Hard self occlusion points

Hi @fmahoudeau! Thank you for sharing ws such interesting solution. I wonder how dow you estimate its stability in near profile poses? Does it supports hard self occlusion situations without omitting predict of some keypoints like openpose did?
Another issue of ShelfNet backbone - how do you estimate its ability for DensePose task in scope of low inference time?