Optics and Precision Engineering, Volume. 32, Issue 22, 3366(2024)

Design of partial overlap point cloud registration network driven by overlap score and matching matrix

Jianbing YI1...2,*, Xin CHEN1,2, Feng CAO1,2, Shuxin YANG1,2 and Jingyong WANG3 |Show fewer author(s)
Author Affiliations
  • 1School of Information Engineering, Jiangxi University of Science and Technology, Ganzhou34000, China
  • 2Jiangxi Province Key Laboratory of Multidimensional Intelligent Perception and Control, Ganzhou341000, China
  • 3Longnan Dingtai electronic Technology Co., Ltd., Ganzhou41000, China
  • show less
    Figures & Tables(24)
    Partial overlap point cloud registration network structure
    Channel space aware Transformer
    Feature concatenate operator and feature add operator
    Feature encoder based on Transformer
    Diagram of adding slack variables
    Flow chart of outlier removal
    Visualization results of toilet registration in ModelNet40 dataset
    Visualization results of table registration in ModelNet40 dataset
    Visualization results of window registration in ShapeNetCore.v2 dataset
    Visualization results of car registration in ShapeNetCore.v2 dataset
    Visualization results of Bunny registration in Stanford Bunny dataset
    Visualization results of different registration algorithms in ModelNet40 dataset
    • Table 1. Experimental results of different algorithms on the ModelNet40 dataset

      View table
      View in Article

      Table 1. Experimental results of different algorithms on the ModelNet40 dataset

      MethodMIE(RMIE(tMAE(RMAE(t
      ICP1524.641 30.252 512.147 00.117 1
      FGR1614.011 40.097 88.274 40.046 2
      DeepGMR1714.361 20.158 97.091 40.077 5
      IDAM614.289 10.190 97.496 60.087 7
      RPMNet51.423 90.013 90.730 40.006 5
      ROPNet81.154 10.010 80.592 80.005 1
      GAP-Net181.487 10.015 2
      GTGMM190.560 60.010 2
      Ours0.906 80.008 20.462 60.003 9
    • Table 2. Experimental results of different algorithms on the ModelLoNet40 dataset

      View table
      View in Article

      Table 2. Experimental results of different algorithms on the ModelLoNet40 dataset

      MethodMIE(RMIE(t
      RPMNet57.3420.124
      Predator75.2350.132
      ROPNet85.0450.049
      REGTR203.9300.087
      UDPReg213.5780.069
      HPNet224.1280.787
      Ours3.3040.031
    • Table 3. Experimental results of different algorithms on the ShapeNetCore.v2 dataset

      View table
      View in Article

      Table 3. Experimental results of different algorithms on the ShapeNetCore.v2 dataset

      MethodRMSE(RRMSE(tMAE(RMAE(t
      IDAM61.7460.0140.7570.006
      ReAgent231.7910.0281.5220.010
      ROPNet80.6730.0060.3200.003
      DORNet241.3760.0090.7340.006
      UTOPIC110.6820.0080.2810.002
      MGFPCR250.5810.0020.2530.001
      Ours0.3370.0040.0970.001
    • Table 4. Experimental results of generalization performance of different algorithms on the ModelNet40 dataset

      View table
      View in Article

      Table 4. Experimental results of generalization performance of different algorithms on the ModelNet40 dataset

      MethodMIE(RMIE(tMAE(RMAE(t
      ICP1525.487 90.250 412.897 60.118 3
      FGR1615.331 70.103 69.411 00.049 1
      DeepGMR1718.360 80.199 89.142 50.099 7
      IDAM617.361 40.215 38.982 30.102 2
      RPMNet51.678 10.016 90.874 90.007 9
      ROPNet81.252 30.012 40.668 70.005 9
      REGTR200.965 30.009 70.509 70.004 6
      DLF260.478 10.004 3
      GANPCR275.158 10.043 1
      Ours0.915 40.009 00.473 10.004 3
    • Table 5. Experimental results of generalization performance of different algorithms on the Stanford Bunny dataset

      View table
      View in Article

      Table 5. Experimental results of generalization performance of different algorithms on the Stanford Bunny dataset

      MethodRMSE(RRMSE(tMAE(RMAE(t
      ICP1513.320.049 210.720.024 2
      Go-ICP2812.920.042 94.520.028 2
      DCP296.440.040 64.780.037 4
      ROPNet83.760.028 92.330.015 9
      Ours2.800.019 72.200.015 1
    • Table 6. Experimental results of different algorithms under Gaussian noise with a standard deviation of 0.01

      View table
      View in Article

      Table 6. Experimental results of different algorithms under Gaussian noise with a standard deviation of 0.01

      MethodMIE(RMIE(tMAE(RMAE(t
      ICP1525.521 90.251 512.911 80.119 0
      FGR1629.281 30.182 818.198 60.188 7
      DeepGMR1718.807 50.202 79.367 20.099 2
      IDAM616.980 50.212 58.936 60.100 3
      RPMNet51.756 90.017 70.917 70.008 4
      ROPNet81.573 20.015 50.838 60.007 4
      MFGNet3112.266 80.110 66.341 90.055 3
      FINet327.206 80.085 33.733 30.040 6
      REGTR201.795 90.017 10.931 00.008 1
      OGMM3314.482 50.189 47.290 10.086 1
      RGM341.736 60.015 90.937 40.007 6
      GeoTransformer351.435 00.018 10.725 20.008 1
      Ours1.251 40.012 20.654 30.005 9
    • Table 7. Experimental results of different algorithms under Gaussian noise with a standard deviation of 0.02

