No dataset too small! Animating 3D motion data to enlarge 2D video databases

This study outlines a technique to leverage the wide availability of high resolution three- dimensional (3D) motion capture data for the purpose of synthesising two-dimensional (2D) video camera views, thereby increasing the availability of 2D video image databases for training machine learning models requiring large datasets. We register 3D marker trajectories to generic 3D body-shapes (hulls) and use a 2D pose estimation algorithm to predict anatomical landmark keypoints in the synthesised 2D video views – a novel approach that addresses the limited data available in elite sport settings. We use 3D long jump data as an exemplar use case and investigate the influence of; 1) varying anthropmetrics, and 2) the 2D camera view, on keypoint estimation accuracy. The results indicated that 2D keypoint determination accuracy is affected by body-shape. Frontal plane camera views result in lower accuracy than sagittal plane camera views.