3D Human Reconstruction

3D human reconstruction aims to recover a person's body shape, pose, clothing, appearance, and motion from images, video, or other sensors, often alongside cameras and surrounding scenes. Methods range from fitting parametric models such as SMPL-X using keypoints and silhouettes to transformer-based mesh predictors and diffusion models that generate missing views or poses; neural fields and 3D Gaussian splatting support detailed, animatable appearance. Multi-view geometry, temporal motion priors, and contact or physics constraints help resolve ambiguity, with benchmarks including 3DPW and EMDB for motion and CAPE and THuman for clothed geometry and avatars. Feed-forward models are making reconstruction faster, but monocular scale ambiguity, occlusion, hands, loose garments, and consistency between people, scenes, and motion remain difficult, especially beyond controlled capture settings.