• Laser & Optoelectronics Progress
  • Vol. 60, Issue 2, 0228007 (2023)
Kun Zhang1, Yawei Zhu1, Xiaohong Wang1, Liting Zhang1, and Ruofei Zhong2,*
Author Affiliations
  • 1College of Information Science and Engineering, Hebei University of Science and Technology, Hebei 050018, Shijiazhuang, China
  • 2College of Resources Environment and Tourism, Capital Normal University, Beijing 100048, China
  • show less
    DOI: 10.3788/LOP212825 Cite this Article Set citation alerts
    Kun Zhang, Yawei Zhu, Xiaohong Wang, Liting Zhang, Ruofei Zhong. Three-Dimensional Point Cloud Semantic Segmentation Network Based on Spatial Graph Convolution Network[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0228007 Copy Citation Text show less
    PCGCN problem description diagram
    Fig. 1. PCGCN problem description diagram
    Overall structure diagram of PCGCN
    Fig. 2. Overall structure diagram of PCGCN
    EdgeConv subnet. (a) 2nd layerneighborhood aggregation; (b) 1st layer neighborhood aggregation; (c) local features graph
    Fig. 3. EdgeConv subnet. (a) 2nd layerneighborhood aggregation; (b) 1st layer neighborhood aggregation; (c) local features graph
    PCGCN residual network diagram
    Fig. 4. PCGCN residual network diagram
    Aggregating characteristics of different layers
    Fig. 5. Aggregating characteristics of different layers
    KNN diagrams. (a) Original KNN; (b) randomly delete nodes; (c) original sampling result; (d) sampling result after improvement
    Fig. 6. KNN diagrams. (a) Original KNN; (b) randomly delete nodes; (c) original sampling result; (d) sampling result after improvement
    Comparison of training results of S3DIS.(a) Train average accuracy; (b) train accuracy; (c) train IoU
    Fig. 7. Comparison of training results of S3DIS.(a) Train average accuracy; (b) train accuracy; (c) train IoU
    Visualization of segmentation results in S3DIS
    Fig. 8. Visualization of segmentation results in S3DIS
    Accuracy of ShapeNet test set
    Fig. 9. Accuracy of ShapeNet test set
    Visualization of segmentation results in ShapeNet
    Fig. 10. Visualization of segmentation results in ShapeNet
    Robustness analysis on ShapeNet
    Fig. 11. Robustness analysis on ShapeNet
    Anti noise analysis on S3DIS
    Fig. 12. Anti noise analysis on S3DIS
    num_pointsk-mmEvery epoch /sMean IoU
    2048200144241.79
    2048155167942.34
    2048153175742.87
    2048123137042.23
    1024207123442.13
    1024205120042.03
    1024203112341.78
    1024257154743.44
    102415390041.58
    102412376241.35
    10249367038.24
    Table 1. Parameter comparison experiment
    ModelMean IoUOverall accuracy
    PointNet40.476.0
    PointNet++46.279.6
    MS+CU47.879.2
    DeepGCN52.283.5
    GAC47.879.8
    DGCNN49.180.8
    Proposed model49.581.3
    Proposed simplified model49.381.0
    Table 2. Semantic segmentation results of S3DIS
    ModelEvery epoch /s
    PointNet3840
    PointNet++25500
    DeepGCN5400
    GAC2939
    DGCNN812
    Proposed model1242
    Proposed simplified model762
    Table 3. Semantic Segmentation time of S3DIS
    batch_sizetrain mAcctest mAcctrain mIoUtest mIoU
    291.689.279.277.3
    892.891.082.281.4
    1694.193.083.582.8
    3294.892.884.982.9
    Table 4. Parameter sensitivity experiment
    ModelKd-NetSK-NetSRI-NetPointNetPointNet++DGCNNProposed modelProposed simplified model
    mIoU /%82.385.073.583.785.185.085.184.7
    Table 5. Semantic segmentation results of ShapeNet
    Kun Zhang, Yawei Zhu, Xiaohong Wang, Liting Zhang, Ruofei Zhong. Three-Dimensional Point Cloud Semantic Segmentation Network Based on Spatial Graph Convolution Network[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0228007
    Download Citation