Optics and Precision Engineering, Volume. 30, Issue 24, 3210(2022)

Robust point cloud registration of terra-cotta warriors based on dynamic graph attention mechanism

Linqi HAI... Guohua GENG, Xing YANG, Kang LI and Haibo ZHANG* |Show fewer author(s)
Author Affiliations
  • School of Information Science and Technology, Northwest University, Xi’an710127, China
  • show less
    Figures & Tables(17)
    Dynamic graph attention mechanism module structure
    Overall network structure of DGATNet
    Registration results of point clouds with low overlap
    Point cloud registration results in the resolution mismatch setting
    Influence of Gaussian noise on different models
    Influence of inlier ratio threshold τ2 (left) and inlier distance threshold τ1 (right) on feature matching recall
    Terracotta warriors point clouds data
    Registration results of head and foot of the terracotta warriors point clouds
    Influence of number of layers of dynamic graph attention on the model
    • Table 1. Evaluation results on the 3DMatch standard dataset

      View table
      View in Article

      Table 1. Evaluation results on the 3DMatch standard dataset

      Samples5 0002 5001 000500250
      Feature Matching Recall(%)↑
      FCGF1797.297.097.296.494.8
      D3Feat1895.895.694.694.393.3
      Predator2796.696.696.596.396.5
      DGATNet(ours)97.697.597.897.597.1
      Inlier Ratio(%)↑
      FCGF1755.552.947.541.233.0
      D3Feat1840.740.642.744.145.0
      Predator2758.058.457.154.149.3
      DGATNet(ours)63.463.561.758.452.8
      Registration Recall(%)↑
      FCGF1785.084.785.582.275.0
      D3Feat1882.284.484.982.579.3
      Predator2789.089.589.888.585.5
      DGATNet(ours)89.489.390.189.187.5
    • Table 2. Evaluation results on the 3DLoMatch dataset

      View table
      View in Article

      Table 2. Evaluation results on the 3DLoMatch dataset

      Samples5 0002 5001000500250
      Feature Matching Recall(%)↑
      FCGF1775.274.573.771.865.9
      D3Feat1867.366.767.066.766.6
      Predator2778.677.476.375.775.3
      DGATNet(ours)78.279.478.878.476.8
      Inlier Ratio(%)↑
      FCGF1720.419.116.514.111.1
      D3Feat1813.213.114.014.615.0
      Predator2726.728.128.327.525.8
      DGATNet(ours)28.430.329.629.026.8
      Registration Recall(%)↑
      FCGF1742.440.138.835.725.2
      D3Feat1837.242.746.943.839.1
      Predator2759.861.261.860.857.8
      DGATNet(ours)60.061.060.860.658.2
    • Table 3. Evaluation results on the 3DMatch resolution-mismatch dataset

      View table
      View in Article

      Table 3. Evaluation results on the 3DMatch resolution-mismatch dataset

      λi/cm2.54.55.56.5
      Feature Matching Recall(%)↑
      FCGF1776.774.912.25.6
      D3Feat1892.492.687.977.9
      Predator2795.394.887.258.2
      DGATNet(ours)96.495.193.791.9
      Inlier Ratio(%)↑
      FCGF1720.216.42.01.6
      D3Feat1842.534.723.318.9
      Predator2742.637.424.610.2
      DGATNet(ours)58.452.549.743.9
      Registration Recall(%)↑
      FCGF1749.647.314.18.2
      D3Feat1877.779.276.060.8
      Predator2785.184.278.557.6
      DGATNet(ours)86.986.885.381.2
    • Table 4. Registration results of head of the terracotta warriors point clouds

      View table
      View in Article

      Table 4. Registration results of head of the terracotta warriors point clouds

      Number of pointsOverlap ratioMethodsRRE/radRTE/mTime/s
      ICP401.1840.2310.069
      FPFH11+RANSAC200.7570.1763.723
      FPFH11+Teaser++380.1810.0490.400
      3 741,7 71940.7%FCGF171.0100.2150.190
      D3Feat180.7510.1269.526
      Predator270.9330.1630.429
      DGATNet0.0710.0160.435
    • Table 5. Registration results of foot of the terracotta warriors point clouds

      View table
      View in Article

      Table 5. Registration results of foot of the terracotta warriors point clouds

      Number of pointsOverlap ratioMethodsRRE/radRTE/mTime/s
      ICP401.5540.0910.052
      FPFH11+RANSAC200.5790.1113.774
      FPFH11+Teaser++380.1260.0460.467
      3 946,8 26030.2%FCGF170.1900.0960.194
      D3Feat180.1490.0659.760
      Predator270.0750.0610.516
      DGATNet0.0600.0110.531
    • Table 6. Influence of number of K-nearest neighbor points in the attention module on the registration results

      View table
      View in Article

      Table 6. Influence of number of K-nearest neighbor points in the attention module on the registration results

      Methods3DMatch
      FMR(%)↑IR(%)↑RR(%)↑
      DGATNet v196.957.487.3
      DGATNet v297.260.087.5
      DGATNet v397.661.288.3
      DGATNet(ours)97.861.790.1
      Methods3DLoMatch
      FMR(%)↑IR(%)↑RR(%)↑
      DGATNet v175.726.755.6
      DGATNet v276.026.656.1
      DGATNet v376.927.458.9
      DGATNet(ours)78.829.660.8
      DGATNet v1:kself,kcrossfully connected
      DGATNet v2:k3,4self=k3,4cross=128,k5,6self=k5,6cross=128
      DGATNet v3:k3,4self=128,k5,6self=128,kcrossfully connected
      DGATNet(ours):k3,4self=64,k5,6self=32,kcrossfully connected
    • Table 7. Ablation experiments of BNHN normalization technique

      View table
      View in Article

      Table 7. Ablation experiments of BNHN normalization technique

      Methods3DMatch(resolution-mismatch)
      FMR(%)↑IR(%)↑RR(%)↑
      Baseline8.21.65.6
      Baseline+BNHN85.230.575.2
      Baseline+Attention87.438.476.6
      DGATNet91.943.981.2
    • Table 8. Ablation experiments of each component in DGATNet

      View table
      View in Article

      Table 8. Ablation experiments of each component in DGATNet

      Methods3DMatch
      FMR(%)↑IR(%)↑RR(%)↑
      Baseline97.247.585.5
      Baseline+BNHN96.346.785.3
      Baseline+Attention97.560.088.3
      Baseline+Attention+Position embedding97.561.288.7
      Baseline+Attention+Sin position embedding97.861.790.1
    Tools

    Get Citation

    Copy Citation Text

    Linqi HAI, Guohua GENG, Xing YANG, Kang LI, Haibo ZHANG. Robust point cloud registration of terra-cotta warriors based on dynamic graph attention mechanism[J]. Optics and Precision Engineering, 2022, 30(24): 3210

    Download Citation

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

    Category: Information Sciences

    Received: May. 12, 2022

    Accepted: --

    Published Online: Feb. 15, 2023

    The Author Email: ZHANG Haibo (zhanghb@nwu.edu.cn)

    DOI:10.37188/OPE.20223024.3210

    Topics