SADNet: Space-aware DeepLab network for Urban-Scale point clouds semantic segmentation

Abstract

Semantic segmentation of urban-scale point clouds can effectively assist people in understanding and perceiving 3D urban scenes. Although a considerable number of deep learning models for the semantic segmentation of point clouds have been proposed, some methods are plagued by information loss caused by sampling and insufficient perception of the spatial relationship between points. To address this issue, this paper proposes an end-to-end space-aware DeepLab deep learning network, named SADNet. In the SADNet, a space-aware attentive residual module (SARM) is incorporated to extract rich point cloud features with the assistance of perceiving spatial relationships between points. Then, in combination with a point cloud atrous spatial pyramid pooling module (PCaspp), SADNet extracts multiscale point cloud features while effectively avoiding information loss from pooling and downsampling. Finally, a dilated local feature extraction (DLFE) module is designed to enhance the detection ability for small objects by dilating the feature map. Furthermore, to validate the superiority of the SADNet, extensive experiments are conducted on two publicly available benchmarks, Sensaturban and Hessigheim 3D. The results demonstrate the state-of-the-art performance on both datasets, which achieves the mean IoU of 66.8% on Sensaturban and overall accuracy of 91.77%, mean F1-score of 82.81% on Hessigheim 3D. Overall, SADNet is a promising approach for urban-scale point cloud semantic segmentation and has the potential to enhance understanding and perception of real-world urban scenes.