How far can you trust a joint angle that no marker ever touched? Two 2026 reviews and two validation studies converge on an uncomfortable answer: accuracy is not one number. It depends on the joint, the plane of motion, and the task — and it can collapse the moment conditions stop being ideal. Here is where markerless kinematics holds up, and where it does not.
Table of Contents
Accuracy is not one number
It is tempting to ask for a single accuracy figure for markerless motion capture — one number to trust or distrust. But the evidence refuses to give one. Pooled across many studies, the best-supported case is narrow: lower-limb sagittal joint angles. Move off that axis — to the upper limb, out-of-plane rotations, or joint kinetics and forces — and the support thins out fast. And the conditions themselves change the answer: lighting, occlusion, camera count and the task being performed all shift the error. Accuracy here is a property of a setup, not of a method.
Paper 01 — Twenty-two validation studies, pooled
The first review pools 22 validation studies, screened down from 1,676 records, spanning tools from Kinect and OpenPose to Pose2Sim, OpenCap and Theia3D. The pattern that emerges is consistent: lower-limb sagittal kinematics are best supported, while kinetic evidence — joint moments and forces — is sparser and more variable. The takeaway is not a single accuracy score but a map of where the confidence actually lives.
Paper 02 — Body-worn or camera-based for joint angles?
The second review sets IMUs against markerless vision, side by side. For sagittal lower-limb gait, body-worn inertial sensors land at roughly 3–6° of error; vision-based systems are competitive but degrade under occlusion and motion blur. The review’s sharper point is that error is not a single sensor’s fault — it comes from the whole measurement chain, from capture through modelling to the reported angle.
Paper 03 — Cutting an 18-camera rig down to six
How much does camera count actually matter? This validation study took an 18-camera markerless setup and cut it to 12, then to six, comparing walking, running and jumping. For walking, joint angles shifted under 2° — a reassuring result for anyone working with a modest rig. Faster tasks (running, jumping) changed more, but the changes stayed modest. Fewer cameras cost less than you might fear, at least for steady gait.
Paper 04 — When the gait pattern stops being steady
The counterweight. Using the SMARTGAIT markerless system on smartphone video, researchers tested five adults (mean age about 83) walking normally and under unexpected perturbation. When the gait pattern stopped being steady, frontal-view gait-event detection fell from an F1 of 0.68 to 0.49 — nearly halved. The lesson pairs with Paper 03: geometry (camera count) is forgiving, but behaviour — a disturbed, non-steady pattern — is not.
Validity is conditional
Put together, the four studies say the same thing from different angles: markerless validity is conditional — specific to the joint, the plane and the task. Cutting an 18-camera rig to six barely moved walking angles; perturbing the gait pattern halved smartphone event detection. The number you cite is only as good as the conditions it was measured in. For a concrete accuracy comparison across estimators, see our breakdown of pose estimation accuracy across six models, and for how two pipelines fare with and without ground truth, validated or not.
Frequently asked questions
Is markerless motion capture accurate enough for research?
For lower-limb sagittal joint angles, the pooled evidence is supportive — that is the best-established case. For joint kinetics, out-of-plane rotations, or the upper limb, the evidence is sparser and more variable. Accuracy is joint-, plane- and task-specific, so “accurate enough” depends entirely on which quantity you need.
How many cameras does a markerless system need?
Fewer than you might think for steady gait. One 2026 study cut an 18-camera rig to six and saw walking joint angles shift under 2°; running and jumping changed more, but modestly. Camera count is relatively forgiving — the bigger risk is non-ideal conditions like occlusion or perturbed movement.
Are IMUs or cameras better for joint angles?
Neither wins outright. For sagittal lower-limb gait, IMUs sit around 3–6° of error; markerless vision is competitive but degrades under occlusion and blur. Error comes from the whole measurement chain, not just the sensor, so the right choice depends on the setting and the joint.
References
[1] Shubbar, A. A. A. (2026). Artificial Intelligence for Markerless Motion Analysis in Sports Biomechanics: A Systematic Review of Validation Studies, Current Challenges, and Future Directions. SportRxiv.
[2] Ceriola, L., Molinaro, L., Taborri, J., Patane, F., & Milani, I. (2026). IMU- and Vision-Based Measurement Techniques for Joint Kinematics: A Narrative Review. Sensors, 26(16), 5063.
[3] Walker, R., Robinson, M. A., Outerleys, J., et al. (2026). The effect of camera configuration on multi-camera markerless motion capture for biomechanical analysis. Journal of Biomechanics.
[4] Graf, V., Haug, V., Seebacher, D., et al. (2026). Validation of AI-based markerless gait event detection during perturbed walking using smartphone videos from two camera perspectives. Frontiers in Sports and Active Living.
Takashi Fukushima — Sports Science & Pose Estimation.
▶ Subscribe on YouTube · Website · ORCID · Contact