Monocular accuracy/stability and minimal camera count for biomechanical IK

Determine the accuracy and stability of biomechanical inverse-kinematics estimates obtained from monocular videos using single-image human pose estimation algorithms that directly predict 2D/3D keypoints (e.g., MeTRAbs-ACAE), and identify the minimal number of cameras required to achieve a specified level of accuracy for joint kinematics within the presented markerless motion capture pipeline.

Background

The paper implements a multi-camera markerless motion capture pipeline that uses MeTRAbs-ACAE to obtain dense, biomechanically motivated keypoints, reconstructs smooth trajectories via an implicit function, and performs inverse kinematics with nimblephysics. Although MeTRAbs can produce 3D joint estimates from single images, the study leverages only 2D detections (plus confidence) within a multi-view setup.

In the Discussion, the authors explicitly note the need to evaluate how well biomechanical fits can be obtained from monocular video using single-image pose estimation approaches and to establish how many cameras are minimally needed for a targeted accuracy. They explicitly defer this investigation to future work, indicating that these questions remain unresolved within the present study.