Darts are found by an on-device keypoint model, and scored through a homography onto the board

Context

The app reads a darts score from handheld photos of the board: no fixed camera, no photo of the empty board before the darts are thrown, often an oblique angle, and darts that hide each other. A player’s photos never leave the phone (CLAUDE.md), so there is no server to run anything on.

The published work that comes closest is DeepDarts (McNally et al., CVPR 2021 Workshops, arXiv 2105.09880). One detector finds four calibration points on the double wire and the point where each dart enters the board. A homography maps those points onto the regulation board, and the score is read in polar coordinates. It reports 94.7% of images scored exactly right on face-on phone photos, but only 84.0% at varied angles. That 84.0% was taken on DSLR photos and chosen as the best of five runs on the test set, so it is optimistic. The main failure is a dart hidden by another dart. No later peer-reviewed work on handheld scoring was found as of 2026-11-11.

The licences constrain what can be reused. This repo is all-rights-reserved and source-available, and the app is on Play with a tip purchase:

Decision

• One keypoint detector does both jobs. In each photo it finds two kinds of point:
  • Calibration points: the boundary points between segments on the double and treble wires. That is many points rather than DeepDarts’ four, so a homography still fits when darts cover some of them.
  • Entry points: where each dart’s shaft enters the board.

• The board is the regulation steel-tip wire board of the WDF Playing and Tournament Rules, rev. 20, §4.01, checked against that PDF on 2026-11-11. The rules fix the bull and 25 rings by their inside diameters (12.7 mm and 31.8 mm), the treble and double wires by their outside edge (107.4 mm and 170.0 mm from the centre), and both narrow rings as 8.0 mm wide measured inside wire to inside wire. A wire may be 1.27–1.85 mm thick, written w. Each wire is a band, in mm from the centre:

  Wire	Inner edge	Outer edge
  Bull ring	6.35	6.35 + w
  25 ring	15.9	15.9 + w
  Inner treble wire	99.4 − 2w	99.4 − w
  Outer treble wire	107.4 − w	107.4
  Inner double wire	162.0 − 2w	162.0 − w
  Outer double wire	170.0 − w	170.0

  A region runs between the wires around it, and a dart scores on the side of the wire its point enters (§3.03). The app takes w as 1.5 mm. A real board’s wire differs from that by at most 0.35 mm, so an edge measured across one wire moves by up to 0.35 mm and the inner treble and double edges, measured across two, by up to 0.7 mm: both inside the 2 mm wire margin below. Older BDO tables give the treble as 107.0 mm; WDF rev. 20 gives 107.4 mm, and that is used. Segments run clockwise from the top starting at 20: 20, 1, 18, 4, 13, 6, 10, 15, 2, 17, 3, 19, 7, 16, 8, 11, 14, 9, 12, 5.

• The homography is fitted with RANSAC. Each entry point is mapped into board millimetres and scored there. Only the entry point is mapped: the homography is correct only on the board’s plane, and a dart’s flight sticks out of it.
• Several photos are merged in board millimetres, one dart to one dart. Each further photo’s Entry points are matched to the darts found so far by an assignment on distance, so two Entry points from the same photo never land on one dart. A pair further apart than a gate of a few millimetres is not a match: those stay separate darts. A matched dart’s position is the average of its Entry points, weighted by confidence. The app asks for another angle only when one of these is true:
  • fewer darts were found than were thrown;
  • a dart’s confidence is low;
  • the photos disagree on a dart’s segment;
  • an entry point sits within about 2 mm of a wire.
• The model is trained off the phone, with permissively licensed code only (amended 2026-11-13, below), in this repo’s ml/ folder. ml/ holds code only: the datasets, the labels and the training weights are git-ignored and never published, because the owner’s photos show their home. The one thing that leaves ml/ is the exported model file that ships inside the app, so scoring needs no network, and Play services does not supply the runtime. Which framework and which runtime is settled in phase M2 by measurement, not here.
• Training data:
  • the DeepDarts dataset, credited on the open-source licences screen;
  • the owner’s own handheld photos from several angles, labelled. No public handheld set exists, and these are what the accuracy bar is measured on.

