When understanding pointnet, there is a layer of shared mlp, but I don’t know what it is, so I checked the information on the Internet, and I can find a lot of good information by using share mlp as a keyword. The following is just a summary.
MLP reference article , this one picture should be enough.
share mlp (picture on the left)
I think the original blogger's picture has no contrast effect
Comparing this with the MLP diagram, we can find that the share MLP is essentially an MLP. It is just a name in the point cloud. It is emphasized that all points use the same parameters.
The one-dimensional convolution reference conv1d is the principle of realizing share MLP, and the convolution calculation is used.