Why is there no T-net in PointNet++ Architecture?

[rastna12]

The paper mentions that after the sampling and grouping layers, the results are passed to PointNet layers. The PointNet Encoder includes a T-Net, but I do not see this in the model architecture source code for PointNet++. Why?

[Na1an]

In my experience related to the issue "rotation invariant", doing data augmentation "ratation input data" or add T-Net is not efficient. Which means, I guess, the efforts for resolving "rotation invariant" by T-Net is failed.

[surpoloyang]

the stn3d and stnkd are T-nets