imGHUM

imGHUM is a generative model of 3D human shape and articulated pose, represented as a signed distance function. The full body is modeled implicitly as a function zero-level-set and without the use of an explicit template mesh. We compute the signed distance s = S(ρ, α) and the semantics c = C(ρ, α) of a spatial point ρ to the surface of an articulated human shape defined by the generative latent code α. Using an explicit skeleton, we transform the point ρ into the normalized coordinate frames as {ρ̃ʲ} for N = 4 sub-part networks, modeling body, hands, and head. Each sub-model {Sʲ} represents a semantic signed-distance function. The sub-models are finally combined consistently using an MLP U to compute the outputs s and c for the full body. The multi-part pipeline builds a full body model as well as sub-part models for head and hands, jointly, in a consistent training loop.

On the right of the Figure, we visualize the zero-level-set body surface extracted with marching cubes and the implicit correspondences to a canonical instance given by the output semantics. The semantics allows e.g. for surface coloring or texturing.