Throw

Darts scoring from a camera, with vision models running in the browser.

throw.maz-ui.com
Screenshot of the Throw project
Status
Live
Year
2026
Role
Design, model training and full development
Stack
  • Nuxt
  • TypeScript
  • ONNX Runtime Web
  • WebGPU
  • YOLO
  • Python
  • PWA
  • WebSocket

Private source code: the architecture and the results are shown here.

The context

Automatic darts scoring systems need several cameras mounted around the board and a dedicated computer. I wanted a phone set in front of the board, with no server and no extra hardware.

What I built

A Nuxt PWA that detects the board and the dart tips in real time with ONNX Runtime Web (WebGPU, WASM fallback). I trained and benchmarked seven generations of my own YOLO pose models on audited datasets and hand-labelled throws, then kept the best pipeline: the difference between two frames first, the model next.

Architecture

  1. 01CaptureThe phone camera films the board, frames stabilized.
  2. 02CalibrationThe board corners are detected to straighten the view.
  3. 03DifferenceBefore and after the throw are compared to isolate the dart.
  4. 04TipThe pose model locates the tip to the millimetre.
  5. 05ScoreThe point is mapped on the board and the game is updated.

Results & impact

  • 71 to 96 ms per inference with WebGPU on iPhone, 175 ms with WASM, fully on the device.
  • 88% of darts right against Autodarts with the difference pipeline, versus 57% from the photo alone.
  • Median error of the board corners down from 7.1 to 2.7 mm across trainings, about 400 tests.

Skills applied

  • Computer vision
  • Model training
  • Full stack product

Got a project in mind?

An application to build, a team to strengthen or an AI product to launch: let's talk.