Photo from Scientific Research

We fit 3D SMPL bodies to the SLP dataset [3], which we use for initializing the physics simulator and for training and testing our deep models. Our method resolves depth ambiguity using a loss between the SMPL mesh and 3D points from the depth image. Examples are shown without the depth loss term, resulting in poses with depth error. Examples are also shown without BetaNet, resulting in bodies with unreasonable shapes.

DOI