About init_tet function of DMTet

[ug-kim]

Can you explain init_tet() for initializing DMTet?

As far as I understand, with SDF, the part where the SDF value is 0 becomes the surface. I wonder if self.sdf.data += (sigma - density.thresh).clamp(-1, 1) means to set the density corresponding to the density_thresh value as surface. (Since self.sdf is initialized to zero at the beginning, and sigma predicts only positive, I understood that it means to set values corresponding to density_thresh as the surface in sigma.)

And I wonder why the density_thresh is multiplied by 25 if the density_activation is softplus. Can you tell me what 25 numbers mean?

@torch.no_grad()
def init_tet(self):

    if self.cuda_ray:
        density_thresh = min(self.mean_density, self.density_thresh)
    else:
        density_thresh = self.density_thresh

    if self.opt.density_activation == 'softplus':
        density_thresh = density_thresh * 25

    # init scale
    sigma = self.density(self.verts)['sigma'] # verts covers [-1, 1] now
    mask = sigma > density_thresh
    valid_verts = self.verts[mask]
    self.tet_scale = valid_verts.abs().amax(dim=0) + 1e-1
    self.verts = self.verts * self.tet_scale

    # init sigma
    sigma = self.density(self.verts)['sigma'] # new verts
    self.sdf.data += (sigma - density_thresh).clamp(-1, 1)

    print(f'[INFO] init dmtet: scale = {self.tet_scale}')

[ashawkey]

@ug-kim Hi, you are right. 25 is simply an empirical scale for softplus activation, to make the surface more accurate.