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: v2023.1.0.3161, patch 14 (explicitly stated in the paper).
Study Overview
Design
This was a validation study using a within-subject, simultaneous-measurement design. Each volunteer completed two 30-second bipedal quiet stance trials, one with eyes open and one with eyes closed, while Theia3D and Vicon recorded synchronously using a shared hardware clock to ensure frame-level temporal alignment. The resulting kinematic data from each system were independently processed into principal movements via PCA, and the two systems’ principal movement structures, variance distributions, and sensitivity to the eyes-open/eyes-closed manipulation were then statistically and qualitatively compared.
Participants
A convenience sample of 13 healthy university students participated (4 males; age 25.0 ± 1.6 years; height 174.9 ± 9.8 cm; weight 73.1 ± 11.9 kg). This study was part of a larger validation project involving 26 volunteers performing a range of movement tasks; participants were included in this specific analysis only if their recordings were complete and valid, since hardware constraints caused frame loss in several high-speed video recordings. The final analyzed sample reflects this completeness-based selection, not data quality or performance. All participants provided informed consent, and the study protocol was approved by the responsible institutional ethics review board (certificate 28/2024).
Data Collection
Participants were equipped with 53 retroreflective markers representing a full-body marker set for the Vicon system (8 Vicon Bonita 10 and 2 Vicon Vero v2.2 cameras, 10 cameras total, 200 Hz). No separate preparation was required for the Theia3D system, which used an 8-camera Blackfly S (Teledyne FLIR Systems) setup at 100 Hz. Both systems were operated and calibrated within Vicon Nexus software and recorded synchronously using a shared hardware clock, with temporal consistency further verified by cross-checking trial start and end timestamps and visually inspecting synchronized marker and trajectory motion. Volunteers completed two 30-second bipedal stance trials, once with eyes open and once with eyes closed, standing upright with arms hanging loosely and gaze directed straight ahead. Vicon data were labeled, gap-filled, and exported as 3D coordinates in a lab-based external reference frame; Theia3D (version 2023.1.0.3161, patch 14) processed video images from its 8 cameras to export 18 reference points in the same lab-based reference frame, interpolating its data to a 200 Hz frame rate to match Vicon.
Analysis
Data from both systems were processed using the MATLAB app “PManalyzer” to remove the first and last 5 seconds of each trial, normalize each trial to unit mean Euclidean norm, and concatenate data across trials and volunteers before submitting it to PCA. The primary analysis used unfiltered kinematic data, a deliberate choice intended to assess native system output without preprocessing that could artificially inflate agreement; a secondary, exploratory analysis applying a 5 Hz low-pass filter was also conducted for comparison. Eigenvectors (principal movement components) and their variance explained were compared via scree plots and qualitative visual comparison; score time series (how each principal movement evolved over the recording) were compared using Pearson correlation coefficients, calculated per subject and trial and then averaged. Sensitivity to the eyes-open/eyes-closed manipulation was assessed using two secondary variables derived from the score time series: PP-rVAR (relative subject-specific variance, reflecting postural movement structure) and PV-STD (standard deviation of movement speed), compared using paired t-tests or Wilcoxon tests depending on normality, with Cohen’s d or rank-biserial correlation effect sizes.
Key Findings
- The two dominant postural movement patterns were nearly identical between systems. The first principal movement, representing anterior-posterior ankle-strategy sway, explained 74% of variance in Vicon and 76% in Theia3D, with near-perfect correlation (r = 99.8%) between systems. The second principal movement, representing medio-lateral sway, explained 9.4% (Vicon) and 9.9% (Theia3D) of variance, with similarly strong correlation (r = 99.0%).
- Agreement remained strong across the seven movement components that captured the bulk of postural variance. The first seven principal movements collectively accounted for 97.0% (Vicon) and 96.5% (Theia3D) of total variance, with correlation coefficients between systems ranging from r = 0.814 to 0.928 despite fundamental differences in how each system places and tracks markers or landmarks.
