Optics and Precision Engineering, Volume. 30, Issue 24, 3210(2022)
Robust point cloud registration of terra-cotta warriors based on dynamic graph attention mechanism
Figures & Tables(17)
Fig. 4. Point cloud registration results in the resolution mismatch setting
Fig. 6. Influence of inlier ratio threshold
Fig. 8. Registration results of head and foot of the terracotta warriors point clouds
Fig. 9. Influence of number of layers of dynamic graph attention on the model
Table 1. Evaluation results on the 3DMatch standard dataset
View tableView in ArticleTable 1. Evaluation results on the 3DMatch standard dataset
Samples 5 000 2 500 1 000 500 250 Feature Matching Recall（%）↑ FCGF［ 17 ］97.2 97.0 97.2 96.4 94.8 D3Feat［ 18 ］95.8 95.6 94.6 94.3 93.3 Predator［ 27 ］96.6 96.6 96.5 96.3 96.5 DGATNet（ours） 97.6 97.5 97.8 97.5 97.1 Inlier Ratio（%）↑ FCGF［ 17 ］55.5 52.9 47.5 41.2 33.0 D3Feat［ 18 ］40.7 40.6 42.7 44.1 45.0 Predator［ 27 ］58.0 58.4 57.1 54.1 49.3 DGATNet（ours） 63.4 63.5 61.7 58.4 52.8 Registration Recall（%）↑ FCGF［ 17 ］85.0 84.7 85.5 82.2 75.0 D3Feat［ 18 ］82.2 84.4 84.9 82.5 79.3 Predator［ 27 ］89.0 89.5 89.8 88.5 85.5 DGATNet（ours） 89.4 89.3 90.1 89.1 87.5
Table 2. Evaluation results on the 3DLoMatch dataset
View tableView in ArticleTable 2. Evaluation results on the 3DLoMatch dataset
Samples 5 000 2 500 1000 500 250 Feature Matching Recall（%）↑ FCGF［ 17 ］75.2 74.5 73.7 71.8 65.9 D3Feat［ 18 ］67.3 66.7 67.0 66.7 66.6 Predator［ 27 ］78.6 77.4 76.3 75.7 75.3 DGATNet（ours） 78.2 79.4 78.8 78.4 76.8 Inlier Ratio（%）↑ FCGF［ 17 ］20.4 19.1 16.5 14.1 11.1 D3Feat［ 18 ］13.2 13.1 14.0 14.6 15.0 Predator［ 27 ］26.7 28.1 28.3 27.5 25.8 DGATNet（ours） 28.4 30.3 29.6 29.0 26.8 Registration Recall（%）↑ FCGF［ 17 ］42.4 40.1 38.8 35.7 25.2 D3Feat［ 18 ］37.2 42.7 46.9 43.8 39.1 Predator［ 27 ］59.8 61.2 61.8 60.8 57.8 DGATNet（ours） 60.0 61.0 60.8 60.6 58.2
Table 3. Evaluation results on the 3DMatch resolution-mismatch dataset
View tableView in ArticleTable 3. Evaluation results on the 3DMatch resolution-mismatch dataset
λi/cm 2.5 4.5 5.5 6.5 Feature Matching Recall（%）↑ FCGF［ 17 ］76.7 74.9 12.2 5.6 D3Feat［ 18 ］92.4 92.6 87.9 77.9 Predator［ 27 ］95.3 94.8 87.2 58.2 DGATNet（ours） 96.4 95.1 93.7 91.9 Inlier Ratio（%）↑ FCGF［ 17 ］20.2 16.4 2.0 1.6 D3Feat［ 18 ］42.5 34.7 23.3 18.9 Predator［ 27 ］42.6 37.4 24.6 10.2 DGATNet（ours） 58.4 52.5 49.7 43.9 Registration Recall（%）↑ FCGF［ 17 ］49.6 47.3 14.1 8.2 D3Feat［ 18 ］77.7 79.2 76.0 60.8 Predator［ 27 ］85.1 84.2 78.5 57.6 DGATNet（ours） 86.9 86.8 85.3 81.2
Table 4. Registration results of head of the terracotta warriors point clouds
View tableView in ArticleTable 4. Registration results of head of the terracotta warriors point clouds
Number of points Overlap ratio Methods RRE/rad RTE/m Time/s ICP［ 40 ］1.184 0.231 0.069 FPFH［ 11 ］+RANSAC［20 ］0.757 0.176 3.723 FPFH［ 11 ］+Teaser++［38 ］0.181 0.049 0.400 3 741，7 719 40.7% FCGF［ 17 ］1.010 0.215 0.190 D3Feat［ 18 ］0.751 0.126 9.526 Predator［ 27 ］0.933 0.163 0.429 DGATNet 0.071 0.016 0.435
Table 5. Registration results of foot of the terracotta warriors point clouds
View tableView in ArticleTable 5. Registration results of foot of the terracotta warriors point clouds
Number of points Overlap ratio Methods RRE/rad RTE/m Time/s ICP［ 40 ］1.554 0.091 0.052 FPFH［ 11 ］+RANSAC［20 ］0.579 0.111 3.774 FPFH［ 11 ］+Teaser++［38 ］0.126 0.046 0.467 3 946，8 260 30.2% FCGF［ 17 ］0.190 0.096 0.194 D3Feat［ 18 ］0.149 0.065 9.760 Predator［ 27 ］0.075 0.061 0.516 DGATNet 0.060 0.011 0.531
Table 6. Influence of number of K-nearest neighbor points in the attention module on the registration results
View tableView in ArticleTable 6. Influence of number of K-nearest neighbor points in the attention module on the registration results
Methods 3DMatch FMR（%）↑ IR（%）↑ RR（%）↑ DGATNet v1 96.9 57.4 87.3 DGATNet v2 97.2 60.0 87.5 DGATNet v3 97.6 61.2 88.3 DGATNet（ours） 97.8 61.7 90.1 Methods 3DLoMatch FMR（%）↑ IR（%）↑ RR（%）↑ DGATNet v1 75.7 26.7 55.6 DGATNet v2 76.0 26.6 56.1 DGATNet v3 76.9 27.4 58.9 DGATNet（ours） 78.8 29.6 60.8 DGATNet v1： DGATNet v2： DGATNet v3： DGATNet（ours）：
Table 7. Ablation experiments of BNHN normalization technique
View tableView in ArticleTable 7. Ablation experiments of BNHN normalization technique
Methods 3DMatch（resolution-mismatch） FMR（%）↑ IR（%）↑ RR（%）↑ Baseline 8.2 1.6 5.6 Baseline+BNHN 85.2 30.5 75.2 Baseline+Attention 87.4 38.4 76.6 DGATNet 91.9 43.9 81.2
Table 8. Ablation experiments of each component in DGATNet
View tableView in ArticleTable 8. Ablation experiments of each component in DGATNet
Methods 3DMatch FMR（%）↑ IR（%）↑ RR（%）↑ Baseline 97.2 47.5 85.5 Baseline+BNHN 96.3 46.7 85.3 Baseline+Attention 97.5 60.0 88.3 Baseline+Attention+Position embedding 97.5 61.2 88.7 Baseline+Attention+Sin position embedding 97.8 61.7 90.1
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
Paper Information
Category: Information Sciences
Received: May. 12, 2022
Accepted: --
Published Online: Feb. 15, 2023
The Author Email: ZHANG Haibo (zhanghb@nwu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20