GIP-Pointnet++: A Point Cloud Semantic Segmentation Network Integrating Channel Gating and Inverted Residuals

Point cloud semantic segmentation is a core task in 3D scene understanding. Most existing methods rely on complex and sophisticated local extractors, facing the bottlenecks of high computational cost and large memory consumption. Based on PointNet++, we proposes a point cloud semantic segmentation framework that integrates channel gating and inverse residuals. Specifically, a channel gated linear unit (CGLU) is introduced as an attention mechanism, embedded into the aggregation abstraction layer of PointNet++, and enhances the discriminative expression ability of geometric features through dynamic calibration of channel responses. The inverted residual bottleneck structure is transferred to the feature propagation stage, and a “dimensional increase - transformation - dimensional reduction” serpentine MLP structure is designed to perform nonlinear recombination in the high-dimensional space, achieving cross-dimensional context aggregation. Experiments on the S3DIS dataset show that the proposed method achieves 86.1% overall classification accuracy, 66.3% average class accuracy, and 58.1% average intersection-over-union, compared to the PointNet++ baseline, with improvements of 3.1%, 4.3%, and 4.9% respectively, verifying the effectiveness of module collaboration.