Invisible or hard self occluded 2D points support

[kallivad]

Hi @lxy5513! Thank you for sharing with this interesting solution!! Having implementing some type of such work now I wonder how are you dealing with points not predicted from 2D points backbone? I mean some of drawbacks of openpose method especially, where for example we can get omitted prediction from head eyes or nose points when person is back to camera. Cause VideoPose3D could handle 3d reconstruction only having sequence of full 2D points set from each frame in sequence.