cutting-edge vision models (CLIP, DINOv2, SigLIP2). Features a graphical user interface for easy use and HTML preview for visual cluster inspection.

✨ Features

• 🎨 Graphical User Interface: Easy folder browsers and visual controls
• 🖼️ HTML Image Preview: View cluster samples in browse before labeling
• 🤖 SOTA Vision Models: CLIP, DINOv2, DINOv2-Large, SigLIP2
• 🔬 Multiple Clustering: K-means, Spectral, DBSCAN
• 📁 Smart Organization: Preserves original filenames
• ✂️ Auto Splits: Train/val/test dataset splitting
• 💾 Export: CSV, JSON formats
• 🔌 Python API: Full programmatic control

🚀 Installation

pip install autoannotate-vision

Or from source:

git clone https://github.com/Metamind-Innovations/autoannotate-vision.git
cd autoannotate-vision
pip install -e .

🎨 Quick Start - GUI

The easiest and most simplified way to use AutoAnnotate-Vision:

autoannotate-images

Note: Windows users need to have the latest C++ Redistributable installed which can be found here

Workflow:

1. 📁 Select input folder with images
2. 📂 Select output folder
3. 🔢 Set number of classes
4. 🤖 Choose model (SigLIP2 or DINOv2 recommended)
5. ▶️ Click "Start Auto-Annotation"

The app will cluster images and open HTML previews in your browser showing sample images from each cluster for easy labeling!

💻 CLI Usage

For extra commands and utilities.

autoannotate-images-cli annotate /path/to/images /path/to/output \
    --n-clusters 10 \
    --method kmeans \
    --model siglip2 \
    --create-splits

Available models: clip, dinov2, dinov2-large, siglip2

🐍 Python API

from autoannotate import AutoAnnotator

annotator = AutoAnnotator(
    input_dir="./images",
    output_dir="./output",
    model="siglip2",  # or "dinov2", "dinov2-large", "clip"
    clustering_method="kmeans",
    n_clusters=5,
    batch_size=32
)

result = annotator.run_full_pipeline(create_splits=True)
print(f"Processed {result['n_images']} images into {result['n_clusters']} classes")

Available models: clip, dinov2, dinov2-large, siglip2 Available clustering methods: kmeans, hdbscan, spectral, dbscan

📁 Output Structure

output/
├── metadata.json
├── labels.csv
├── cats/              # Your class names
│   ├── IMG_001.jpg   # Original filenames preserved!
│   └── ...
├── dogs/
└── splits/            # train/val/test
    ├── train/
    ├── val/
    └── test/

🧠 Model Comparison

Model	Speed	Quality	Notes
CLIP	⚡⚡	⭐⭐⭐	General-purpose, good for diverse datasets
DINOv2	⚡⚡⚡	⭐⭐⭐⭐	Fast, self-supervised, excellent for objects
DINOv2-Large	⚡	⭐⭐⭐⭐⭐	Best quality, slower, great for fine details
SigLIP2	⚡⚡⚡	⭐⭐⭐⭐⭐	Latest Google model - Recommended 🌟

Recommendation: Start with SigLIP2 for best results, or DINOv2 for faster processing.

🔧 Features & Improvements

• ✅ Fast Image Processing: All models use optimized processors (use_fast=True) for better performance
• ✅ Normalized Embeddings: All embeddings are L2-normalized for consistent similarity measurements
• ✅ Batch Processing: Efficient batch processing with configurable batch sizes
• ✅ GPU Support: Automatic GPU detection and usage when available
• ✅ Progress Tracking: Real-time progress bars for all operations
• ✅ HTML Previews: Interactive HTML preview for visual cluster inspection before labeling

🔍 Pre-Push Checklist

Before pushing code:

# Format code
black src/autoannotate tests

# Run tests
pytest tests/ -v

# Typing
mypy src/autoannotate --ignore-missing-imports

🤝 Contributing

1. Fork the repository
2. Create feature branch
3. Format with Black: black src/autoannotate tests
4. Check typing with mypy: mypy src/autoannotate --ignore-missing-imports
5. Run tests: pytest tests/ -v
6. Push and create PR

📄 License

MIT License - see LICENSE file.

🙏 Acknowledgments

Built with PyTorch, Transformers, scikit-learn. Vision models: CLIP, DINOv2, SigLIP2.

Made for the RAIDO Project, from MetaMind Innovations