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 v2025.1.2, as stated in the preprint.
Study Overview
Design
A perspective article documenting hardware, synchronization, calibration, and data-management issues encountered across three Theia3D data-collection campaigns, including a direct comparison of Theia3D's recommended dynamic chessboard calibration against three object-based calibration alternatives.
Deployment Scenarios
Three data-collection campaigns informed the article's recommendations: an indoor study with 60 participants performing 70 tasks under ambient lighting; an outdoor rugby tackling study with 15 participants completing 10 tackling trials in both daylight and night-time conditions (temperatures as low as approximately -2°C); and a second indoor study with 20 participants performing 24 tasks. All three used an eight-camera Sony RX0 II setup.
Calibration Comparison
Three object-based calibration approaches, static chessboard calibration, fixed-object calibration using a laboratory force-plate platform, and active wand calibration using a standard infrared calibration wand, were each compared against Theia3D's recommended dynamic chessboard calibration as a reference, using camera-centre position and relative rotation errors calculated from five independent data-collection sessions with an eight-camera setup.
Analysis
Camera-centre positions and relative rotation errors were calculated from the translation vectors and rotation matrices in the Theia3D calibration files, and mean errors for each object-based calibration method were compared against the dynamic chessboard reference across the five sessions.
Key Findings
- Static chessboard and fixed-object calibration closely matched Theia3D's recommended dynamic chessboard method. Mean 3D camera-centre errors were 36.76 mm and 38.17 mm respectively, with relative rotation errors of 0.37° and 0.38°, suggesting both are practical backup options when dynamic calibration is unavailable or fails.
- Active wand calibration produced substantially larger errors than the other two alternatives. Mean translation error reached 86.40 mm and rotation error reached 1.39°, roughly 2.3 and 3.7 times larger than the static chessboard and fixed-object methods, an issue the authors attribute to the wand's reference points spanning only a small fraction of the capture volume.
- Camera synchronization failures were a recurring, session-threatening issue. In approximately 10% of experimental sessions, dynamic chessboard calibration failed outright with a timecode-related error that rendered all task recordings from that session unusable, since Theia3D provides no real-time feedback to catch the problem during acquisition.
- Cable tension at elevated camera mounts was a common cause of undetected camera drop-out. The weight of Ethernet cabling running to cameras mounted 1 to 2.5 meters high could progressively loosen power-over-Ethernet connections, silently disconnecting a camera from the network until the issue surfaced during post-processing.
- Data storage requirements grew far beyond the raw video footage alone. A single campaign generated roughly 3 terabytes of raw video, but renaming, camera-order verification, and Theia3D's built-in video-organization step each required a further full copy of the dataset, pushing total storage needs to roughly four times the original raw video size.
What This Means for Labs Running or Scaling Up Theia3D Systems
For labs setting up a new Theia3D system or scaling up an existing one, this article's central message is that reliable markerless data collection is as much an operational discipline as a software one: treating calibration and synchronization checks as active, in-session quality-control steps, rather than assumptions to verify later, can prevent the kind of invisible data loss that only becomes apparent during post-processing. The calibration comparison also gives labs a concrete, low-cost contingency plan, a static chessboard or a known fixed object such as a force-plate platform, for the specific situation where dynamic chessboard calibration isn't available or has failed.
The Role of Theia3D in Movement Analysis
Theia3D was the markerless system used throughout all three data-collection campaigns described in this article, and its calibration files provided the camera-position and rotation data used to compare the dynamic chessboard method against the three object-based alternatives. The authors frame their recommendations as practical safeguards for getting the most reliable data out of a Theia3D deployment, rather than as a critique of the software itself.
Study Limitations and Future Directions
The authors note that their observations reflect a specific hardware configuration, an eight-camera Sony RX0 II system mounted on tripods at an average height of approximately 1.6 meters in a roughly circular arrangement, and may not generalize directly to systems using a different number of cameras, alternative placements, or other capture geometries. All data-collection sessions involved single-person capture, so the article does not address how multi-person recording and the inter-subject occlusion it introduces might affect tracking performance. The comparison between calibration methods was limited to camera-position and rotation error; the authors identify quantifying how each calibration approach propagates into downstream 3D keypoint reconstruction, segment pose, and joint kinematics during real participant movement as a clear next step.
Citation: Gholami, B., Edgington, H. C. A., Hollinger, D. S., Zhai, X., & Liew, B. X. W. (2026). Deploying Theia3D in biomechanics research: Three critical considerations for accurate and efficient workflows. SSRN. https://papers.ssrn.com/abstract=7194327
Link: Read the full preprint
What Does It Actually Take to Run a Reliable Markerless System?
This article is a reminder that getting accurate markerless data isn't only about the underlying algorithms, it's also about cable management, synchronization checks, calibration contingencies, and storage planning done right the first time. If your lab is setting up or scaling a Theia3D deployment and wants to avoid the kind of preventable data loss this article describes, talk to our team about building a reliable end-to-end workflow.