PointMSCA: A point cloud analysis model based on attention feature fusion mechanism

Analyzing point clouds presents challenges because of their irregular and unordered data structure. To effectively capture the 3D geometries, previous studies have primarily focused on developing local geometric extractors through methods such as convolution and attention mechanisms. Enhancing the performance of local geometry extractors can be achieved by augmenting the number of convolution layers and constructing deeper networks. However, this approach comes with the drawback of substantially heightened computational costs and increased memory usage, consequently resulting in longer reasoning times. This paper proposes a point cloud network PointMSCA based on channel attention feature fusion mechanism, which can effectively reduce the number of parameters and improve the accuracy. The experimental results show that the class-level mean IoU is 84.9% and the instancelevel mean IoU is 86% on ShapeNetPart. This approach reaches a performance level similar to existing techniques while utilizing fewer parameters.