Equivariance in PointMLP
smiles724 opened this issue · comments
Hi thanks for your sharing.
However, I am curious about your initialization of point embeddings. You used a Conv1D to initialize the point feature, where the input channel is 3, the coordinates of each point.
I wonder whether this kind of initialization can lead to different results when the coordinate system changes. In other words, how do you ganrantee the equivariance of your network?
Please see more direct figure in your paper. You directly change (1024, 3) to (1024, 64).
