MBMRNet: Point Cloud Semantic Segmentation Based on Multi-branch Attention and Multi-resolution Feature Fusion

In the existing point cloud semantic segmentation methods, the study of local feature learning plays a dominant role. However, due to the absence of context information, the accuracy of semantic segmentation is seriously affected. In this study, we propose a novel network MBMRNet based on attention mechanisms. Firstly, we introduce the gradient information of point cloud and use relative position encoding to obtain spatial information. Then, we adjust the attention weight of different features by learning the point neighborhood and the relationship between different channels through a multi-branch attention (MBA) module. Finally, considering the differences between different resolution features at different stages, we design a multi-resolution feature fusion (MRFF) module to make full use of low resolution and high resolution features for dynamic point-level feature fusion. The experimental results on several public benchmarks demonstrate that MBMRNet outperforms existing methods, achieving 73.2% and 83.2% on the S3DIS Area 5 and NPM3D datasets, respectively.