The club is only as findable as the room.
A golf shaft is a dark quarter-inch stick moving past 100 mph — the hardest thing in the bay to track. What decides how well we track it isn't really the algorithm; it's the room it's swung in. A bright bay with a plain wall gives up the whole swing arc; a dark sim room hides half of it. And where the room can't help, a few grams of retroreflective tape on the shaft — lit by a small ring at the lens — lift club tracking to instrument grade.
Club and ball tracking is part of the full-swing pipeline that's still being built. This page explains the method we're developing, not a finished feature — every part of it is being measured against a hand-annotated corpus of real swings before it goes near a release. See how we earn the numbers →
The room does half the work
Before any code runs, the room has already decided how easy the club is to see. It's your choice — nobody should repaint a wall or bin a sim room to use a camera — but two things matter here, and they aren't equal.
- 1
Light is the hard limit
Missing photons can't be recovered in software. A dark room forces long exposures (more blur) and high gain (more noise), and the shaft either vanishes into dark surrounds or washes out against a bright mat. No detector fixes that after the fact.
- 2
A quiet background is second
Static clutter — window frames, screens, club racks — invites false locks, but it's manageable: the tracker can veto anything that was already sitting in the empty scene. Manageable isn't free, but unlike darkness it isn't a wall.
- ✓
An honest ladder
Best to worst: bright light and a plain wall (full performance), bright but busy, dark but plain, and — the typical home simulator — dark and busy. That last one is the hard case, and it's exactly where tape earns its place.
The point of the tape. The retroreflective option below is deliberately environment-independent — it makes the hardest room behave like the ideal one. That's why it's the right answer for a black sim room, and largely unnecessary in a bright bay with a quiet backdrop.
Illustrative — light first, background second; tape lifts tier 4 to tier 1.
A few grams of tape, instrument grade
Vision-only tracking is the default, and no one has to touch their clubs. But a hard room takes its toll on any photometric detector long before it troubles a marker — so for a dark bay, a taped shaft changes the problem entirely.
- ✓
Bands become bright blobs
Retroreflective tape lit by a small light at the lens throws that light straight back, saturating as bright, compact dots whatever the background — dark wall or blown-out mat alike. A banded shaft is trackable exactly where a plain one is invisible.
- ✓
Known spacing is geometry
Bands at measured spacings turn every frame into direct geometry: projected spacing against known spacing gives the shaft's foreshortening outright, the line of bands gives its angle without leaning on the hands, and the clubhead follows as a known extrapolation even when the head can't be seen.
- ✓
A pattern, not a lucky flash
Along-shaft distance ratios survive projection, so the pattern reads the same at any club angle. A specular glint is one blob; several bands in a line at the recorded ratios are unmistakably ours — and because blur smears across the shaft while the gaps run along it, the pattern survives the impact streak.
- !
One physical rule
Retro tape returns light to its source within a ~1–2° cone, so the light must sit at the lens — a small LED ring, not the ceiling. Visible-band only; never IR, which fights launch monitors.
Optional by design. The untaped tracker stays the product — and every taped swing doubles as dense ground truth for proving the unmarked detector, so both improve together. Taping is an opt-in high-accuracy mode for hard rooms, not a requirement. How to mark up a club →
Illustrative — a “2–1–3” band pattern, lit from the lens and read as geometry.
The fastest move in sport
Most rooms aren't the ideal bay and most clubs aren't taped — so the tracker still has to find a plain shaft in whatever light the room gives it. That's the product default, and it's a hard problem: a full swing lasts under two seconds, the downswing a fraction of that, and at those speeds an ordinary camera fights two things at once.
- !
Severe motion blur
Past 100 mph a thin shaft isn't a crisp line — it smears into a faint, semi-transparent wedge of grey across the frame.
- !
Background clutter
Alignment sticks, trouser creases, shadows and seams in the turf all look line-like — and tempt a naive tracker into "finding" a club that isn't there.
Illustrative — the club becomes a faint fan; the ground lies in wait.
How the computer looks
Rather than search the whole frame at once, the tracker works outward from what it already knows for certain.
- 1
Anchor on the hands
It first finds the golfer's hands from the body-pose skeleton. That becomes the pivot everything else is measured from.
