Improved Indoor 3D Point Cloud Semantic Segmentation Method Based on PointNet++

With the rapid advancement of point cloud-based applications, including autonomous driving and robotic navigation, 3D point cloud semantic segmentation has gained increasing research attention. PointNet++ is a widely recognized neural architecture for point cloud analysis. However, its inefficient local feature learning and limited global context capture reduce segmentation accuracy. To overcome these limitations, this paper presents an enhanced PointNet++-based model for indoor 3D point cloud segmentation. Specifically, FastKAN (Kolmogorov-Arnold Networks) replaces all MLP layers, utilizing an attention mechanism during feature aggregation to emphasize important features while preserving both local and global details, thereby enhancing segmentation performance. Additionally, the SA (Shuffle Attention) mechanism is integrated, employing channel grouping and rearrangement to effectively capture both local and global information while ensuring computational efficiency. Experimental results on the Stanford S3DIS dataset show that the proposed model achieves an overall accuracy of 86.5%, reflecting a 3.6% improvement over PointNet++, and a mean Intersection over Union (mIoU) of 57.3%, 3.8% higher than PointNet++.