Note: These findings illustrate how differences in system architecture, camera configuration, and keypoint detection modelling may influence accuracy, particularly in high-speed, multi-planar movements such as baseball pitching. Importantly, these results should not be interpreted as demonstrating inherent superiority of one ML system over another, but rather as evidence of how specific design characteristics manifest in measured biomechanical outcomes of baseball pitching. Thus, these findings reflect the distinct operational designs of each ML system rather than a hierarchy of performance accuracy.

Why This Matters
‍

Pitching is one of the most complex and high-velocity movements in sport. Historically, only marker-based lab systems could capture it accurately. But these setups can be expensive, slow, and impractical for real-world or team environments.

This paper delivers the strongest independent evidence to date that multi-camera markerless systems can estimate pitching kinematics and key injury-relevant metrics with practical reliability, particularly when camera placement is optimized (as in a Theia3D deployment).

For labs, teams, and player-development groups, this matters because:

• You can analyze real, game-speed mechanics without markers or lab constraints
• You can scale biomechanics to larger groups, not just 1-2 athletes at a time
• You gain insight into stride mechanics, trunk/pelvis rotation, and arm action – variables that drive velocity and load
• Validation against marker-based data means data you can trust

This study provides an important evidence base for evaluating systems like Theia3D in elite baseball environments.

‍

Study Overview

• Participants: 18 NCAA D1/D2 pitchers
  
• Setting: Petco Park (MLB stadium)
  
• Task: Ten max-effort fastballs per athlete
  
• Systems tested:
  
  
  • Theia3D (10 Qualisys Miqus Video high-speed cameras, portable and lab-style configuration)
    
  • Hawk-Eye (5 in-stadium cameras, fixed positions)
    
  • Marker-based (MB) (18 Qualisys Arqus cameras, reference system)
    
• Outputs analyzed:
  
  
  • 3D joint positions
    
  • Kinematic waveforms (pelvis rotation, trunk rotation, shoulder IR/ER velocity, etc.)
    
  • Twelve discrete performance & injury-risk metrics (stride length, lead knee flexion, max trunk rotation velocity, shoulder distraction force, elbow valgus torque, etc.)
    

The authors used industry-standard metrics to quantify agreement across systems: root mean square error (RMSE), mean per-joint position error (MPJPE), concordance correlation coefficient (CCC), and Bland-Altman analysis.
‍

Key Findings
‍

1. Theia3D showed lower joint-position errors across most variables
‍

Mean per-joint position error (MPJPE)

• Theia3D: 52.0 ± 12.3 mm
  
• Hawk-Eye: 56.6 ± 9.4 mm

Errors were smallest at the hip (~40–43 mm) and largest at the elbow (~70–75 mm), expected for high-speed upper-extremity motion.
‍

2. Theia3D produced smaller RMS errors in several high-value pitching variables
‍

Theia3D out-performed or matched the in-stadium system in:

• Knee flexion
  
• Pelvis rotation
  
• Trunk rotation
  
• Pelvis rotational velocity
  
• Shoulder rotational velocity

This consistency across both lower-body and trunk-based metrics reinforces Theia’s suitability for velocity-generation and timing-chain analyses.
‍

3. Stride length showed strong agreement with marker-based (CCC > 0.85)
‍

Stride length is one of the most stable and high-value mechanical indicators of consistency and timing.

Both systems slightly underestimated stride length (–0.04 to –0.05 m), but demonstrated excellent consistency.
‍

4. Upper-extremity velocities and torques showed expected variability
‍

• Rapid internal rotation of the shoulder amplifies small joint-position errors
  
• Shoulder IR velocity and arm-torque measures had the widest LOA and lowest CCC values

This aligns with broader literature showing the difficulty of modeling rapid, multi-planar shoulder motion, even in marker-based systems.
‍

5. Theia3D demonstrated consistent agreement across a broader range of metrics
‍

While the in-stadium system performed well in certain discrete measures, Theia3D showed:

• Narrower limits of agreement in elbow flexion
  
• Higher concordance in several trunk and pelvis metrics
  
• Overall smaller average errors in many kinematic and kinetic signals
  

This supports Theia3D’s versatility in both lab and field-based baseball biomechanics.
‍

What This Means for Baseball Performance, Biomechanics, and Return-to-Throwing
‍

For MLB/NCAA Performance Staff
‍

You can use Theia3D to analyze:

• Pelvis-trunk sequencing
• Stride mechanics
• Timing chain events between SFC → BR → FT
• Day-to-day consistency and workload management

These variables showed the strongest agreement and are foundational to velocity and repeatability.
‍

For Sports Scientists & Biomechanics Labs
‍

The study reinforces Theia3D’s:

• Accuracy across lower-extremity and trunk variables
• Validity of rotational velocities tied to timing and kinematic efficiency
• Utility in high-speed dynamic tasks beyond gait and jumping

It expands the evidence base supporting markerless biomechanics outside of tightly controlled lab setups.
‍

For Clinicians Focused on Injury Risk
‍

While shoulder kinetics demonstrated expected variability, the study provides:

• The first published evaluation of kinetic outputs from any ML system during pitching
• Practical insight into how markerless estimation compares to MB baselines
• A foundation for continued use in return-to-throw protocols where trends matter as much as absolutes

Access the full study here.
‍

Interested in Markerless Biomechanics for Baseball?
‍

Theia3D is the only markerless system validated across running, jumping, cutting, gait, and now high-speed baseball pitching in both lab and stadium environments.

Contact us today to book a demo.

‍

‍

‍