- 2
Search outward in rays
It casts search lines out from the hands in a circle, only as far as a club could possibly reach.
- 3
Look for a ridge
Along each ray it looks for a long, high-contrast run of dark pixels — the signature of a real shaft, not a stray mark.
Illustrative — rays sweep outward from the hands.
Where the naive way fails
Blind to the golfer's body, that basic search falls into two classic traps.
- ✗
Panicking when cropped
To guess scale it assumed a fixed height. Crop the feet out of frame and that guess collapses — the search widens to the whole image and gets lost in the noise.
- ✗
Locking onto tiny lines
A club is longer than an arm, yet the basic search would happily lock onto a two-inch sleeve crease or watch strap beside the hands — bright, sharp and completely wrong.
Illustrative — a crop and a crease, both leading the search astray.
Scale and gate by the arm
The fix is to tie the search to the skeleton itself. The golfer's own arm sets both how far to look and how short is too short.
- ✓
Sizing from the arm, not a guess
We measure the arm on screen, shoulder to wrist. A club is always around 1.8–2.2× that length, so the search radius is set precisely — however the video is framed or cropped.
- ✓
A physical-length gate
A club is longer than an arm, full stop. Anything shorter than one arm length is rejected outright, so sleeve folds and watch straps never reach the matching stage.
Illustrative — the arm sets both the reach and the floor.
A swing is a double pendulum
Mechanically, the lead arm and club behave like two linked pendulums. That gives a physics-based guardrail for where the club can plausibly be.
- ✓
A predictable wrist-cock angle
The angle between lead arm and club follows a known curve through the swing — near straight at address, hinged near the top, holding lag, releasing back to straight at impact.
- ✓
A moving angle guardrail
Knowing the swing phase and the arm's angle, the model predicts where the club should be, allows a soft margin around it, and rejects anything that grossly breaks the physics.
Illustrative — the physics narrows the search to a sector.
Don't fight the blur — predict it
At the bottom of the downswing the club is a ghost. So we stop hunting for a crisp line and change the rules for the high-speed zone.
- ✓
Expect a wedge, not a line
From the club's speed and the camera's exposure time, the model works out how wide the blurred fan should be — then looks for that fan instead of a sharp edge.
- ✓
Add up the faint energy
Any single pixel in the smear is too faint to trust, but summing the light across the whole predicted sweep adds up to a confident detection.
Illustrative — energy summed across the predicted sweep.
Two tracks that check each other
Because the tracker produces both a measured path and a predicted one, it can grade its own work — and show that working on replay.
- 1
Measured vs predicted
One path comes from the pixels, the other from the physics model. The gap between them is a running health check on the tracker.
- 2
A "ghost" overlay for testing
During checks, a faint dashed prediction can be drawn behind the solid detected club, making any drift easy to spot by eye.
- 3
A clubhead trail on replay
Replays draw a fading, glowing trail of the last few clubhead positions, so swing path and release read at a glance.
Illustrative — measured and predicted, side by side, with the trail.
Tuned to the footage in front of it
A few touches let the tracker adapt to real, messy video rather than assuming ideal conditions.
- 1
Exposure, measured not assumed
A quick first pass measures how wide the blur smear actually is on this footage, so the model uses the real exposure rather than a fixed default.
- 2
Eased for unusual swings
It gauges how clean the footage is and relaxes or tightens its assumptions for non-standard stances, instead of forcing one setting on everyone.
- 3
Handedness, worked out for you
Left- or right-handed is read from the direction the club first travels in the backswing — nothing to set.
- 4
Club lifted off the background
A still background is estimated and subtracted, and the moving body is masked out, so the faint club smear stands clear of everything else.
Still being proven. These refinements are built and improving, not signed off. Next on the bench: projecting the physics through the camera's true 3D geometry, calibrating against motion-sensor ground truth, and extending the model to down-the-line views — each held to the same validation bar before it ships.
Illustrative — isolating the club from a moving scene.
The room, the tape, the physics.
Good club tracking starts with the room, leans on a few grams of tape where the room can't help, and — for every plain shaft in between — falls back on what a body and a swing can physically do. All of it is drawn from the literature, kept open to inspect, and held against real swings before it ships. It's a prototype under validation, with a first release targeted for late 2026.