Uncertainty-Aware Mapping from 3D Keypoints to Anatomical Landmarks for Markerless Biomechanics

A research team from the University of Cassino and Southern Lazio has published a novel AI framework that introduces predictive uncertainty estimation to the critical step of mapping 3D skeletal keypoints to precise anatomical landmarks. This process is foundational for modern markerless biomechanics, which uses standard video instead of intrusive physical markers. Current pipelines treat AI-generated keypoint estimates as perfectly accurate, lacking any mechanism for quality control. The new model, detailed in the arXiv preprint, quantifies two types of uncertainty: 'observation noise' from the input data and 'model uncertainty' stemming from the AI's own limitations.

Using the AMASS motion capture dataset for validation, the team found that the model's uncertainty scores, particularly those representing model uncertainty, are highly predictive of actual landmark positioning errors. The scores show a strong monotonic correlation with error (Spearman ρ ≈ 0.63). This allows the system to perform automatic frame-wise quality control. For instance, by selecting only the top 10% most confident frames, the average landmark error can be reduced to approximately 16.8 mm. Furthermore, the framework can detect catastrophic failures—errors greater than 50 mm—with a high degree of accuracy, achieving a Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of about 0.92.

The research indicates that in this mapping task, failures are primarily driven by the AI model's own limitations rather than simple input noise. The system's reliability rankings remain stable even when inputs are artificially degraded with Gaussian noise or simulated missing joints. This work, submitted to Pattern Recognition Letters, establishes predictive uncertainty not just as a theoretical metric, but as a practical, automated tool for building more robust and trustworthy markerless biomechanical analysis pipelines, paving the way for more reliable applications in sports science, rehabilitation, and movement analysis.