Reconstructing a clothed human from monocular video as separate, watertight clothing and body geometries that stay physically plausible (no interpenetration) across motion. Covers hybrid explicit-implicit human representations that combine a signed distance field with parametric body priors, region aggregation to split visible surfaces into clothing and body while completing invisible body regions from the parametric template, deformation via skeletal skinning plus learned non-rigid displacement fields, and occlusion-aware differentiable rasterization losses that keep the body inside the clothing. The payoff is animatable, separately editable clothing and body assets extracted from a single ordinary video.
D3-Human reconstructs disentangled 3D clothing and body geometry of a moving person from a single monocular video, so the results can be reused for animation and clothing transfer. It represents the …