- Higher-order components showed weaker agreement and more Theia3D-specific noise. Components beyond the first seven each contributed less than 1% of variance individually, and correlation between systems declined in these smaller components, with Theia3D-derived time series showing visibly higher noise than the smoother Vicon-derived signals, noise the authors attribute to pose estimation variability such as occlusions, background artifacts, or joint misclassification inherent to markerless tracking.
- Both systems detected the same expected postural response to closing the eyes. The structural variable (PP-rVAR) showed no significant differences between eyes-open and eyes-closed conditions in either system, indicating postural movement structure remained stable; the movement-speed variable (PV-STD) showed statistically significant increases in sway speed with eyes closed in both systems, consistent with established literature on visual input and postural stability.
- Filtering improved agreement in higher-order components but altered what those components represented. An exploratory analysis applying a 5 Hz low-pass filter before PCA increased correlation coefficients between corresponding principal movements, but also shifted the orientation of higher-order eigenvectors, meaning filtered and unfiltered results captured somewhat different postural features and were not directly comparable, a key reason the authors based their primary findings on unfiltered data.
What This Means for Postural Control and Balance Researchers
For researchers using or considering PCA-based “principal movement” analysis of postural control, this study provides direct evidence that Theia3D can substitute for marker-based systems like Vicon when the focus is on dominant postural strategies, ankle- and hip-driven sway patterns that together capture the large majority of postural variance. That said, the findings also draw a clear boundary: agreement weakens meaningfully in smaller, higher-order movement components, which are more susceptible to markerless-specific noise. Researchers whose questions depend on these finer-grained components, or who are working with naturally small-amplitude, low-velocity movements like quiet standing, should treat Theia3D output with appropriate caution and consider whether filtering or other preprocessing is warranted for their specific application, while recognizing that filtering itself can alter the structure of higher-order components.
The Role of Theia3D in Movement Analysis
In this study, Theia3D served as the markerless motion capture system under validation, generating whole-body 3D pose estimates from an 8-camera video array without requiring markers or extensive setup time. Its output, 18 reference points per frame, synchronized with Vicon’s 53-marker data in the same lab-based reference frame, was processed independently through the same PCA-based analytical pipeline used for the marker-based data, allowing the two systems’ resulting principal movement components to be directly compared. This design let the researchers isolate how much of any disagreement between systems stemmed from Theia3D’s underlying motion tracking versus the broader analytical approach, since both datasets went through an identical PCA workflow.
Study Limitations and Future Directions
The authors note several limitations. Hardware-related frame loss in several video recordings reduced the usable sample from a larger pool of 26 volunteers down to 13 valid datasets; while exclusions were based solely on recording completeness rather than data quality or performance, the authors acknowledge that potential bias from this selection cannot be entirely ruled out. Theia3D-derived data showed higher noise than Vicon, particularly in higher-order components, and because the primary analysis intentionally used unfiltered data to preserve native system output for validation purposes, this noise may have affected the precision of some results. The study also focused exclusively on quiet bipedal stance, a relatively small-amplitude, low-velocity task; while this makes for a particularly rigorous test of signal fidelity, the authors note that further validation is needed for more dynamic postural tasks, such as perturbed stance or gait. The eyes-open and eyes-closed conditions were always conducted in the same fixed order rather than randomized, so the authors cannot fully exclude the possibility that order effects, such as learning, adaptation, or fatigue, contributed to the observed differences between conditions. Finally, visual inspection of principal movement animations was used to support interpretation of higher-order components but was not used as a primary validation criterion, and the authors explicitly acknowledge the subjectivity of that approach.
Citation: Federolf, P., Kühne, M., Schiel, K., Reimeir, E., Debertin, D., Calisti, M., & Mohr, M. (2025). Validation of markerless (Theia3D) against marker-based (Vicon) motion capture data of postural control movements analyzed through principal component analysis. Journal of Biomechanics, 189, 112831.
Link: https://doi.org/10.1016/j.jbiomech.2025.112831
Could Markerless Capture Simplify Your Postural Control Research?
Quiet standing is a deceptively demanding test for any motion capture system, and this study shows Theia3D holding up against marker-based capture even in that subtle, low-amplitude setting. If your research involves balance, postural strategies, or principal-component-based movement analysis, talk to our team about how markerless capture could streamline your protocol without compromising the data you depend on.