Study at a glance
Summary
Disclaimer: The following summary details independent academic research. Theia3D is intended for research, sports performance and educational use only. It is not cleared as a medical device and is not intended to be used for the diagnosis, treatment, mitigation or prevention of any disease, injury or medical condition.
Editor's note: Theia3D version v2024 (Theia3D Apollo), as stated in the paper.
Study Overview
Design
A systematic evaluation of how camera geometry affects agreement between a two-camera markerless motion-capture pipeline and an eight-camera Theia3D reference system, using all 28 possible two-camera combinations drawn from the same eight-camera array across five athletic tasks.
Participants
Twenty-three female varsity soccer players (mean age 20.40 years; height 1.66 m; mass 63.31 kg) cleared for athletic participation with no current musculoskeletal condition or concussion.
Tasks and Data Collection
Participants completed a single trial of Y-balance (bilateral), one trial of L-hop (bilateral), three repetitions of single-leg and bilateral countermovement jumps, three broad jumps, and three repetitions (from a 35-repetition set) of bodyweight squats. An eight-camera Vicon Vue system (1,920 x 1,080 pixels, 60 Hz) recorded all trials, processed through Theia3D using its default 17-segment model to generate reference kinematics. The same video was independently reprocessed for every one of the 28 possible two-camera combinations using a separate markerless pipeline built around 2D pose detection, triangulation, an LSTM-based marker-augmentation model, and OpenSim inverse kinematics.
Analysis
Camera geometry for each two-camera combination was quantified from calibration data, including camera height, distance to origin, azimuth, field of view, and horizontal angular separation, and combinations were grouped into six geometry subgroups: Front, Back, Left, Right, Same Quadrant, and Diagonal. Agreement between each two-camera combination and the Theia3D reference, and consistency among two-camera configurations, were assessed using root mean square error (RMSE), mean absolute error, bias, limits of agreement, and the coefficient of multiple correlation (CMC).
Key Findings
- Front and Back camera configurations, with moderate horizontal angular separation, showed the most favorable agreement with the Theia3D reference, particularly for sagittal-plane hip and knee kinematics (RMSE under 8 degrees, CMC approaching 1.00).
- Same Quadrant camera configurations, with the smallest angular separation of any group, showed the poorest agreement, especially for frontal- and transverse-plane joint angles (RMSE exceeding 10 degrees, CMC under 0.60).
- Diagonal camera configurations had the largest angular separation of any group but did not consistently produce better agreement, suggesting that excessive angular separation alone does not guarantee improved reconstruction and may reduce shared visibility between the two views.
- Y-balance and L-hop tasks showed reduced agreement and higher trial dropout than the other tasks, attributed primarily to self-occlusion during limb-crossing movements and participants moving outside one or both cameras' field of view during these larger-displacement tasks.
- Ankle joint angles showed consistently poor agreement across nearly every camera configuration, and hip flexion/extension showed a large but highly consistent systematic offset (RMSE of 17 to 27 degrees despite waveform similarity with CMC above 0.90); the authors attribute both patterns substantially to differences in how the two systems define ankle and pelvic segment coordinate systems, rather than to camera geometry or reconstruction error alone.
What This Means for Movement Science Researchers and Sports Performance Teams
For teams considering a lower-cost, two-camera markerless setup for sport-specific movement assessment, this study offers concrete, geometry-specific guidance: front-or-back camera placement with moderate angular separation is best suited to bilateral, sagittal-dominant tasks like squats and jumps, same-quadrant placement should be avoided, and tasks involving limb crossing or large lateral movement, like balance or hop tests, call for camera placement that prioritizes continuous visibility of the foot and ankle and full coverage of the movement path over angular separation alone. The large, consistent hip flexion offset observed here is also a useful reminder that apparent disagreement between two markerless systems can reflect differences in how each one defines its underlying segment coordinate systems, not necessarily a difference in the movement each is capturing.
The Role of Theia3D in Movement Analysis
In this study, Theia3D served as the high-quality eight-camera markerless reference system against which a separate, custom-built two-camera markerless pipeline was evaluated, rather than as the system under direct validation itself. The authors are explicit that this role means the study's reported results reflect agreement with Theia3D specifically, not an assessment of absolute kinematic accuracy against a true gold standard, and that future work should validate reduced-camera workflows directly against marker-based motion capture as well.
Study Limitations and Future Directions
The authors note several limitations. The eight-camera Theia3D system, while a high-quality markerless reference, should not be interpreted as a definitive biomechanical gold standard, so the reported agreement reflects consistency with Theia3D rather than absolute accuracy; some of the observed discrepancies may also reflect differences in model and joint coordinate system definitions between the two systems rather than true reconstruction error. Camera-to-participant distance, focal length and zoom, lighting, clothing, and task-specific segment visibility may all have contributed to configuration-dependent differences beyond angular separation alone. The two-camera workflow remained sensitive to occlusion, field-of-view loss, and limited segment visibility, particularly for frontal- and transverse-plane motions and tasks involving limb crossing, such as the Y-balance task. The findings are specific to the particular two-camera pose-estimation and modeling pipeline used in this study and may not generalize to other markerless systems, since camera-orientation effects can be algorithm- and task-dependent. The sample consisted only of healthy female varsity soccer athletes, limiting generalizability to male athletes, youth or older adults, populations with movement-related conditions, and individuals with altered movement patterns.
Citation: Zhao, X., Zhang, Y., & Graham, R. B. (2026). Influence of camera geometry on 3D joint angle estimation in markerless motion capture. Frontiers in Sports and Active Living, 8, 1850018. https://doi.org/10.3389/fspor.2026.1850018
Link: Read the full study
How Should You Place Your Cameras in a Reduced-Camera Markerless Setup?
This study shows that, for a two-camera markerless pipeline, where you place your cameras can matter as much as how many you use, with front-and-back placement favoring sagittal-plane accuracy in bilateral tasks and same-quadrant placement consistently underperforming. If your work involves designing a reduced-camera or portable markerless setup and you want to understand how camera geometry could affect your own results, talk to our team about what this study could mean for your camera configuration.