REP3D: 3D Human Motion Capture Dataset for Athletic Movement

The field of human 3D pose estimation suffers from a small population of diverse public motion capture datasets, each with a low number of environments and subjects. We propose a new dataset including 45 participants and 22 environments, using motion capture technology that allows data collection in arbitrary locations. The dataset is composed of video and motion capture data for athletic actions selected from golf and baseball, recorded from a plurality of angles and distances. The annotation process for semi-automatically aligning video data with ground truth 3D joint locations is fully outlined. The performance of a modern human 3D pose estimation model on a subset of the dataset is reported.