Coupled Geometry-Sequence State-Space Model for Point Cloud Semantic Segmentation

Abstract

The complex spatial structures and varying scales of point cloud data make semantic segmentation a highly challenging task. While downsampling and feature fusion are essential steps in mainstream networks, existing models often suffer from local geometry loss and multiscale semantic conflicts during these processes. To alleviate these issues, we propose coupled geometry-sequence (CGS)-Mamba, a novel state-space model (SSM)-based network for point cloud semantic segmentation. Specifically, we propose the CSG block (CGS-Block). It extracts fine-grained local geometric features and effectively captures global long-range dependencies. Furthermore, we introduce the adaptive gated fusion (AGF) module in the decoder, which can mitigate semantic conflicts during cross-layer feature fusion through on-demand feature absorption. Experiments show that CGS-Mamba achieves mean intersection over union (mIoU) scores of 86.83% and 75.46% on the Tinto and Toronto-3-D datasets, respectively, along with overall accuracy (OA) of 94.59% and 93.35%. Notably, compared to baselines, our model achieves significant mIoU improvements of 1.53% and 2.53% on the two datasets, respectively, while nearly halving the parameters to 16.16 M. The code and models are available at https://github.com/MMYY-LL/CGS-Mamba