PPFNet: Global Context Aware Local Features for Robust 3D Point Matching 阅读

It's based on PointNet but the important thing is the N-tuple loss, which is a many-to-many loss. The basic idea is still the reproject error sum. So the authors call them point pair features (PPF). This gives a fast representation of objects for 3D reconstruction. Minors thing such as CGF, 3D Match are standard in the community.