yuturuishi/ytlabel

ytlabel – Open‑source image/video annotation tool with AI‑assisted labeling

What it is

• A lightweight web‑based application for manually labeling images (rectangles, polygons, etc.) and videos, then exporting the results in YOLO‑compatible formats.
• It can also call a user‑provided large‑model API (e.g., LMStudio, vLLM, Ollama, Alibaba Cloud) to generate AI‑auto‑labels.
• The tool focuses on data‑annotation and dataset export; model training is handled by a separate project (ytstudio).

Key capabilities

• Manual labeling – rectangle, polygon, multi‑class tags; import from LabelMe datasets.
• AI auto‑labeling – plug‑in for any locally hosted LLM/VLM API you configure; no external data is sent.
• Dataset export – YOLO format with configurable train/val/test split.
• File manager – browse, upload, download, and delete images and annotation files.
• Offline‑first – all static assets are bundled; the server can run on an isolated intranet.
• Keyboard shortcuts – Ctrl+S to save, Ctrl+Shift+D to clear a label.

How to get started

# clone the repo, then create a virtual environment
python -m venv venv
source venv/bin/activate   # or venv\Scripts\activate on Windows

# install dependencies (mirrored from Tsinghua PyPI source)
pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple

# run the Flask server
python app.py --host 0.0.0.0 --port 9924

Open a modern browser at http://127.0.0.1:9924 and you’ll see the labeling UI.

Project layout (high‑level)

• app.py – entry point, starts the Flask + SocketIO server.
• AiUtils.py – helper class that forwards images to the configured large‑model API and parses the returned annotations.
• templates/ – HTML pages (index.html, ai_config.html, file_manager.html).
• static/ – CSS, JavaScript, icons, and a bundled Socket.IO client library for offline use.
• uploads/ – runtime storage for uploaded media, generated annotation files, and user‑specific config.

Technology stack

• Backend: Python 3.9+, Flask, Flask‑SocketIO.
• Frontend: HTML / CSS / JavaScript, Socket.IO for real‑time updates.
• Image handling: OpenCV and Pillow (PIL).
• Packaging: PyInstaller spec (ytlabel.spec) for optional standalone binaries.

License

• The code written for ytlabel is released under the MIT license.
• Third‑party libraries keep their original licenses.
• The AI auto‑label feature merely forwards requests to a user‑controlled model API; ytlabel itself does not ship any model weights.

Who might use it

• Researchers or engineers building custom computer‑vision datasets who need a simple, self‑hosted labeling UI.
• Teams that want to keep data on‑premise and still benefit from large‑model assistance without sending images to cloud services.
• Anyone looking for a Flask‑based, cross‑platform labeling tool that can be packaged for offline deployment.