Attention

We design an attention block that projects the attention vector into a subspace represented by the combination of learnable prototypes. The spatial supports of prototypes are designed to have different dilations and shapes (clustered using the training data), considering the irregularity of 3D data.