Facebook DensePose

I was reading Facebook DensePose paper and i did not understand the method of choosing sampled points for each segmented part of a person in 2d image and getting correspondense on SMPL 3D model, i would like to know how did they decided for sampled points and did they give those points any identity for localization.

here is the link https://arxiv.org/pdf/1802.00434.pdf