Point Cloud Semantic Segmentation Method Based on Boundary Feature Enhancement

Point cloud semantic segmentation plays a pivotal role in various applications, including autonomous driving, robotics, and augmented reality, by assigning semantic labels to each 3D point. Accurate segmentation of boundary regions is particularly crucial in high-precision tasks, as it directly impacts system safety and efficiency. However, existing methods often struggle with boundary areas, leading to blurred segmentation contours and reduced overall accuracy. In this paper, we propose a novel Boundary Feature Enhancement (BFE) module designed to improve segmentation performance specifically in challenging boundary regions. The BFE module seamlessly integrates into any encoder-decoder network architecture without requiring significant modifications. It operates by predicting a boundary strength for each point, categorizing boundaries into easy-todetect and hard-to-detect classes. For the latter, the module employs a self-attention mechanism that leverages both local neighbors and globally similar points to enhance feature discriminability. Additionally, we introduce a contrastive learning loss to align the predicted boundaries with ground truth, further refining segmentation accuracy. Comprehensive experiments conducted on the S3DIS performance, particularly in challenging boundary regions. The proposed approach not only enhances boundary accuracy but also boosts overall segmentation metrics, highlighting its effectiveness and generalizability across different network architectures.