Learning When to Correct: Selective Calibration for Monocular Kinematics

Abstract

Post hoc calibration of markerless kinematics enables accessible movement analysis from monocular video, but conventional pipelines apply corrections unconditionally, regardless of initial estimate fidelity. Across 82 subjects, four capture domains, four pose estimators, and nine correction paradigms (3.48M evaluations), we show that Extremely Randomized Trees (ExtraTrees) outperform deep tabular, generative, and temporal alternatives in predicting kinematic residuals from 21 dimensional biomechanical biomarkers, reducing macro MAE from 12.54° to 6.22° (50.4%). Despite this improvement, unconditional ExtraTrees exhibits a critical overcorrection dilemma, pushing 34.7% of frames initially within a 5° tolerance outside the tolerance and producing a 76.8% preservation failure rate at 2°. Factorial transfer analysis further shows that changing the upstream pose estimator induces a domain shift comparable to changing the physical capture environment (-1.76° versus -1.32°). To address this problem, we extract an eight dimensional forest consensus representation from tree variance, prediction quantiles, and leaf topology to decide whether to preserve the incumbent estimate or adopt the ExtraTrees candidate. Trained via nested out of fold benefit regression, our reference free selective router reduces macro MAE to 5.94°, lowers preservation failure from 34.7% to 16.5%, and outperforms unconditional ExtraTrees in 15 of 16 domain and estimator combinations.