Study at a glance

In 202 children with complex gait patterns, one of the largest pediatric comparisons of its kind to date, Theia3D showed good sagittal-plane agreement with marker-based motion capture, including strong gait-classification concordance, but systematically underestimated hip and knee rotation and compressed the between-child variability seen in the frontal and transverse planes.

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 v2024.1.0.4409 (Theia3D Apollo), as stated in the paper.

Study Overview

Design

A concurrent, within-visit comparison of marker-based and Theia3D markerless motion capture in children undergoing a gait assessment visit, evaluating agreement in joint angle trajectories across the gait cycle and in an established gait pattern classification system derived from sagittal knee and ankle kinematics.

Participants

Two hundred and two children with a heterogeneous range of underlying diagnoses (mean age 12.1 years; 115 boys, 87 girls), drawn from 72 unique diagnoses, most commonly cerebral palsy, along with talipes equinovarus, arthrogryposis, and spina bifida, among others. Most participants ambulated without an assistive device; a smaller subset used an assistive device, predominantly a posterior walker.

Data Collection

Each child completed a static standing trial, a dynamic walking trial used to verify marker placement, and a series of dynamic walking trials, all in a barefoot condition while using any necessary ambulation aid, with at least three complete strides per leg retained for analysis. Reflective markers were placed on anatomical landmarks by trained movement specialists following an established gait model template, and marker-based data were captured using twelve Vicon Vantage V16 infrared cameras. Markerless video was captured simultaneously using eight FLIR Blackfly S cameras and processed through Theia3D (Apollo, v2024.1.0.4409) using its default upper and lower body model with three degrees of freedom enabled at the knee, yielding twelve bilateral kinematic variables across the pelvis, hip, knee, and ankle.

Analysis

Time-continuous kinematic trajectories were compared between systems using statistical parametric mapping with false-discovery-rate correction, alongside per-participant root-mean-square error (RMSE), normalized RMSE, and Cohen's d effect sizes, bootstrapped across the sample. A linear mixed-effects model tested whether age, body mass index, sex, and assistive device use predicted RMSE between systems. Each child's gait was additionally classified into an established sagittal-plane gait pattern category, based on knee and ankle z-scores relative to a normative dataset, and agreement between the two systems' classifications was assessed using Cohen's kappa.

Key Findings

  • Sagittal-plane agreement was the strongest of the three planes tested, with RMSEs below 8 degrees at the hip, knee, and ankle, and below 5 degrees at the knee and ankle specifically. Effect sizes for these sagittal-plane measures, along with pelvic obliquity and hip abduction/adduction, were small (Cohen's d between 0.03 and 0.37).
  • Frontal- and transverse-plane agreement was considerably weaker, with transverse-plane RMSEs at the hip, knee, and ankle all exceeding 10 degrees on average, and moderate effect sizes (Cohen's d between 0.34 and 0.61); markerless motion capture also showed systematically less rotation and consistently underestimated pelvic tilt compared with the marker-based system.
  • Markerless output showed markedly reduced between-child variability in hip and knee rotation compared with marker-based data, and this pattern held up on direct inspection of children with pronounced hip rotation in the study cohort. Among children selected for having some of the highest degrees of hip rotation in the sample, RMSEs rose to roughly 20 to 25 degrees on average, compared with just over 10 degrees across the full cohort, with the marker-based system appearing to more plausibly capture these pronounced rotational deviations.
  • Gait pattern classification showed moderate overall agreement between systems (kappa = 0.60; 67% trial-level concordance), with agreement varying considerably by specific pattern. Percent agreement was highest for true equinus (83%), jump gait (77%), and crouch gait (69%), and lowest for recurvatum (52%) and ankle crouch (23%); underlying knee and ankle classification agreement was somewhat stronger (kappa = 0.63 and 0.68, respectively).
  • Higher body mass index and use of an assistive device were both associated with greater RMSE between systems, while older age was associated with modestly lower RMSE. Each one-unit increase in BMI was associated with roughly a 1.6% increase in average error, assistive device use was associated with about an 18% increase in average error, and each additional year of age was associated with roughly a 1.5% reduction in average error.

What This Means for Movement Science Researchers and Care Teams

For teams considering markerless motion capture as part of a pediatric gait assessment service, this large-cohort comparison offers a plane-specific answer: Theia3D's sagittal-plane kinematics and gait classification showed good enough agreement with marker-based motion capture to support reasonable confidence in sagittal-plane analyses and the efficiency and setup-time benefits that come with eliminating reflective markers, particularly valuable for a pediatric population where minimizing visit burden matters. The clearer caution is in the frontal and especially transverse planes, where systematic underestimation of hip and knee rotation, and the apparent suppression of true between-child variability, mean rotational findings from markerless data should currently be treated with caution rather than used as a stand-in for marker-based measurement in care-planning decisions that depend on precise axial-plane assessment.

The Role of Theia3D in Movement Analysis

Theia3D was directly evaluated throughout this study, with its sagittal, frontal, and transverse joint kinematics compared against a simultaneously recorded marker-based reference across one of the largest and most diagnostically varied pediatric cohorts tested to date. Because this study also compared how the two systems classified each child's gait pattern, not just the raw joint angles, it offers a particularly practical picture of where Theia3D's current pediatric performance supports confident use and where further model development is still needed, especially for children with pronounced rotational gait deviations.

Study Limitations and Future Directions

The authors note that marker-based motion capture itself is not an objective, error-free reference; it depends on modeling assumptions such as hip joint center localization that carry their own uncertainty, which complicates attributing every observed difference specifically to the markerless system. All data in this study were collected in a single laboratory, which limited the researchers' ability to assess how lighting or camera configuration, both previously identified as factors affecting markerless agreement, might influence results in other settings. While the diagnostic makeup of the cohort was broad, diagnoses were too sparsely distributed across specific conditions to allow the researchers to stratify their statistical models by diagnosis in this analysis. The authors suggest that markerless motion capture's apparent underestimation of hip and knee rotation, and its reduced between-child variability in these planes, likely reflects the training data used to build the underlying pose-estimation model rather than a true physiological finding, and they recommend continued efforts to improve joint center estimation and expand markerless model training using more diagnostically and kinematically varied pediatric gait data. The authors identify further stratifying by diagnosis and gait severity, conducting multi-site data collection to assess environmental factors, and evaluating additional composite gait outcome measures as clear next steps.

Citation: Chafetz, R. S., Warshauer, S., Waldron, S., Kruger, K. M., Donahue, S., Bauer, J. P., Sienko, S., Bagley, A., & Courter, R. (2026). Validation, characterization, and utility of markerless motion capture in a large cohort of pediatric patients with complex gait patterns. Journal of Biomechanics, 204, 113359. https://doi.org/10.1016/j.jbiomech.2026.113359

Link: Read the full study

How Well Does Markerless Capture Scale to a Large, Complex Pediatric Cohort?

This study tests Theia3D against marker-based motion capture in one of the largest and most diagnostically varied pediatric cohorts published to date, finding strong sagittal-plane agreement and gait classification concordance, alongside clearer gaps in transverse-plane rotation. If your work involves evaluating markerless motion capture for pediatric gait assessment and you want to understand where its current evidence base is strongest, talk to our team about what this study could mean for your population.

Recent Posts