https://istvansarandi.com/dozens/
Describes how many different pose datasets with their different skeletons can be used to train one single model. Essentially comes up with a common skeleton, using autoencoders.
The latest of 3D human pose estimation: https://github.com/dluvizon/scene-aware-3d-multi-human
Factor graphs review, for robotics