A 3D Point Cloud Segmentation Algorithm Based on Bidirectional Feature Enhancement

To address poor boundary segmentation in point clouds, we develop a feature enhancement algorithm leveraging bidirectional learning. We first enhance the network's learning of point cloud spatial and color features through vector and bidirectional coding, enabling recognition of geometrically similar regions in 3D scenes. Furthermore, the algorithm utilized global bilateral linear regularization to collect additional local features from each point, enhancing its global perception capability. According to experimental results, the suggested method achieved 64.2% mIoU on S3DIS Area 5, outperforming some classic networks and surpassing RandLA-Net by 1.7%. This demonstrates the applicability and effectiveness of the segmentation algorithm proposed in this paper.