Mining local geometric structure for large-scale 3D point clouds semantic segmentation

In this paper, we concentrate on how to preserve fine-grained geometric structure information when extracting local contextual features for efficient large-scale point clouds semantic segmentation. Firstly, the Local Geometric Structure Representation Block is proposed to model fine-grained geometric structures for individual points by fully utilizing relative and global geometric relationships in the neighborhood. Then, we design a Parallel Attentive Fusion Module focusing on geometric structure and semantic information respectively, which reduces the feature’s ambiguity and preserves local geometric structure information. Furthermore, thanks to these two modules, we present a lightweight Local Contextual Features Extractor through the bilateral structure to obtain more discriminate local contextual features. Finally, a deep network named LGS-Net is introduced to predict point’s classes. Extensive experiment shows that our network surpasses the state-of-the-art approaches for semantic segmentation on two public large-scale point clouds datasets Semantic3D, SensatUrban, and achieve competitive performance on S3DIS. Especially, our LGS-Net with minimal model size outperforms the state-of-the-art network by 1.2% on the Semantic3D dataset. Thorough ablation studies and visualizations are presented to understand our network.