PointSSM: State space model for large-scale LiDAR point cloud semantic segmentation

LiDAR point cloud semantic segmentation is the foundation of numerous practical applications. Recently, the Mamba, as a promising alternative to Transformer, has been getting intense attention in this field. However, the most of existing Mamba-based methods have to crop the input point clouds into patches, which limits its global modeling ability and hinders its further application in large-scale LiDAR point cloud processing. To this end, we thoroughly investigate the difficulties of Mamba in large-scale LiDAR point cloud learning and resolve this bottleneck by combining Mamba with convolution. Specifically, we introduce convolution as an information propagator to address the long-range collapse issue, which effectively enhances the global modeling ability of Mamba and enables it to handle the large-scale point clouds without patches. Besides, we redesign the bidirectional Mamba and serialization strategy to expand the receptive field of Mamba for point cloud semantic segmentation task. Furthermore, we further investigate the selectivity of Mamba, and exploit Mamba in the down-sampling stage for feature aggregation. To evaluate the effectiveness of our method, extensive experiments are conducted on two indoor and two outdoor public point cloud datasets. The results demonstrate the superiority of our method compared with state-of-the-art networks. • A novel Mamba-based method is proposed for large-scale point cloud segmentation. • MambaConv is proposed to strengthen global context modeling of point cloud representation. • DSamba is proposed to adaptively aggregate features during downsampling.