It is clear from your paper that the input to the first MLP for feature extraction is a 1D vector containing the features from the vertex's neighbors. However, I do not understand how you build a 1D vector from a set of 1 or more neighbors?
In your code, the PointSet Pooling class uses as an input a [N, M] tensor with the point features. This wouldn't be consistent with the paper. Can you please explain how this tensor 2D tensor is transformed into a 1D tensor that you use as an input for the feature extraction MLP?
Additionally, Do you use the same MLP for all sets of neighbors? Or do you use a different MLP for each set?
Dear authors,
It is clear from your paper that the input to the first MLP for feature extraction is a 1D vector containing the features from the vertex's neighbors. However, I do not understand how you build a 1D vector from a set of 1 or more neighbors?
In your code, the PointSet Pooling class uses as an input a [N, M] tensor with the point features. This wouldn't be consistent with the paper. Can you please explain how this tensor 2D tensor is transformed into a 1D tensor that you use as an input for the feature extraction MLP?
Additionally, Do you use the same MLP for all sets of neighbors? Or do you use a different MLP for each set?