      View table
      View in Article

      Table 7. Experimental results of different algorithms under Gaussian noise with a standard deviation of 0.02

      MethodMIE(RMIE(tMAE(RMAE(t
      ICP1530.709 60.462 015.704 20.221 5
      FGR1626.311 10.224 053.048 30.438 2
      DCP2930.709 60.462 015.704 20.221 5
      IDAM613.651 70.167 97.102 70.084 7
      RPMNet51.862 30.019 60.970 40.009 4
      PREDATOR72.806 50.027 24.782 20.050 0
      ROPNet82.004 60.019 71.064 00.009 5
      MFGNet3112.725 10.119 66.568 90.060 2
      MLIPPCR371.783 70.018 00.942 70.008 8
      Ours1.797 70.017 40.954 50.008 4
    • Table 8. Experimental results of different algorithms under Gaussian noise with a standard deviation of 0.03

      View table
      View in Article

      Table 8. Experimental results of different algorithms under Gaussian noise with a standard deviation of 0.03

      MethodMIE(RMIE(tMAE(RMAE(t
      ICP1530.964 10.467 915.742 40.225 8
      FGR1647.788 30.533 529.070 70.254 2
      DCP2913.890 50.176 37.010 20.088 5
      IDAM613.593 90.169 67.074 30.085 2
      RPMNet52.231 80.023 51.160 00.011 2
      PREDATOR77.704 30.082 84.055 80.039 6
      ROPNet82.568 60.025 11.359 30.012 2
      MFGNet3113.857 80.129 47.110 90.065 0
      MLIPPCR372.403 00.024 41.270 50.011 9
      Ours2.277 50.022 01.200 40.010 7
    • Table 9. Experimental results of different algorithms under Gaussian noise with a standard deviation of 0.04

      View table
      View in Article

      Table 9. Experimental results of different algorithms under Gaussian noise with a standard deviation of 0.04

      MethodMIE(RMIE(tMAE(RMAE(t
      ICP1531.172 00.467 415.800 50.227 0
      FGR1656.843 50.617 335.528 50.293 7
      DCP2913.970 70.176 57.042 20.088 6
      IDAM613.891 50.173 97.145 90.087 4
      RPMNet52.610 30.028 01.357 70.013 3
      PREDATOR710.222 50.107 05.330 90.051 1
      ROPNet83.154 20.031 01.673 40.015 1
      MFGNet3114.702 90.137 57.588 60.069 2
      MLIPPCR372.998 50.030 51.573 80.014 8
      Ours2.577 70.024 81.382 10.012 1
    • Table 10. Real time experimental results of different algorithms on the ModelNet40 dataset

      View table
      View in Article

      Table 10. Real time experimental results of different algorithms on the ModelNet40 dataset

      Method512(Points)1 024(Points)2 048(Points)
      ICP150.1280.1620.231
      Go-ICP2814.05314.14014.212
      FGR160.1720.1870.185
      PointNetLK380.0590.0740.084
      ROPNet80.0220.0310.053
      Ours0.0400.0520.086
    • Table 11. The impact of each improvement step of this algorithm on the ModelNet40 dataset on the model

      View table
      View in Article

      Table 11. The impact of each improvement step of this algorithm on the ModelNet40 dataset on the model

      MethodAOTO
      MIE(RMIE(tMIE(RMIE(t
      Baseline1.203 10.013 71.573 20.015 5
      Method 11.141 30.013 11.419 80.014 0
      Method 21.022 20.011 21.349 80.013 0
      Method 30.914 70.010 31.251 50.012 2
    • Table 12. Overlap precision and recall ablation experiment of this algorithm on the ModelNet40 dataset

      View table
      View in Article

      Table 12. Overlap precision and recall ablation experiment of this algorithm on the ModelNet40 dataset

      Method重叠点(0.05/0.07)代表性重叠点(0.05/0.07)
      AOTOAOTO
      OPOROPOROPOROPOR
      Baseline0.810/0.9390.836/0.8220.806/0.9360.838/0.8240.913/0.9870.382/0.3470.908/0.9840.382/0.348
      Method 10.810/0.9390.835/0.8230.805/0.9360.837/0.8240.912/0.9860.381/0.3470.908/0.9840.382/0.348
      Method 20.826/0.9540.853/0.8370.820/0.9500.853/0.8380.919/0.9890.384/0.3480.914/0.9860.385/0.349
      Method 30.826/0.9540.854/0.8370.820/0.9510.854/0.8380.933/0.9920.391/0.3490.927/0.9890.391/0.350
    Tools

    Get Citation

    Copy Citation Text

    Jianbing YI, Xin CHEN, Feng CAO, Shuxin YANG, Jingyong WANG. Design of partial overlap point cloud registration network driven by overlap score and matching matrix[J]. Optics and Precision Engineering, 2024, 32(22): 3366

    Download Citation

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

    Category:

    Received: Jun. 3, 2024

    Accepted: --

    Published Online: Mar. 10, 2025

    The Author Email: YI Jianbing (yijianbing8@jxust.edu.cn)

    DOI:10.37188/OPE.20243222.3366

    Topics