Simulated Point Clouds Explicitly Guided Semantic Segmentation

Point cloud semantic segmentation (PCSS) is crucial for smart city management but remains a challenging task due to the irregular and sparse nature of the data. While recent advancements in PCSS focus on improving network architectures, less attention has been given to the data aspect. In image analysis, synthetic data have proven useful, but generating point clouds that match real-world distributions remains difficult. In contrast, it is accessible to obtain unlimited high- density, noise-free point clouds through simulators. To enhance PCSS from the data aspect, we propose Simulated Point Clouds Explicitly Guided Semantic Segmentation (SimPCSS), a plug-and-play super- vised learning scheme. Specially, we generate labeled point clouds in various scenarios using an autonomous driving simulator and train a segmentation model. Then, multi-scale features with high confidence are then extracted to construct prior guidance through the confidence update strategy. We further introduce an imitation learning strategy, which injects the above prior guidance into the segmentation process of low-quality point clouds, improving performance. The proposed method is model-agnostic, requiring only minor adjustments to existing network architectures. Experiments conducted on both synthetic and real-world data sets with various models (MinkUnet and PTv3) demonstrate that SimPCSS effectively leverages high-quality point clouds to improve the segmentation of low-quality point clouds. The code and data are available at https://github. com/WHU-USI3DV/SimPCSS.