imSMPL: A Reimplementation of imGHUM Trained on SMPL-X Pose Parameters

For my CSCE-689 finaly project, I implemented a neural signed distance field conditioned on the SMPL-X body model. This project is an adaption and implementation of imGHUM, which learns a signed distance field of the GHUM model, to SMPL-X.

In this project, I restricted to a simpler subproblem of learning the entire SMPL-X body as a signed distance field. In particular, I train a network conditioned on the body pose parameters, and represent only the arms and torso of the model.

In order to train this neural model, 100 animations from the AMASS dataset are used, sampling up to 200 poses per animaiton. 16k training data points are generated for each pose, of which half are sampled uniformly from the body surface. The remaining points are sampled uniformly in space, or in a tight normal distribution arounf the body surface. As in imGHUM, training losses serve to enforce the correct 0 level-set, correct surface normals, correct inside/outside labels, and a gradient magnitude of 1 with respect to input coordinates. Compared to imGHUM, I use a simple single MLP with no spatial input encoding, and train to convergence.