Learning SKEL-specific pose-dependent deformations from BioAMASS

Develop SKEL-specific pose-dependent deformations using the BioAMASS dataset, rather than inheriting pose-correctives from SMPL, to improve the accuracy of SKEL’s skin surface deformation under pose.

Background

In the current SKEL implementation, the pose-dependent deformations are inherited from SMPL by mapping each SKEL degree of freedom to a corresponding SMPL parameter.

The authors note that this transfer is not optimal and can create artifacts in extreme poses; although SKEL can match SMPL meshes with acceptable error, more accurate pose-correctives tailored to SKEL are needed.

The paper explicitly defers the learning of SKEL-specific pose-dependent deformations to future work, suggesting BioAMASS as the training resource.