Markerless 3D on a Phone: Useful or Marketing?

Open any of the current crop of 3D golf swing analysis app listings and you get the same promise: pelvis rotation, thorax rotation, hip-shoulder separation, pelvis sway, lead wrist extension, all from the phone already in your bag. No suit, no sensors, no £8,000 of Gears or a bay with four cameras bolted to the ceiling. Just film down the line, wait forty seconds, get a 3D skeleton you can spin around with your thumb.

The honest answer, after a season of pointing these things at the same swings, is that they are genuinely useful for about four numbers and quietly fictional about the fifteen they display with the most confidence. That is not a small thing. Four reliable numbers is more than most club golfers have ever had. But the apps do not tell you which four, and the interface gives a ±1° readout for lead wrist extension the same visual weight as a pelvis rotation figure that is actually trustworthy. That is where the marketing lives.

What single-camera 3D actually does

There is no depth sensor doing the work here. What you have is a 2D video and a neural network trained on very large datasets of humans, which infers where your joints are in three dimensions from a flat image. The model has learned what human bodies look like from every angle, so given a 2D projection it produces a plausible 3D pose. The technical term is monocular pose lifting, and the key word is “plausible.”

Think about what that means at impact. Your trail arm crosses in front of your torso. Your hands overlap. From a down-the-line camera at 240fps, there are frames where the network is guessing at a joint it cannot see, using a learned average of where that joint usually is on a human doing something that looks like this. It will produce a number. The number will have a decimal place. It is a prior, not a measurement.

Here’s the filter I use. If the movement is large, slow relative to frame rate, and happens in a plane the camera can see, the app is probably right. If the movement is small, fast, or requires resolving depth of a partially occluded segment, treat the number as decoration.

The four numbers you can actually use

Pelvis rotation at top of backswing. Big movement, 35 to 50 degrees for most club golfers, visible from either camera angle. Sportsbox 3D and the Hackmotion-adjacent 3D tools land within a few degrees of each other on this, and within roughly 5° of a marker-based reference in the published comparisons. Good enough to tell a 16-handicap that they are turning 22° and that is the problem.

Thorax rotation at top. Same reasoning. 75 to 95 degrees is a typical tour range, most amateurs are 55 to 75. An app reading you at 58° when you believe you are making a full turn is giving you real information.

Pelvis rotation at impact, or more usefully, rotation velocity ordering. Whether your pelvis peaks before your thorax is a gross temporal question. The sequence is either right or it is obviously not, and a 3D app reading down-the-line video gets the ordering right consistently even when the absolute peak velocity values wander by 15 to 20%.

Head and pelvis sway in the target direction. Centimetres of gross translation, tracked across a whole swing. Sportsbox reports pelvis sway in inches; the number moves in the right direction when you actually change something, and that is the test that matters.

Four things. Turn, upper turn, sequence, lateral movement. Notice what they have in common: they are all the coach’s eye, quantified. You could see all four on a phone video with no AI at all, just less precisely.

The numbers that are theatre

Lead wrist extension and radial deviation. This is the big one, because wrist condition is fashionable right now and it explains clubface behaviour better than almost anything else. HackMotion sells a physical sensor strapped to your lead wrist specifically because wrist angles are hard to measure. A single camera resolving your lead wrist to 2° of extension through a 100mph transition, with the trail hand wrapped over it, is not happening. I ran the same swing through a 3D app and a HackMotion on the same day:

Lead wrist extension at top of backswing
  HackMotion sensor:        48° extended
  3D app (DTL, 240fps):     31° extended

Lead wrist at impact
  HackMotion sensor:         9° extended
  3D app (DTL, 240fps):      2° flexed

Seventeen degrees of disagreement at the top, eleven at impact, and the two disagree about the direction of travel of the thing that actually controls your face angle. If you practise off the app number, you are practising toward someone else’s guess.

Hip-shoulder separation (the “X-factor”) to a single degree. This one is subtler, because both components are measurable. The problem is that separation is a difference of two noisy figures, so the errors add. If pelvis is ±4° and thorax is ±5°, your separation figure carries roughly ±6°. The app shows you 41°. It means somewhere in the high thirties to high forties. Chasing a 3° improvement in that number across a six-week block is chasing noise.

Anything involving the club. Several apps now overlay a shaft and report club path, face angle, attack angle from video. A 6° in-to-out path and a 2° in-to-out path look almost identical on a 240fps phone camera at impact, because the club travels roughly 15cm between frames. If you own a Garmin R10 or Rapsodo MLM2PRO, its radar-derived path number is imperfect (R10’s club path is the weakest thing it measures, easily ±3°) but it is still measuring the club rather than inferring it. Trust the launch monitor over the video overlay.

