LiDAR-to-SAR Point Cloud Segmentation via Unsupervised Domain Adaptation Network

Synthetic Aperture Radar (SAR) 3D point cloud reconstruction improves target identification by mitigating issues like overlay masking and shadows found in 2D SAR image projections. Nevertheless, challenges arise from limited SAR data availability and complexities in deciphering and labeling, hindering research in SAR 3D reconstruction point cloud semantic segmentation. To address these hurdles, we propose an alternative training approach—shifting from LiDAR to SAR point clouds. Leveraging benchmark datasets in the LiDAR domain, we advocate using LiDAR point clouds for training to counter the scarcity of SAR training sets. However, applying a segmentation model across different domains leads to a performance decline, particularly in cross-modal SAR-reconstructed point clouds, attributed to distinct roughness introduced by outliers. This paper introduces a 3D semantic segmentation framework based on unsupervised domain adaptation (UDA) for cross-modal learning from LiDAR to SAR. Additionally, we present a simple yet effective geometric transformation data augmentation technique to handle highly imbalanced data distribution. Experimental results confirm the feasibility and effectiveness of our proposed method for SAR 3D reconstructed point cloud semantic segmentation.