3D motion capture accuracy versus 2D swing video

What the marker suit is actually buying

0.2 mm, under 0.2°Vicon repeatability quoted in the golf study

A marker-based golf laboratory is a room of infrared cameras looking at retroreflective balls. In Aimée Smith's Loughborough study, fifteen golfers off handicaps of 1 to 29 hit their own drivers inside a thirteen-camera Vicon running at 250 Hz, wearing fifty-five 14 mm markers, with five more on the club and reflective tape on the ball so the instant of impact was found rather than guessed. The rig repeated known distances to within 0.2 mm and known angles to under 0.2 degrees.

Gears, a golf-specific system, publishes a comparable claim: eight cameras at 360 frames per second, measurements as fine as 0.2 mm, with the grip and the head both tracked, so shaft deflection is measured rather than inferred. What that money buys is rotation about the long axis of a body segment, and a clock fast enough that contact is a recorded frame rather than the gap between two.

Seventeen degrees, from the same set of markers

16.72 ± 6.20°2D minus 3D X-factor at the top, fifteen golfers

Smith's group then did something unusual. They took one set of trials and worked out the trunk numbers twice: once from full 3D segment angles, once by projecting the same marker vectors flat onto the laboratory planes, which is what a single camera effectively does to a golfer. No marker moved. At the top of the backswing the two methods put X-factor 16.72 ± 6.20 degrees apart, the flattened figure being the larger.

Trunk flexion came out roughly 7 degrees greater flattened (−6.54 ± 3.64, p < .001), and lateral bend toward the trail side 8.65 ± 2.91 degrees smaller at impact (p = .002). Axial rotation at the top differed by 10.97 ± 5.89 degrees, short of significance at p = .068. None of it is a constant you could subtract afterwards: one golfer off 9 differed by about 25 degrees, another off 2 by about 5. Two earlier papers had put the gap near 3 degrees, and near 0.4 the other way.

What survives being flattened

Not everything falls apart. Smith reports the 2D and 3D curves as similar in shape across the swing, the differences at takeaway small and never significant: flattening damages magnitude far more than pattern. Barzyk and colleagues filmed eleven people jumping on an iPhone X at 4K and 60 frames per second, beside a twelve-camera Vicon: sagittal-plane correlations ran 0.96 at the hip, 0.99 at the knee and 0.87 at the ankle, with mean absolute errors of 2.6, 2.1 and 4.5 degrees.

Angles outside that plane were not validated at all. The Kinovea work is narrower still than the way it gets quoted. Puig-Diví's team filmed a wire model of a lower limb swinging like a pendulum from 5 m and compared the coordinates against the same shape drawn in AutoCAD. Error was smallest square-on at 90 degrees and largest at 45. A rigid wire has no skin to slide and no depth to lose, so read it as a test of the software, not of a swing.

Markerless 3D lands between the two

8.5° and 13°Markerless against markers, long-axis rotation

Kanko's group at Queen's ran both instruments at once on thirty walkers: eight video cameras feeding a markerless algorithm, seven infrared cameras tracking markers, the same trials. Joint centres landed 2.4 cm apart at the ankle, 2.2 cm at the knee and 3.6 cm at the hip, and on the axes carrying the large rotations the segment angles agreed within about 1 to 2 degrees. The exception is the one that matters to a golfer.

Rotation about the long axis of a segment, the axis your shoulder turn and hip turn live on, came out 8.5 degrees apart at the thigh and 13 degrees at the shank. One camera fares worse again. Zhao's team tested Capture4D, a monocular system, against Qualisys on nine tennis servers and reported joint-position error of 69.5 to 88.3 mm, worst at the trail elbow when the arm ran along the camera's optical axis. The paper calls 70 to 130 mm acceptable for coaching without citing where that band came from.

The reference has an error bar too

Kanko's discussion is candid about what marker-based capture gets wrong. Markers placed by different operators following different protocols have been shown to shift lower-limb joint angles by up to 5 degrees. Skin-mounted clusters slide over the bone beneath them, translating up to 1.5 cm at the shank and 2.5 cm at the thigh and rotating up to 8 degrees. Estimates of where a joint centre sits run as far as 5.1 cm out.

Which is why the authors decline to call the differences errors: ground truth is not known, they write, and both systems contribute. Pebble Swing reads joint angles out of the frames and the backswing-to-downswing tempo ratio out of the motion, and any figure resting on rotation toward or away from the lens is a bracket. So distrust a single turn number in degrees, from any instrument. A three-degree change sits inside every error band above. A twenty-degree change does not.