MFB: A 3D point cloud segmentation model based on Mamba

Mamba-based approaches have performed extremely well in point cloud segmentation tasks, but still have limitations regarding computational efficiency and global context understanding. To solve this problem, we propose MFB (Mamba Fusion Block), which composes of two parts, one is similar to the previous Mamba block, and the other is our integrated linear attention and Mamba-like Linear Attention (MLLA). Among them, Linear Attention implements a mechanism based on rotational position encoding for processing point cloud features, which can better obtain the feature details from point cloud data and improve the model’s performance for point cloud segmentation. MLLA, while maintaining the original structure of Mamba Block, further enhances the model’s capacity to learn features at multiple scales. The experimental outcomes indicate that Our proposed approach outshines the existing Mamba-based methods on the ShapeNetPart dataset.