Point cloud recognition based on lightweight embeddable attention module

There are challenges in point cloud recognition tasks such as modeling the relationship between points, rotation invariance, disorder, and so on. Many of the previous methods focus on the local structure modeling of point cloud and ignore the global features. Meanwhile, a large number of points need to be input at one time to obtain sufficient information that causes the computational burden. In this paper, we propose a Dual Branch Attention Network (DBAN) considering both global and local information, where a learnable way is used to guide feature aggregation. The main component of DBAN is that we design an effective channel attention module called Lightweight Embeddable Attention Module (LEAM), which not only can be embedded in the existing backbone of the point cloud recognition networks easily but also overcome the contradiction between performance and complexity of general point cloud models. Extensive experiments on challenging benchmark datasets verify that our method performs better with fewer input points in point cloud recognition tasks such as object classification and segmentation. (c) 2021 Published by Elsevier B.V.