Camthink Neoeyes Ne301 End-to-end Edge-ai Toolkit

Overview

After receiving the CamThink NeoEyes NE301, I realized that while it supports model detection, a very important step was missing — image annotation after data collection.
I spent three days building an AI-based tool to solve this problem. With just 20 images, you can complete model training and perform image annotation automatically.
I also plan to further quantize and add more models to expand the annotation capabilities. The project has now been open-sourced on GitHub, and I hope it can be helpful to the community

AI Tool Stack is an end-to-end edge-AI toolkit for NeoEyes NE301 (and similar) devices, covering data collection, labeling, training, quantization, and deployment. It can be self-hosted and continuously updated.

Key Features

• Data ingest & management: MQTT image ingest; project/image lists.
• Annotation workbench: shortcut-driven labeling, class management, dataset import/export (COCO / YOLO / project ZIP), annotation history.
• Training & testing: YOLO-based training and model testing.
• Quantization & deployment: one-click NE301 model package export.
• Export: YOLO training dataset, full annotation ZIP.

Requirements

Docker & docker-compose (required)：(Optionally) If you need to generate the NE301 quantization model package, please pull the image in advance:

docker pull camthink/ne301-dev:latest

Quick Start (Docker)

docker-compose build

docker-compose up

The frontend is built in-container (react-scripts) and the backend is FastAPI. Check container logs for the exposed frontend URL (usually http://localhost:8000 or mapped port).

Local Development (optional)

Run frontend and backend separately:

cd frontend && npm install && npm start

cd backend  && pip install -r requirements.txt && uvicorn main:app --reload

API routes live in backend/api/routes.py. Frontend config is in frontend/src/config.ts.

Typical Workflow

Data Import / Export

• Import: COCO (.json), YOLO (.zip), project export ZIP (images/ + annotations/ + classes.json)
• Export: YOLO dataset, full annotation ZIP

Keyboard Shortcuts (common)

• Navigate images: A / ←, D / →
• Tools: R rectangle, P polygon, V select
• Class select: 1–9 for first 9 classes
• Toggle annotations: H
• Undo/Redo: Ctrl/Cmd + Z, Ctrl/Cmd + Shift + Z
• Save: Ctrl/Cmd + S

Roadmap

• Model hub: standalone quantization and downloadable model library
• Device management: NE101/NE301 connection, data debugging, OTA

Ports & Environment

• Default ports (docker-compose):
  • Backend / API: 8000
  • Frontend (react-scripts in container): 3000 (proxied/mapped, check logs)
• Key environment variables (examples):
  • API_BASE_URL: frontend API base (e.g., http://localhost:8000)
  • MQTT_BROKER_HOST: MQTT broker host/IP (e.g., localhost or broker service name in compose)
  • MQTT_BROKER, MQTT_TOPIC: NE301 MQTT settings
  • DATASETS_ROOT: backend dataset storage path
  • Add more as needed in .env or compose env overrides.

Dataset Formats (import/export)

• COCO: single JSON with images/annotations/categories.
• YOLO: ZIP with images/ + labels/ (.txt).
• Project ZIP: images/ + annotations/*.json + optional classes.json (id/name/color).
• Exports: YOLO dataset ZIP; full annotation ZIP (images + annotations + classes.json).

Troubleshooting

• Build fails (frontend) missing assets: ensure all referenced CSS/TSX exist (DatasetImportModal.css etc.).
• Port conflicts: adjust mapped ports in docker-compose.yml.
• Slow annotation layer on fast navigation: caching added; still slow → reduce image size (JPEG/WebP) or check network/CPU.

Testing

• If tests exist: run frontend lint/tests via npm test or backend tests via pytest (add when available).
• If absent: mark as TODO in issues; manual smoke: load project, navigate images, import/export datasets, train/quantize happy-path.

Contributing

• Issues & PRs welcome.
• Style: follow repo defaults (eslint/prettier for frontend, black/flake8 if configured for backend).
• Branching: fork & PR or feature branches; please keep changes scoped and include repro steps for bug fixes.

Security

• Do not commit secrets/tokens.
• For security reports, contact maintainer privately (add email/contact here).

License

• Add your license of choice (e.g., MIT/Apache-2.0). Add LICENSE file and reference it here.