ASSA-Net: Semantic Segmentation Network for Point Clouds Based on Adaptive Sampling and Self-Attention

Abstract

Point cloud semantic segmentation is widely used in scene analysis. We propose a point cloud semantic segmentation network based on adaptive random sampling and self-attention. The network extracts local centroids using random sampling, enriches feature information of the centroids using the proposed adaptive optimization module, and then learns correlations and differences between feature vectors using a feature aggregation module based on the self-attentiveness mechanism to make feature cross-fertilization more adequate, which effectively improves the performance of semantic segmentation. Experimental results on S3DIS show that the network consumes less computing time, but improves the Mean Intersection over Union (mIou) by 14.4% and overall accuracy (oAcc) by 6.4% over the baseline network PointNet++.