How to Film Your Golf Swing So AI Can Actually Read It

You paid for the app. You stood on the range at Moor Park in the drizzle, propped the phone against your bag, hit six seven-irons, and uploaded the best one. The analysis came back saying your hip rotation at impact was 71 degrees and your shaft lean was negative. Negative. You know it wasn’t negative, because the ball went 165 yards and drew slightly.

The AI wasn’t wrong about what it saw. It was wrong about what it could see, which is a different problem entirely, and it’s yours to fix.

Here’s the thing nobody selling swing analysis wants to lead with: pose-estimation models are only as good as the pixels you hand them. Every consumer golf AI in 2026, whether it’s Sportsbox 3DGolf, Swing AI, HackMotion’s video companion, or the vision layer bolted onto your Garmin R10 app, runs some variant of a keypoint detector. It finds your wrists, elbows, hips, shoulders, ankles. Then it infers everything else. Give it a clip where your trail hip is behind a bag for eight frames and it doesn’t tell you it lost the hip. It guesses, smooths the guess into the curve, and reports a number to one decimal place.

Four capture faults account for nearly every clip that gets rejected outright or, worse, silently mis-tracked. None of them need new kit.

Fault one: your frame rate is lying to you

This is the big one and almost everyone gets it wrong, because phones default to something sensible for filming your kids and catastrophic for filming a golf swing.

Run the numbers. A tour-level downswing takes roughly 0.25 seconds from transition to impact. A club-golfer downswing is a touch slower, call it 0.28. At 30fps, that’s about 8 frames for the entire downswing. Eight. At 60fps you get 17. At 240fps you get 67.

Now the part that actually matters: clubhead speed for a 12-handicapper with a driver is around 92mph, which is 41 metres per second. At 30fps, the clubhead travels 1.37 metres between frames. Your ball is 4.3cm across. The club is never photographed anywhere near the ball. Impact, as a visual event, does not exist in your video.

Downswing frames captured (0.28s), by frame rate:

 30 fps   ████                                     8 frames
 60 fps   ████████                                17 frames
120 fps   ████████████████                        34 frames
240 fps   ████████████████████████████████        67 frames

Clubhead travel between frames at 92mph:
 30 fps   1.37 m
 60 fps   0.68 m
120 fps   0.34 m
240 fps   0.17 m

Sportsbox 3DGolf asks for 60fps minimum and openly recommends 120. Most 2D analysis apps will accept 30 and then produce transition and impact numbers built from two or three real frames plus interpolation. The tempo ratio it hands back (you’re looking for something near 3:1 backswing to downswing) is arithmetically incapable of being accurate when the denominator is eight frames.

Go into your camera settings now. iPhone: Settings, Camera, Record Video, pick 1080p at 120fps rather than 4K at 30. The resolution drop costs you nothing here and the frame rate gain is everything. Samsung Galaxy S24 and up: 1080p/60 in the main video mode, or use Pro Video for 120. Pixel: Settings, Video, 1080p/60, then Slow Motion for 120 or 240 if your model supports it.

Do not use your phone’s built-in Slo-Mo mode and then upload the file. It’s the single most common mistake in the whole process. Slo-Mo exports a 30fps container with the slow section baked in, so the AI reads a 30fps video of a man swinging in treacle and reports your tempo as 6.2:1. Film in standard video at a high frame rate instead, and let the app do the slowing.

Fault two: the camera is standing where a person would stand, not where a camera should

Pose-estimation models were trained on humans photographed from roughly chest height. Golf capture guidance exists for a reason and the reason is geometric, not aesthetic.

Two positions, and only two, produce readable golf footage.

Down-the-line. Camera on the target line extended behind the ball, lens height at hand height (roughly belt buckle to mid-thigh, around 90cm for most adults), distance 2.5 to 3 metres. The critical part: the lens must sit on the line through the ball and the target, not a foot inside it or outside it. A 30cm error at 2.5m distance skews your apparent shaft plane by around 7 degrees, which is more than the difference the app will flag as a fault.

Face-on. Camera perpendicular to the target line, aimed at the centre of your stance, same lens height, 2.5 to 3 metres away. This is the view for weight shift, sway, hip slide, and wrist angles.

What kills both: tripod height. Setting the phone on your golf bag puts the lens at maybe 75cm and tilted up about 15 degrees, and that tilt is the thing. Vertical camera tilt introduces keystone distortion that models read as a real change in your spine angle. A 15-degree upward tilt can add 5 to 8 degrees of apparent spine tilt away from target at address, and the AI will dutifully tell you that you’re set up too far behind the ball.

A cheap tripod with a phone clamp costs £15 and fixes this permanently. The SelfieCom or similar 1.6m extendable models are fine. Get one that lets you set an actual height rather than one that stands at whatever height it stands at.

The pillar piece Filming a Swing an AI Can Actually Read walks through the full setup geometry with distance tables for different club lengths, which matters more than you’d think: your driver setup needs about 40cm more distance than your wedge setup to keep the club in frame at the top.

