Uncertainty-Aware Mapping from 3D Keypoints to Anatomical Landmarks for Markerless Biomechanics
Summary: arXiv:2603.26844v1 Announce Type: cross
Abstract
In recent years, the field of markerless biomechanics has gained significant attention, particularly in its reliance on 3D skeletal keypoints extracted from video data. However, many current biomechanical mappings treat these estimates as deterministic, lacking a principled mechanism for frame-wise quality control. This paper explores the potential of utilizing predictive uncertainty as a quantitative measure of confidence for mapping 3D pose keypoints to 3D anatomical landmarks, a crucial step that precedes inverse kinematics and musculoskeletal analysis.
Introduction
The advent of markerless motion capture technologies has revolutionized biomechanics, allowing for more flexible and efficient data collection. Yet, the inherent uncertainty in the extraction of skeletal keypoints poses challenges in ensuring the accuracy of subsequent analyses. This research aims to address these challenges by modeling both observation noise and model limitations within a temporal learning framework.
Methodology
Using synchronized motion capture ground truth data from the AMASS dataset, the study evaluates uncertainty at both frame and joint levels. The evaluation is conducted through several quantitative measures, including:
- Error–uncertainty rank correlation
- Risk–coverage analysis
- Catastrophic outlier detection
Results
The findings demonstrate that uncertainty estimates, especially those associated with model uncertainty, show a strong monotonic relationship with landmark error. Specifically, the Spearman correlation coefficient (ρ) is approximately 0.63, indicating a significant association. This relationship enables the selective retention of reliable frames, effectively reducing error to approximately 16.8 mm at 10% coverage. Furthermore, the method allows for accurate detection of severe failures, with a receiver operating characteristic area under the curve (ROC-AUC) of approximately 0.92 for errors greater than 50 mm.
Discussion
Reliability ranking remains stable even under controlled input degradation scenarios, such as the introduction of Gaussian noise and simulated missing joints. In contrast, the uncertainty arising from observation noise offers limited additional benefits, suggesting that the predominant failures in keypoint-to-landmark mapping are primarily influenced by model uncertainty. These insights highlight the importance of predictive uncertainty as a practical, frame-wise tool for automatic quality control in markerless biomechanical pipelines.
Conclusion
This work establishes a novel approach to incorporate predictive uncertainty into markerless biomechanics, providing a robust framework for improving the reliability of biomechanical analyses. By focusing on model uncertainty, researchers and practitioners can enhance the quality control of 3D keypoint mappings, ultimately leading to more accurate biomechanical assessments and insights.