Video analysis spots early motor problems in infants

Early detection of motor impairment in infants currently relies on expert observation. This study benchmarked three pose-estimation frameworks (MeTRAbs-ACAE, SAM 3D Body, Sapiens) on multi-view video of 8 infants to assess their ability to extract whole-body kinematics without markers. SAM 3D Body delivered the best 3D reconstruction accuracy (19–28 mm Procrustes error) and enabled biomechanical models to distinguish representative movement patterns linked to motor development. While Sapiens excelled at keypoint detection (22.8 pixel reprojection error), it provided less complete 3D information. The work establishes proof-of-concept for video-based, scalable screening but acknowledges current limitations in pose estimation accuracy for infant biomechanics.