SSMT-2026-0006 — Markerless pose estimation from ordinary video recovers joint kinemati

SSMT-2026-0006v12026-07-26 Type M · Measurement Computer Vision draft

Claim, as marketed

“Lab-grade biomechanics from your phone.”

Canonical claim

Markerless pose estimation from ordinary video recovers joint kinematics with accuracy comparable to marker-based optical motion capture.

Scope

Verdict differs sharply by configuration (8-camera vs 2-phone vs single-camera) and by plane. A vendor restricting its claim to sagittal-plane and spatiotemporal outputs would be Supported today.

Verdict

Partially supported Genuinely good in the sagittal plane; not usable for rotations; and 'from your phone' is the least-evidenced configuration.

Real evidence, narrower than the claim.

Claim drift — 4 of 7

▲Population▲Setting–Comparator▲Outcome▲Magnitude–Reproducibility–Independence

Reproducibility is aligned and unusually good — test-retest RMSD 3.0° ± 1.0 is better than validity error, which is the right shape for tracking within-person change.

Evidence located

~15

primary studies examined

4/6

key studies independent

true single-camera study (n=19)

largest n vs criterion

Findings

Reliable — at or near marker-based noise. Spatiotemporal parameters (gait speed, step length, stance time) rate GRADE High and are the strongest evidence in the field. Knee flexion/extension: RMSD 3.3–3.4° in gait, 4.2–4.8° running. Hip abduction/adduction: 2.6–3.0°.

Not reliable. All transverse-plane rotations — knee internal/external rotation 13.2° in gait, hip rotation 6.9–13.6°, ankle rotation 11.6°. One review reports knee transverse ICC of 0.11–0.13. These errors exceed the physiological range of the movement itself.

Out-of-plane generally. Across eight dynamic tasks (n=41), normalised RMSE ran 29–136% with correlations from −0.09 to 0.80. A negative correlation means the system can report the trend backwards.

Counter-intuitive weak point. Hip flexion/extension is one of the worst sagittal angles (10–11° gait, 12.1–20.6° jumping), attributed to pelvis-segment definition differences rather than tracking failure.

Frontal-plane knee matters commercially. 3.2° in gait but 6.0–9.8° in direct comparison — and dynamic knee valgus is a headline marketed use case.

Single vs multi-camera — the surprising part. True monocular achieved RMSD 5.5° ± 1.1, statistically indistinguishable from a two-camera rig (p > 0.05). But that is one independent study, n=19, healthy adults, simulated pathology, on a lab floor. Where camera count has been tested directly it still matters: significant deviations fell from 16 degrees of freedom at 2 cameras to 9 at 8 cameras, and one three-way comparison found single-view ~3.5× worse at the ankle (15.23° vs 4.36°) — a finding that came from a vendor-funded study, i.e. a concession against interest.

What would change this verdict

Two or more independent, adequately powered (n≥40) studies of true monocular consumer-phone video against marker-based mocap in real patient populations and uncontrolled settings (variable lighting, ordinary clothing), showing sagittal RMSE below 5° and frontal-plane knee below 5°. Much cheaper alternative: a vendor narrowing its claim to sagittal and spatiotemporal outputs, which the current evidence already supports.

Practitioner read

Phone-based motion capture is genuinely useful for what you would measure with a stopwatch and a side-on camera — walking speed, step length, knee bend — where it is within a few degrees of a research lab. It is not trustworthy for rotations or frontal-plane knee position; errors there run 10–20 degrees and can point the wrong way, so do not use it to judge knee valgus for return-to-sport clearance. Good for tracking whether this patient's sagittal pattern is changing over time.

Could not be verified

Six items returned HTTP 403; their numbers come from abstracts or institutional repositories rather than full text.
One frequently-quoted upper-body figure (CMC 0.48, RMSE 36.7°) appeared only in a search snippet and could not be traced to a primary source — do not use it.
One headline RMSE is reported inconsistently across secondary sources (5.8 ± 1.8 vs 5.9 ± 1.2) and was not adjudicated.

Listed rather than dropped. Counts are floors, not censuses.

Sources

Uhlrich SD, et al. (2023). PLOS Comput Biol 19(10):e1011462 — OpenCap, developer-authored ↗
Varcin F, Boocock MG (2026). Artif Intell Med 173:103332 — systematic review, 16 studies ↗
Horsak B, et al. (2025). J Biomech 193:112986 — monocular vs 2-camera, n=19 ↗
Liang H, et al. (2026). Sports Biomech — n=41, eight dynamic tasks ↗
Park YS, et al. (2026). Medicina 62(2):418 — three-way, vendor-funded ↗
Scataglini S, et al. (2024). Sensors 24:3686 — systematic review & meta-analysis ↗

Every source must resolve at a DOI or PubMed ID.

Changelog

v1 · 2026-07-26 · Initial grade.

Cite this card

SSMT. "Markerless pose estimation from ordinary video recovers joint kinematics with accuracy comparable to marker-based optical motion capture." Claim SSMT-2026-0006 v1, graded 2026-07-26. https://sportssciencetech.com/c/SSMT-2026-0006

Disagree with this grade? Every card carries “what would change this verdict” so the specification is explicit. Send evidence and it will be logged and the card re-versioned in public — including when it moves a grade in a manufacturer's favour. Submit evidence →