anno-mage 2.0.3

Semi-automatic image annotation tool with AI-assisted object detection

Anno-Mage: A Semi Automatic Image Annotation Tool

Semi-automatic image annotation toolbox powered by PyTorch object detection models. Available as both a desktop app (Tkinter) and a web app (FastAPI + React).

Interfaces

Desktop App

A Tkinter GUI that runs locally. Annotate images from a directory with keyboard navigation and a precision zoom panel.

Web App

A browser-based interface with a modern dark/light UI, REST API, and dataset directory browsing. See web/README.md for full documentation.

---

Desktop App

Dependencies

• Python 3.12
• PyTorch + Torchvision (for RetinaNet inference)
• Pillow, pascal-voc-writer

Custom snapshots from legacy Keras (.h5) or TensorFlow (frozen_inference_graph.pb) models can be placed in snapshots/keras/ and snapshots/tensorflow/ respectively and will be listed in the model menu.

Installation

1. Clone this repository.

2. Create and activate a virtual environment:

   python -m venv my_venv
   source my_venv/bin/activate  # On Windows: my_venv\Scripts\activate
   

3. Install dependencies:

   pip install torch torchvision pillow pascal-voc-writer
   

4. Run the app:

   python main.py
   

The default model (RetinaNet ResNet50 FPN V2) downloads its weights automatically from the PyTorch model hub on first run.

Usage

1. Click Open Dir to select a folder of images, or Open for a single image.
2. Use Add All Classes to load all 80 COCO labels, or pick individual ones from the dropdown.
3. Select Auto suggestion mode and click Detect to run object detection.
4. To annotate manually, select a class from the list, then drag a bounding box on the canvas.
5. Use ← → arrow keys to move between images. Annotations save automatically on navigation.
6. Final annotations are written to:
   • annotations/annotations.csv
   • annotations/annotations_voc/<image>.xml (Pascal VOC)

To annotate a custom (non-COCO) dataset, update the label map in config.py before running.

Tested on

• Windows 10
• Ubuntu 16.04
• macOS High Sierra

---

Web App

See web/README.md for installation, usage, and API reference.

Quick start:

# Backend (port 8000)
cd web/backend && python main.py

# Frontend (port 3000)
cd web/frontend && npm install && npm run dev

Or use the convenience script:

cd web && bash start.sh

---

PyPI Distribution

Install from PyPI

pip install anno-mage
anno-mage

The app opens in your browser automatically. Annotations are saved to ~/.anno-mage/annotations/.

Publish a Release

Releases publish automatically to PyPI when a version tag is pushed. GitHub Actions builds the frontend, packages everything, and publishes via PyPI Trusted Publishers (no tokens required).

One-time PyPI setup:

1. Go to your PyPI project → Manage → Publishing → Add a new publisher
2. Set: GitHub repo virajmavani/semi-auto-image-annotation-tool, workflow release.yml, environment pypi

To release:

git tag v2.0.1
git push origin v2.0.1

That's it — the workflow in .github/workflows/release.yml handles the rest.

Build Locally

To build the package without publishing:

Prerequisites:

pip install build
npm install  # inside web/frontend if not already done

bash build_release.sh

This compiles the React frontend, copies the build into anno_mage/static/, and produces wheel and sdist artifacts in dist/.

---

Output Formats

Both interfaces produce identical output:

Format	Location	Description
CSV	annotations/annotations.csv	image_path,x1,y1,x2,y2,label per row
Pascal VOC XML	annotations/annotations_voc/	One XML file per image

---

Acknowledgments

• Meditab Software Inc.
• PyTorch / Torchvision for the RetinaNet implementation
• Computer Vision Group, L.D. College of Engineering

Join the developers channel

Slack: https://join.slack.com/t/annomage/shared_invite/zt-dh4ca9du-4VOcwUMCSNA6lmyG~tNUPg