1: \begin{abstract}

2: In point cloud generation and completion, 

3: previous methods for transforming latent features to point clouds are generally based on fully connected layers (FC-based) or folding operations (Folding-based). 

4: However, 

5: point clouds generated by FC-based methods are usually troubled by outliers and rough surfaces. 

6: For folding-based methods, their data flow is large, convergence speed is slow, and they are also hard to handle the generation of non-smooth surfaces. 

7: In this work, 

8: we propose AXform, an attention-based method to transform latent features to point clouds. 

9: AXform first generates points in an interim space, using a fully connected layer. 

10: These interim points are then aggregated to generate the target point cloud. 

11: AXform takes both parameter sharing and data flow into account, 

12: which makes it has fewer outliers, fewer network parameters, and a faster convergence speed. 

13: The points generated by AXform do not have the strong 2-manifold constraint, which improves the generation of non-smooth surfaces. 

14: When AXform is expanded to multiple branches for local generations, 

15: the centripetal constraint makes it has properties of self-clustering and space consistency, 

16: which further enables unsupervised semantic segmentation. 

17: We also adopt this scheme and design AXformNet for point cloud completion. 

18: Considerable experiments on different datasets show that our methods achieve state-of-the-art results.

19: \end{abstract}

20: