Markerless (ML) motion capture has emerged as a viable option to marker-based (MB) motion capture in estimating movement biomechanics, but limited data exists on the accuracy of ML systems during high-speed throwing. This study evaluated the accuracy and reliability of an in-stadium (Hawk-Eye) and a portable (Theia3D) ML motion-capture system in quantifying baseball pitching kinematics and kinetics relative to an MB reference. Eighteen collegiate pitchers were simultaneously recorded using all three systems. Mean per-joint position error (MPJPE), statistical parametric mapping (SPM), root mean square error (RMSE), Bland-Altman analysis, and concordance correlation coefficients (CCC) were used to assess agreement. Both ML systems demonstrated measurable discrepancies across variables, with MPJPE values of 56.6 ± 9.4 mm (Hawk-Eye) and 52.0 ± 12.3 mm (Theia3D). Stride length exhibited the strongest agreement with MB in both systems (CCC > 0.85), whereas shoulder rotational variables showed greater variability. Error magnitudes in joint positions and kinematic waveforms were comparable to those reported for other ML systems during dynamic movements. These results highlight the influence of system configuration, camera deployment, and pose-estimation models on biomechanical accuracy. Overall, both configurations showed potential for estimating pitching biomechanics, underscoring the trade-offs between criterion and ecological validity in markerless motion capture.
Keywords: Bland-Altman analysis; Hawk-Eye; Theia3D; criterion validity; statistical parametric mapping.