Cross self-attention network for 3D point cloud

It is a challenge to design a deep neural network for raw point cloud, which is disordered and unstructured data. In this paper, we introduce a cross self-attention network (CSANet) to solve raw point cloud classification and segmentation tasks. It has permutation invariance and can learn the coordinates and features of point cloud at the same time. To better capture features of different scales, a multi-scale fusion (MF) module is proposed, which can adaptively consider the information of different scales and establish a fast descent branch to bring richer gradient information. Extensive experiments on ModelNet40, ShapeNetPart, and S3DIS demonstrate that the proposed method can achieve competitive results. (C)& nbsp;2022 Elsevier B.V. All rights reserved.

Automated summary

This paper introduces a cross self-attention network (CSANet) for raw point cloud classification and segmentation tasks. It has permutation invariance and can learn the coordinates and features of point cloud simultaneously. To capture features at different scales, a multi-scale fusion (MF) module is proposed, which adaptively considers information from different scales and brings richer gradient information, achieving competitive results.