  A Roboflow dataset is added only after its page has been opened and both its licence and where its images came from are confirmed.

Alternatives

• Classical computer vision only: red and green ring masks, an ellipse fit, and the segment boundaries as calibration points (Zgonc & Batagelj, ROSUS 2023). It finds the board well. It finds a dart’s Entry point only by comparing against a photo taken before the dart from the same fixed camera, which a handheld phone does not have. Kept as a cross-check on the calibration, not as the way darts are found.
• Fixed-camera systems (Autodarts, Scolia): these use three cameras mounted around the board and compare frames from before and after each dart. The app has neither the mounted cameras nor the before frame. What carries over is the idea that more views settle a hidden dart, which is what the second angle is for.
• A cloud vision model (any hosted multimodal API): the photo would leave the phone, the app would need INTERNET and someone to pay per request, and nothing published shows such models locate a dart to the millimetre a wire needs.
• Ultralytics YOLO, the easiest route to train: AGPL-3.0, and its terms make the trained weights AGPL too. Using it means releasing the app under AGPL or buying a licence.

Consequences

• Most of the work is data. Hundreds of labelled handheld photos from several angles have to be taken and labelled before any accuracy figure means anything. Phase M1 is that work and phase M2 trains on it; both come before the app shows anything real.
• A dart on a wire needs 1–2 mm of accuracy. Lens distortion, a calibration point hidden by a dart, or the wrong spot on a dart taken for its Entry point each cost that much. So the app will sometimes be unsure, says so, and lets the player correct a dart by tapping the board.
• Steel-tip regulation boards only. A soft-tip board has different geometry and needs its own ADR.
• The app ships the model, which adds roughly 5–20 MB.
• CC BY 4.0 requires attribution. The DeepDarts dataset’s credit belongs on the licences screen from the first build that ships a model trained on it.

Amendment — 2026-11-13: the Calibration points, DeepDarts’ labels and the licence rule

Phase M1 (phase-M1.md) writes the geometry above as code (ml/src/darts_ml/board.py) and settles three things this ADR left open.

• A Calibration point is the centre of a crossing: where the centre line of the radial wire between two segments meets the centre line of one of the four ring wires. That is 20 × 4 = 80 points. Each radial wire lies on a segment boundary, at 9° plus a multiple of 18° clockwise from the top, and each ring wire’s centre radius follows from the bands above with w = 1.5 mm:

  Ring wire	Name	Centre radius (mm)
  Outer double	do	169.25
  Inner double	di	159.75
  Outer treble	to	106.65
  Inner treble	ti	97.15

  A point is named <ring>-<a>-<b>, with a and b the segments either side in clockwise order, such as do-20-1 and ti-5-20. The centre of a crossing is what the eye finds on a Photo; a wire’s edge is not.

• DeepDarts’ labels come from IEEE DataPort’s labels_pkl.zip, which is CC BY 4.0, and not from the dataset/labels.pkl in wmcnally/deep-darts, which carries that repository’s no-licence status. DataPort’s models.zip is not used: those weights were trained with the unlicensed code, and M2 starts from its own permissively licensed initialisation.
• The licence rule is permissive licences only (Apache, MIT, BSD and the like) for the code we adapt, train with, or ship. Copyleft code (GPL, AGPL) is none of those. The earlier wording, “Apache or MIT code only”, contradicted this ADR’s own table, which counts BSD-3 ncnn as usable.
  • A runtime library bundled inside a dev tool’s wheel under a runtime exception or the LGPL is allowed, and listed in ml/README.md’s licence table. Today that is libgfortran (GPL-3.0 with the GCC Runtime Library Exception) and libquadmath (LGPL-2.1-or-later), both in NumPy’s Linux wheel. They are allowed because ml/ never ships: the app ships only the exported weights.
  • labelme, which is GPL-3.0, runs beside ml/ through uvx to label Photos and never enters the lockfile, training or the app.