Input transform and point cloud with features on the points

[adosar]

Hello and congrats for this excellent work!

I want to use PointNet but I don't know whether I should remove the first T-Net from the architecture. In my task, only rigid transformations are meaningful but this is not guaranteed by the T-Net, since it predicts a 3x3 affine transformation. Moreover, my input includes additional feature channels, i.e. the input has shape (N, 3+F).

Does it make sense to remove the first T-Net and just increase the number of input channels in the next layer to take account the additional features?