Pelvis rise and depth to the millimetre. Vertical and anterior-posterior pelvis motion, “posting up,” is real and matters. But depth is exactly the axis a single camera struggles with, and rise is often contaminated by the app’s estimate of your standing height. The shape of the curve is informative, the numbers on the axis are not.

Two worked examples

A 14-handicap, mid-fifties, chronic slice, filmed down the line on an iPhone 15 at 240fps from 8 feet, hosel-height. The app reported pelvis rotation 26° at top, thorax 61°, separation 35°, lead wrist extension 44° at top and 19° at impact, and “early extension: 4.2 inches.”

The usable read: he is not turning. 26° of pelvis is restricted, 61° of thorax against it means his separation is being generated mostly by coiling into a blocked pelvis. That is a real finding, and it matches what you see with your eyes. Give him a turn drill and a mobility screen.

The unusable read: the 44° wrist extension. If that were accurate, he would be wide open at the top and his slice would have a clean mechanical explanation. Put a HackMotion on him and it came out nearer 60°, which changes the prescription from “quiet the hands” to “genuinely reduce cupping in the takeaway.” Same golfer, same swing, different practice plan, because one device measured the wrist and the other inferred it.

Second example. An 8-handicap filmed face-on, chasing sequencing. Session one: the app reported peak pelvis rotational velocity 340°/s and peak thorax 580°/s, with pelvis peaking 0.06s before thorax. Session two, three weeks later, after genuine work on ground sequencing: peak pelvis 395°/s, thorax 610°/s, gap 0.09s.

Here is the thing. The absolute velocities are almost certainly wrong. Marker-based systems typically put amateur peak pelvis velocity higher than 340°/s, and the app’s frame-differencing approach systematically underestimates peaks. But the direction of change and the widening gap are real, because the same error is present in both sessions. That is the most valuable property these apps have, and it is barely mentioned in the marketing: they are decent at deltas and poor at absolutes. Film in the same spot, with the same phone, at the same frame rate, and the change between sessions carries information even when the readings themselves do not.

Setup discipline changes the numbers more than the app does

Move the camera a foot closer and the perspective distortion changes the inferred pose. Film from 4 feet high instead of hip height and your reported spine tilt moves by several degrees. Drop from 240fps to 60fps and the club travels 60cm between frames near impact, so every impact-frame number becomes an interpolation. Wear a baggy waterproof and the pelvis estimate gets worse, because the model is finding the outline of the jacket.

Practical rules that cost nothing: same spot every time, mark it with a tee or a tripod footprint. 240fps minimum, which every iPhone since the 8 and most recent Androids do. Tight clothing, ideally a plain shirt tucked in. Full body in frame with a bit of headroom. Bright, even light: shadows across the torso genuinely degrade joint detection. Film three swings and use the middle one, because run-to-run variation on the same app with the same golfer routinely exceeds 5° on rotation figures.

If you want a sense of how much of what these apps report survives contact with a second opinion, the comparison in AI Swing Analysis Apps, Tested on the Same Swing puts several of them against one identical recording, which is the only way to see how differently they read the same body.

So what do you actually practise?

Take the gross rotation numbers and treat them as a scoreboard for turn. If your pelvis is under 30° at the top, work on that, and use the app to confirm the change over weeks rather than swings. Use the sequencing order as a pass or fail, not as a set of targets. Use sway and head movement as trend lines.

For wrist condition, either buy the sensor or use the old, reliable method: a face-on video, a still at the top, and an honest look at whether the back of the lead hand is cupped. Your eyes on a still frame are more trustworthy than a model guessing at an occluded joint, and they are free.

Club delivery belongs to your launch monitor. Even a £450 radar gives you ball flight numbers that are grounded in physics, and if your R10 says your path is 5° out and your face is 2° open, that is a better description of your slice than any pose estimate.

The apps are not a con. £150 a year for a tool that reliably tells a 20-handicap they are turning their hips half as far as they think is decent value, and the coaching-communication side of it (seeing your own skeleton next to a reference) genuinely moves people faster than words do. What is oversold is the precision, and the precision is what the pricing page leads with. You are buying a good rotation measurement wrapped in twenty numbers that look identical to it and are not.

Two years from now, multi-frame temporal models and better training data will probably drag wrist and club numbers into usable territory. They are not there in September 2026, and an app that presented its own confidence intervals would be doing you a favour it currently declines to do.