Study at a glance

In 23 healthy adults performing walking and eight lower-extremity exercises, Theia3D, using ten cameras, achieved the lowest kinematic error of thirteen single-view and two multi-view markerless tracking methods benchmarked against marker-based motion capture, with multi-view systems outperforming single-view ones by roughly 2 to 3 degrees overall.

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 not specified in the paper.

Study Overview

Design

A benchmarking study comparing thirteen single-view and two multi-view markerless motion tracking methods, including Theia3D, against simultaneously recorded marker-based motion capture for lower-extremity kinematics during walking and lower-extremity functional exercises.

Participants

Twenty-three healthy adults (10 women, 13 men; mean age 36 years) with no gait impairments.

Data Collection

Motion data were collected simultaneously using a marker-based reference system (20 infrared cameras with 45 reflective markers) and video for markerless analysis (10 synchronized RGB cameras). Participants completed at least one minute of overground walking and three repetitions each of eight functional exercises: lateral step-down, step-up and step-down, drop jump, countermovement jump, squat, step and hold, single-leg squat, and sit-to-stand.

Analysis

Marker-based ground-truth kinematics were calculated in OpenSim using the Rajagopal musculoskeletal model. Thirteen single-view computer-vision methods and two multi-view methods, OpenCap (two cameras) and Theia3D (ten cameras), were each compared against this reference using root-mean-square differences (RMSD) across joint angles and activities. Linear mixed-effects models assessed the effects of motion-capture method, joint, and activity on tracking accuracy, with biological variables (age, BMI, weight, height, sex) tested as covariates.

Key Findings

  • Theia3D achieved the lowest overall tracking error of any method benchmarked. Median RMSD was 3.9 degrees for Theia3D, compared with 5.4 degrees for OpenCap, and 7.1 to 7.2 degrees for the best-performing single-view methods.
  • Multi-view methods significantly outperformed single-view ones across joints and activities. Theia3D showed lower RMSDs than OpenCap by a median of 1.3 degrees and lower than the top single-view method by 3.0 degrees overall (p < 0.0001), with Theia3D's advantage over the top single-view method ranging from 1.3 to 7.5 degrees depending on the specific activity.
  • Adding a biomechanical model to a vision-only method's output did not systematically improve kinematic accuracy. Testing this directly on the best-performing single-view method, median RMSD differences before and after adding the biomechanical model were close to zero (0.0 degrees [-0.5, 0.4]), with the direction of improvement varying by joint and activity rather than consistently favoring the biomechanically constrained output.
  • No biological variable significantly predicted tracking accuracy for any method. After correcting for multiple comparisons, age, BMI, weight, height, and sex showed no significant association with RMSD across the markerless methods tested.
  • Kinematic accuracy correlated strongly with a standard computer-vision benchmarking metric, the Procrustes-aligned mean per-joint position error (PA-MPJPE), in both an established external benchmark dataset (r = 0.85) and the study's own dataset (r = 0.96), suggesting that computer-vision accuracy metrics may serve as a reasonably useful proxy for biomechanical kinematic accuracy during early-stage method development.

What This Means for Movement Science Researchers and Lab Methodologists

For researchers and labs weighing markerless tracking options, this independent, multi-method benchmark offers useful context: of every system tested, including thirteen current single-camera computer-vision methods, Theia3D's ten-camera setup produced the closest agreement with marker-based ground truth, and multi-view systems in general showed a consistent, meaningful accuracy advantage over single-view approaches. At the same time, the authors are candid that this advantage, while statistically significant, came in at roughly 2 to 3 degrees, a gap researchers and labs should weigh against the added cost, calibration time, and data-storage demands of a multi-camera system for applications where that level of precision may not be essential. The finding that adding a biomechanical model after pose estimation didn't reliably improve accuracy is also a useful methodological caution for anyone assuming that step is a free accuracy gain.

The Role of Theia3D in Movement Analysis

Theia3D was one of two multi-view markerless systems directly benchmarked in this study, evaluated using a ten-camera configuration alongside twelve other single-view computer-vision methods and one other multi-view method, with lower-extremity kinematics from all methods compared against a common marker-based reference across walking and eight functional exercises. Of every method tested in this independent academic benchmark, Theia3D produced the lowest overall tracking error relative to the marker-based ground truth.

Study Limitations and Future Directions

The authors note that all methods were benchmarked only in participants without gait impairments, and that root-mean-square differences may vary somewhat in populations with gait impairments, though they have no specific reason to expect the broader conclusions about biomechanical modeling and multi-view fusion would not generalize. Testing was conducted in a controlled laboratory setting without researchers or other personnel in the camera field of view, so these findings are most directly relevant to single-person data capture rather than more complex real-world or in-the-field scenarios with cluttered backgrounds or multiple people in view. The hypothesis that biomechanical modeling does not improve accuracy was tested using a parametric mesh-based computer-vision method and may not generalize to tools like OpenCap and Theia3D, which do not fit meshes to video data in the same way; prior work cited by the authors suggests mesh-based approaches can be slightly less accurate at joint-center estimation despite offering more information. The authors recommend continued benchmarking as new computer-vision methods emerge and have made their dataset and code publicly available to support that ongoing comparison.

Citation: Li, Z., Shin, S., Phan, V., Meinders, E., & Halilaj, E. (2025). Impact of multi-view fusion and biomechanical modeling on markerless motion tracking. IEEE Transactions on Biomedical Engineering. https://doi.org/10.1109/TBME.2025.3622032

Link: Read the full study

How Does Theia3D Stack Up Against the Latest Computer-Vision Methods?

This independent benchmark found that Theia3D's multi-camera approach produced the lowest kinematic error of fifteen markerless tracking methods tested, including more than a dozen newer single-camera computer-vision approaches, though the authors note that accuracy gains from multi-camera systems should be weighed against their added complexity for any given application. If your work involves choosing between markerless tracking approaches and you want to understand these tradeoffs for your own research, talk to our team about what this benchmark could mean for your setup.

Recent Posts