The proposed algorithm is a unique combination of a GCN and a novel rotation-invariant local

We appreciate positive and constructive comments and address the main concerns raised by the reviewers below. Note that our training procedure takes the original 3D points and, consequently, is free from information loss. The manual feature extraction steps in RIConv and ClusterNet may incur the loss and lead to performance degradation. Therefore, the accuracy of A-CNN on z/SO(3) is as low as 35.8% according to our experiment based CNN (G-CNN) [A4] are designed for meshes and their target tasks are different from ours. In practice, computing PCAs at every level does not affect the overall accuracy at all.