MoonLabel

An object-detection and image-caption dataset tool.

Powered by Moondream VLM

---

Overview

MoonLabel is both a Python library and a tiny web UI to generate object-detection and image-caption datasets quickly.

1. Use the library to auto-label folders of images and export YOLO, COCO, VOC, or Captions.
2. Or launch the UI, switch between Detection/Caption, and export with one click.

Backends supported: Moondream Cloud, Moondream Station, or fully local (Hugging Face).

Demo

demo.mp4

---

Features

• 📦 Library + UI — moonlabel package with an optional web UI.
• 🌐 FastAPI server — Served by a single moonlabel-ui command.
• ⚛️ Modern frontend — React, TypeScript, TailwindCSS, Vite.
• 🖼️ Object detection — Choose between Moondream Cloud, the open-source Hugging Face model, or the native Moondream Station app.
• 📝 Image caption datasets — Export captions as captions.jsonl alongside images.
• ⚡ GPU-accelerated & offline — Local and Station modes automatically use available hardware acceleration (CUDA / MPS).

Install

• Library only (Cloud/Station by default):

pip install moonlabel

• Library + UI server:

pip install "moonlabel[ui]"

• Local inference (Hugging Face) extras:

pip install "moonlabel[local]"

• Both UI and local inference:

pip install "moonlabel[ui,local]"

Quick Start (UI)

pip install "moonlabel[ui]"
moonlabel-ui    # opens http://localhost:8342

Choose backend in Settings:

• Moondream Cloud: paste API key
• Moondream Station: set endpoint (default http://localhost:2020/v1)
• Local (Hugging Face): install local extras and select Local

In the Home page:

• Use the top toggle to select Detection or Caption dataset.
• For Caption, choose length (short/normal/long) and generate/export.

Quick Start (Library)

from moonlabel import create_dataset

# Cloud
create_dataset("/path/to/images", objects=["person"], api_key="YOUR_API_KEY")

# Station
create_dataset("/path/to/images", objects=["car"], station_endpoint="http://localhost:2020/v1")

# Local (after: pip install "moonlabel[local]")
create_dataset("/path/to/images", objects=["bottle"])  # no key needed

By default this exports YOLO. Choose formats via export_format:

# YOLO (default)
create_dataset("/path/to/images", objects=["person"], export_format="yolo")

# COCO
create_dataset("/path/to/images", objects=["person", "car"], export_format="coco")

# Pascal VOC
create_dataset("/path/to/images", objects=["cat", "dog"], export_format="voc")

# Image Captioning
create_dataset("/path/to/images", export_format="caption")
# With length (short|normal|long)
create_dataset("/path/to/images", export_format="caption", caption_length="normal")

Output layouts:

• YOLO: images/, labels/, classes.txt
• COCO: images/, annotations/instances.json, classes.txt
• VOC: images/, annotations/*.xml, classes.txt
• Caption: images/, annotations/captions.jsonl

Moondream Station Mode

The backend can connect to a running Moondream Station instance for fast, native, on-device inference.

1. Download, install, and run Moondream Station.
2. Ensure the endpoint matches your Station configuration (default: http://localhost:2020/v1).

Local Mode (Hugging Face)

The backend can run fully offline using the open-source vikhyatk/moondream2 checkpoint.

1. pip install "moonlabel[local]"
2. In the UI, select Local (no API key required).

The first detection will trigger a one-off model download to ~/.cache/huggingface/; subsequent runs reuse the cached weights.

GPU / Device selection

The backend chooses the best device automatically in the following order: CUDA → Apple Silicon (MPS) → CPU.

Override via environment variable before launching the backend:

# Force GPU
export MOONDREAM_DEVICE=cuda

# Force Apple Silicon
export MOONDREAM_DEVICE=mps

# CPU only
export MOONDREAM_DEVICE=cpu

Project Structure

moonlabel/
├── src/moonlabel/             # Python package (library + server)
│   ├── dataset.py             # create_dataset API
│   ├── infer.py               # Moondream wrapper (cloud/station/local)
│   ├── types.py               # shared types 
│   ├── utils.py               # helpers 
│   ├── yolo.py                # YOLO label writer
│   ├── coco.py                # COCO writer
│   ├── voc.py                 # Pascal VOC XML writer
│   └── server/                # FastAPI app + static assets
│       ├── api.py
│       ├── cli.py             # moonlabel-ui entrypoint (port 8342)
│       └── static/            # embedded UI build (no npm for users)
├── ui/                        # Frontend source (for maintainers)
│   └── dist/                  # Built files to embed
├── scripts/embed_ui.py        # Copies ui/dist → src/moonlabel/server/static
├── Makefile                   # make ui-build, ui-embed, release
└── pyproject.toml

---

Roadmap / TODOs

Below are planned enhancements and upcoming features. Contributions welcome!

• Local Hugging Face model support – Offline inference with optional GPU acceleration.
• Moondream Station integration – Native Mac/Linux app support for on-device inference.
• Batch uploads – Label multiple images in one go, with progress tracking.
• Additional export formats – COCO JSON and Pascal VOC alongside YOLO.

---

License

This project is licensed under the terms of the Apache License 2.0. See LICENSE for details.