Custom YOLO Backend for Label Studio

This backend provides ML-assisted labeling capabilities to accelerate your annotation workflow, supporting both object detection and instance segmentation tasks.

🏗️ Project Structure

• Dockerfile: The Dockerfile for building the backend container.

• docker-compose.yml: The docker-compose file for running the backend.

• _wsgi.py: WSGI app initializer.

• start.sh: bash script to start the whole process.

• model.py: The Python code for the ML backend model.

• requirements.txt: The list of Python dependencies for the backend.

🚀 Quick Start

1. Clone the repository:

   git clone https://github.com/seblful/label-studio-yolo-backend.git
   cd label-studio-yolo-backend

2. Create and prepare your model directory:

   mkdir model
   cp /path/to/your/model.pt model/
   mv model/your_model.pt model/best.pt

3. Edit .env with your settings:

   - LABEL_STUDIO_URL=http://host.docker.internal:${PORT}
   - LABEL_STUDIO_API_KEY=<your_api_key>
   - TASK_TYPE=<detection> or <segmentation>

   Important Notes:

   • Default LABEL_STUDIO_URL is http://host.docker.internal:8080
   • Get IP using 'ifconfig' (Linux/Mac) or 'ipconfig' (Windows)
   • ⚠️ Never use localhost as the container is isolated from the host
   • Get your API_KEY from: Label Studio -> Account Settings -> Access Token

4. Deploy using the following commands:

   docker pull pytorch/pytorch:2.3.0-cuda12.1-cudnn8-runtime
   docker compose build
   docker compose up

5. Add model in project settings:

   From the project settings, select the Model page and click Connect Model.

   Add the URL http://locallhost:{PORT} and save the model as an ML backend.

   

6. Label in interactive mode

   To use this functionality, activate Auto-Annotation.

🇨🇳 For Users in China

To ensure smooth deployment in regions with internet restrictions:

1. Configure Docker daemon with proxy:

{
  "registry-mirrors": ["https://registry.docker-cn.com"]
}

📋 TODO

• add support for keypoints and obb.

Training

Model training is not included in this project. This will probably be added later.

Contributing

Contributions to this project are welcome. To contribute, please submit an issue or pull request.