Remote Sensing Technology and Application

Semantic Segmentation Method of Street Point Cloud based on Deep-Supervised Multi-Scale Self-Attention

Guangchao LIU, Chenxiao ZHANG, Lei HU

Remote Sensing Technology and Application, 2025, 40(6): 1626-1636. DOI: 10.11873/j.issn.1004-0323.2025.6.1626

方法	扩张率	近邻点数量	OA/%	mIoU/%
单编码器	1	16	96.66	77.71
单编码器	2	32	96.81	81.35
单编码器	3	48	96.68	79.41
双编码器	1,2	16,32	97.08	81.41
双编码器	2,3	32,48	97.15	81.46
三编码器	1,2,3	16,32,48	97.16	82.17

Table 4 Results of ablation study on Toronto3D dataset

Other figure/table from this article

Fig.1 Network structure of 3D point cloud semantic segmentation based on deep-supervised multi-scale self-attention
Fig.2 The structure of the dilated nearest neighbor self-attention module
Fig.3 The idea of dilated k-nearest neighbors
Fig.4 Structure of dilated down-sampling module， nearest neighbor up-sampling module and multi-scale attention aggregation module
Table 1 Semantic segmentation results of different models on Toronto 3D dataset
Fig.5 Visualization results on Toronto 3D dataset and CSPC dataset
Table 2 Semantic segmentation results of different models on CSPC dataset Scene2
Table 3 Semantic segmentation results of different models on CSPC dataset Scene5
Fig.6 Confusion matrix of Toronto3D dataset
Fig.7 Confusion matrix of CSPC dataset Scene2
Fig.8 Confusion matrix of CSPC dataset Scene5
Fig.9 Visualization results of some mispredictions on Toronto3D dataset
Fig.10 Visualization results of some mispredictions on CSPC dataset
Table 5 Test efficiency of each method on Toronto3D dataset