Semi-supervised point cloud semantic segmentation via cross-learning for sewer inspection

In recent years, notable progress has been made in sewer inspection using point cloud semantic segmentation. While deep learning-based methods have shown considerable promise for fully supervised point cloud semantic segmentation, the labor-intensive and costly process of point labeling remains a challenge. This study proposes a semi-supervised point cloud semantic segmentation (SPCSS) method based on cross-learning principles. Our SPCSS method integrates a Local Feature Extraction Network (LFEN) and a Global Feature Extraction Network (GFEN) to address the unique challenges of sonar point clouds, which exhibit sparse axial sampling (due to slow sonar traversal) and dense radial noise (from acoustic scattering). Both labeled and unlabeled point clouds are processed by LFEN and GFEN, generating predictions for each individual point. Unlabeled points are then assigned pseudo-labels derived from the outputs of LFEN and GFEN, enabling the mutual updating of network parameters and fostering cross-learning between the two networks. This cross-learning mechanism captures both local and global features, thereby addressing the non-uniform spatial distribution of point clouds. Furthermore, our SPCSS method incorporates adaptive equalization sampling and reweighting strategies to mitigate performance degradation for rare but critical categories (e.g., external outliers, sedimention) caused by class imbalance and sparse labeled data. Experimental results demonstrate that our SPCSS method outperforms other semi-supervised approaches and achieves performance on par with state-of-the-art supervised learning methods.