Urban-scale point cloud semantic segmentation with transformer

Abstract

Semantic segmentation of urban-scale point clouds is widely used in aviation, unmanned aerial vehicles, and autonomous driving. However, owing to the many points in the urban-scale point cloud dataset and massive computation in the learning process, traditional networks often have poor segmentation performance and high costs. In this study, we adopted the Point Transformer network as the baseline and integrated random point sampling and attentive pooling into a new transitiondown block, embedded in the encoder structure of the baseline to improve the speed and accuracy of semantic segmentation. On the challenging SensatUrban dataset, the Point Transformer network and the proposed network obtained mIoU values of 71.1% and 76.8%, respectively. The results show that the proposed network effectively improves the shortcomings of the Point Transformer network and achieves better semantic segmentation performance of urban-scale point clouds.