Fault three: you’re wearing golf clothes

This one sounds absurd and it’s the fault I’d bet money on in your specific case.

Keypoint detectors find your hips by finding the boundary between your torso and your legs. A navy quarter-zip over navy trousers gives the model a single continuous dark blob from shoulder to ankle. There is no edge. It places your hip keypoint by inference from your shoulder and knee positions, and the confidence score drops, and most apps don’t show you the confidence score.

I’ve watched the same golfer, same swing, same session, produce a 14-degree difference in reported hip rotation at P6 purely by changing from a dark hoodie to a fitted polo. Nothing about the swing changed. The pixels changed.

What to wear:

Rain gear is the worst offender. If you’re filming in a waterproof, don’t bother analysing it, the hip and shoulder keypoints will be fiction. A £12 fitted polo in a colour that isn’t the colour of your trousers is genuinely the highest-ROI purchase in this entire article.

Same logic for the background. Standing in front of a green hedge in a green polo is the same problem at torso scale. A driving range bay with a dark net behind you and you in a light top is close to ideal.

Fault four: nothing in the clip establishes scale, and half of it is missing

Two related problems, both about what’s inside the frame.

Framing first. The AI needs your clubhead at the top of the backswing and your clubhead at the finish, both inside the frame. Crop either and the app either rejects the clip (Sportsbox is strict here and will tell you) or, more commonly, clips your swing arc at the frame boundary and reports a backswing that’s 20 degrees short. Leave a full club length of headroom above your hands at address. Film in portrait for down-the-line, landscape for face-on, and check the finish position specifically, because that’s the one people cut off.

Scale second, and this is the one that separates 2D from 3D tools. Anything reporting distances in real units (pelvis sway in inches, hand depth in centimetres) needs to convert pixels to metres. Sportsbox does this partly by asking for your height during setup. If you fudged your height, or you’re in a hat, or you skipped the calibration step, every linear measurement it gives you is scaled wrong by that proportion. A 6ft golfer entered as 5’10” gets sway readings about 3% short, consistently, forever.

Some apps want a calibration object. If yours asks for one, use it, don’t skip it because the swing looked good and you wanted to see the numbers.

And the frame needs to include roughly one second before you start back and one second after the finish. Motion-onset detection uses the still period to establish your address baseline. Start the video mid-waggle and the app anchors your address position to whatever it caught, which throws every subsequent angle off the same amount.

What good actually looks like

Here’s a capture spec you can screenshot into your notes:

CAMERA
  Frame rate     120 fps (60 minimum, never 30)
  Resolution     1080p  (4K adds nothing, costs frame rate)
  Mode           Standard video, NOT Slo-Mo
  Shutter        1/1000s or faster if your app allows manual
  
POSITION
  Down-the-line  On target line extended, 2.5-3.0 m back
  Face-on        Perpendicular to target line, 2.5-3.0 m
  Lens height    Hand height at address (~90 cm)
  Tilt           Zero. Level the phone.
  
LIGHT
  Best           Bright overcast, or indoor bay with even lighting
  Avoid          Low winter sun behind you, deep shade, dusk
  Why            Motion blur. Short exposure needs light.

SUBJECT
  Top and trousers in contrasting mid-tones
  One layer, fitted
  Full swing arc in frame including finish
  1 second static before takeaway, 1 second after finish

Shutter speed deserves a line of its own. Automatic exposure on a dull February afternoon in Yorkshire will happily select 1/60s, and at that exposure your hands travel 15cm while the sensor is open. They arrive as a smear. Pose models handle smeared limbs badly, and this is exactly why your indoor winter clips analyse worse than your summer ones despite identical setup. Apps like Filmic Pro (£14.99 one-off on iOS) give you manual shutter control. Failing that, film outdoors in decent light, or add light to your bay.

Practise the capture, then practise the swing

Film one down-the-line and one face-on clip with a 7-iron, following the spec above. Upload both. Before you read a single number, ask yourself whether the tool tracked you: most apps overlay a skeleton, and you can scrub frame by frame watching whether the hip and knee joints stay attached to your body through transition. If a joint jumps, twitches, or briefly appears in mid-air, throw out every number the app derived from that segment.

Then, and only then, look at the numbers that your capture quality actually supports. With a clean 120fps down-the-line clip you can trust shaft plane, hand path, head movement, and tempo ratio. With face-on at the same rate you get sway, lift, and lead wrist position. If you filmed at 60, treat impact-adjacent measurements as indicative rather than precise, and work off the P4 and P6 positions instead, which move slowly enough to be genuinely captured.

The golfer who fixes their capture usually discovers their swing was never the problem they thought it was. The early extension the app kept flagging turns out to have been a tilted tripod. That’s a slightly deflating discovery and a very useful one, because you’ve been drilling something that wasn’t broken while the actual fault sat unmeasured in the frames your camera never took.