Opto-Electronic Engineering, Volume. 50, Issue 10, 230166-1(2023)

Multi-resolution feature fusion for point cloud classification and segmentation network

Zhiyong Tao1, Heng Li1、*, Miaosen Dou1, and Sen Lin2
Author Affiliations
  • 1School of Electronic and Information Engineering, Liaoning Technical University, Huludao, Liaoning 125100, China
  • 2School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang, Liaoning 110159, China
  • show less
    Figures & Tables(17)
    Multi-resolution graph convolution module algorithm flow chart
    Network framework. (a) Classification network; (b) Segmentation network
    The operation procedure of graph convolution
    The process of learning multi-resolution point cloud features
    The operation of feature fusion
    The results of the part segmentation visualization. (a) Groud truth; (b) Ours
    Comparison of segmentation details. (a) Groud truth; (b) Ours; (c) Baseline
    Noise robustness testing
    • Table 1. Experimental parameter setting

      View table
      View in Article

      Table 1. Experimental parameter setting

      参数项分类网络分割网络
      输入点数10242048
      多分辨率点云点数[896,768,640,512][896,768,640]
      图卷积分支k取值2020
      训练周期250300
      优化器SGDSGD
      训练批次3232
      测试批次1616
      初始学习速率0.10.003
    • Table 2. Comparison of classification accuracy with different methods on ModelNet40 dataset

      View table
      View in Article

      Table 2. Comparison of classification accuracy with different methods on ModelNet40 dataset

      方法输入点数/103mAcc/%OA/%
      VoxNet[8]体素-83.085.9
      MVCNN[11]多视图--90.1
      PointNet[12]坐标186.089.2
      PointNet++[13]坐标+法线5-91.9
      文献[24]坐标+法线189.891.6
      文献[25]坐标+法线1-93.0
      3D-GCN[26]坐标1-92.1
      DGCNN[15]坐标190.292.9
      LDGCNN[16]坐标190.392.9
      DDGCN[27]坐标190.492.7
      DRNet[28]坐标1-93.1
      DGANet[29]坐标189.492.3
      PCT[19]坐标1-93.2
      AFM-Net[30]坐标189.492.85
      文献[31]坐标189.0292.5
      Our坐标191.293.4
    • Table 3. Comparison of classification accuracy with different methods on ScanObjectNN dataset

      View table
      View in Article

      Table 3. Comparison of classification accuracy with different methods on ScanObjectNN dataset

      方法输入mAcc/%OA/%
      PointNet[12]坐标63.468.2
      PointNet++[13]坐标75.477.9
      DGCNN[15]坐标73.678.1
      DRNet[28]坐标78.080.3
      GBNet[32]坐标77.880.5
      PRANet[33]坐标79.182.1
      Ours坐标81.783.3
    • Table 4. Part segmentation results on the ShapeNet Part dataset

      View table
      View in Article

      Table 4. Part segmentation results on the ShapeNet Part dataset

      方法PointNet[12]PointNet++[13]文献[25]3D-GCN[26]LDGCNN[16]DGANet[29]DGCSA[34]DGCNN[15]本文
      飞机83.482.483.883.184.084.684.284.083.6
      78.779.077.584.083.085.773.383.483.4
      帐篷82.587.787.986.684.987.882.386.788.4
      74.977.378.777.578.478.577.777.878.4↑
      椅子89.690.890.890.390.691.091.090.689.7
      耳机73.071.877.374.174.477.375.374.780.5
      吉他91.591.091.890.991.091.291.291.291.8
      85.985.987.986.488.187.988.687.588.6
      台灯80.883.784.283.883.482.485.382.881.6
      手提电脑95.395.395.995.695.895.895.995.795.8↑
      摩托65.271.671.866.867.467.858.966.369.6↑
      马克杯93.094.195.194.894.994.294.394.994.4
      手枪81.281.380.981.382.381.181.881.183.7
      火箭57.958.759.659.659.259.756.963.562.5
      滑板72.876.476.675.776.075.775.474.582.0
      桌子80.682.682.482.881.982.082.782.683.0
      mIoU83.785.185.485.185.185.285.385.285.4
    • Table 5. Effect of different k values on model performance

      View table
      View in Article

      Table 5. Effect of different k values on model performance

      kOA(%)用多分辨率分支补偿后OA/%提升/%
      520.735.1+14.4
      1085.488.3+2.9
      1591.992.1+0.2
      2092.593.4+0.9
      2592.192.3+0.2
    • Table 6. Ablation experiments of multi-resolution GCN module

      View table
      View in Article

      Table 6. Ablation experiments of multi-resolution GCN module

      实验GCN分支M-R分支融合mAcc/%OA/%
      1××89.992.5
      2××84.089.1
      3×89.992.6
      491.293.4
    • Table 7. The effect of different resolution point cloud on network performance

      View table
      View in Article

      Table 7. The effect of different resolution point cloud on network performance

      不同分辨率的点云mAcc/%OA/%
      [512,384,256,128]90.492.7
      [640,512,384,256]90.692.8
      [768,640,512,384]90.993.0
      [896,768,640,512]91.293.4
    • Table 8. Comparison of the noise robustness of the several methods

      View table
      View in Article

      Table 8. Comparison of the noise robustness of the several methods

      噪声水平下降程度
      3D-GCNAdaptConvDGCNNOurs
      0.020.71.8↓1.4↓0.9↓
      0.042.2↓2.2↓2.2↓1.8
      0.064.6↓3.3↓3.2↓3.1
      0.088.4↓6.5↓5.76.4↓
      0.114.9↓10.813.1↓11.7↓
    • Table 9. The impact of different number of feature extraction modules on network performance

      View table
      View in Article

      Table 9. The impact of different number of feature extraction modules on network performance

      模块数量mAcc/%OA/%每轮训练时间/s模型参数量/M
      389.792.4632.8
      491.293.41393.6
      590.693.13234.8
    Tools

    Get Citation

    Copy Citation Text

    Zhiyong Tao, Heng Li, Miaosen Dou, Sen Lin. Multi-resolution feature fusion for point cloud classification and segmentation network[J]. Opto-Electronic Engineering, 2023, 50(10): 230166-1

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Research Articles

    Received: Jul. 7, 2023

    Accepted: Sep. 20, 2023

    Published Online: Jan. 22, 2024

    The Author Email: Li Heng (李衡)

    DOI:10.12086/oee.2023.230166

    Topics