Gram-Schmidt on regression CNN

[matfcp]

Hey Dmitry, how are you? Hope you're well!

I read your article on Towards Data Science about 6D representation of rotation matrices for better continuity on regression tasks.

I'm using a PointNet-based architecture for regressing rotation angles. My targets are the rotation matrices, my inputs are pointclouds, and I must ask: how can I equip my network with Gram-Schmidt orthogonalization so I can have a 6D representation from the first two columns of the rotation matrix? Should I use it as a regularizer?

Thanks in advance!