Check out section 4.2 of the paper: 3D is very similar to 2D, using the same strategy of big regions -> small regions -> voxels.

(There's also a bit of extra logic to skip regions which are occluded in Z, plus a final pass to render normals using automatic differentiation)