Joint Path Alignment Framework for 3D Human Pose and Shape Estimation From Video

3D human pose and shape estimation (3D-HPSE) from video aims to generate sequence of mesh that depict body in the video. Current deep learning based 3D-HPSE networks takes input have focused on improving temporal consistency among joints by supervising acceleration error between predicted ground-truth motion. However, these methods overlooked geometric misalignments persistent discrepancy paths drawn joints. To this end, we propose Joint Path Alignment (JPA) framework, a model-agnostic approach mitigates introducing Temporal Procrustes Regularization (TPAR) loss performs group-wise joint movement paths. Unlike previous rely solely per-frame supervision for accuracy, our framework adds sequence-level accuracy with TPAR performing analysis sequences Our experiments show JPA advances network exceed state-of-the-art performances benchmark datasets both smoothness metric.