Self-attention Mechanism and Pyramid Feature Fusion for Point Cloud Segmentation

This study introduces RandLASAMP-Net as an advanced version of RandLA-Net, designed to facilitate efficient per-point semantics inference for large-scale 3D point clouds. Existing algorithms encounter difficulties in processing large point clouds due to costly sampling or pre/post-processing necessities. To enhance the feature aggregation module of RandLA-Net, which relies on a rudimentary attention mechanism, we propose the integration of a self-attention mechanism. This mechanism enables the model to focus on different regions of the point cloud, enhancing its capacity to extract significant features from small and unstructured point sets. Furthermore, we introduce a multi-scale pyramid module that adapts to the UNet-shaped architecture for the point cloud to better utilize encoding features. Our experimental results demonstrate that these enhancements improve the accuracy of semantic segmentation while maintaining computational and memory efficiency for large-scale point clouds. Our goal is to develop a reliable model that can efficiently handle large-scale point clouds with high accuracy.