Random rotation augmentation neglects normals

[johnpeterflynn]

On the following line you apply random rotation to help teach your network to be invariant to absolute position. This is important because even though you formulated your input EdgeConv operator to be invariant to absolute position, it still contains absolute orientation. However you neglect to apply this random rotation to your normals as well. This not only means you're learning on absolute orientation but also that your normals and positions no longer match. I suspect fixing this issue could lead to noticeable improvements in your segmentation IoU scores.

https://github.com/VisualComputingInstitute/dcm-net/blob/e021c69d7a27fc985450b6afadef0a3327161a1c/transform/random_rotation.py#L18