Optics and Precision Engineering, Volume. 32, Issue 18, 2823(2024)
3D point cloud classification and segmentation based on dual attention and weighted dynamic graph convolution
Figures & Tables(18)
Fig. 1. Point cloud classification and segmentation networks based on DAM and WEM
Fig. 8. Comparison of the impact of
Fig. 9. Comparison of the impact of
Fig. 11. Comparison of the impact of
Table 1. Main parameter settings of the experiment
View tableView in ArticleTable 1. Main parameter settings of the experiment
Dataset Points Optimizer Learning rate Cosine Annealing Batch size Test batch size Epochs ModelNet40 1 024 AdamW 0.001 1×10-5 16 8 250 ShapeNet Part 2 048 AdamW 0.001 1.5×10-5 8 4 200
Table 2. Comparison of 3d point cloud classification results on modelnet40 dataset
View tableView in ArticleTable 2. Comparison of 3d point cloud classification results on modelnet40 dataset
Methods Input Type OA/% mAcc/% MVCNN［ 1 ］Image 90.1 — VoxNet［ 2 ］Voxel 85.9 — PointNet［ 5 ］Points（1k） 89.2 86.2 PointNet++［ 6 ］Points（5k）+Normal 91.7 89.5 PointWeb［ 11 ］Points（1k）+Normal 92.3 89.4 Point-Transformer［ 21 ］Points（1k） 93.7 — PointConv［ 23 ］Points（1k）+Normal 92.4 — PointASNL［ 24 ］Points（1k）+Normal 93.2 — SpiderCNN［ 25 ］Points（1k）+Normal 92.4 — GCN3D［ 20 ］Points（0.768k） 93.0 90.3 RF-DGCNN［ 22 ］Points（1k） 93.62 90.86 KPConv［ 9 ］Points（1k） 92.9 — DGCNN［ 7 ］Points（1k） 92.9 90.2 Linked-DGCNN［ 12 ］Points（3k） 92.6 89.7 Ours Points（1k） 93.36 90.74
Table 3. Comparison of the results for the parameters and forward time of different classification networks
View tableView in ArticleTable 3. Comparison of the results for the parameters and forward time of different classification networks
Methods Parameters/M Forward Time/ms OA/% PointNet［ 5 ］3.48 3.2 89.2 PointNet++［ 6 ］1.48 4.8 91.7 DGCNN［ 7 ］1.84 6.3 92.9 Point-Transformer［ 21 ］1.73 136.4 93.7 Ours 2.08 8.3 93.36
Table 4. Comparison of 3d point cloud segmentation results on shapenet part dataset
View tableView in ArticleTable 4. Comparison of 3d point cloud segmentation results on shapenet part dataset
Methods mIoU/% ariplane Bag Cap Car Chair Earphone Guitar Knife PointNet［ 5 ］83.7 83.4 78.7 82.5 74.9 89.6 73.0 91.5 85.9 PointNet++［ 6 ］85.1 82.4 79.0 87.7 77.3 90.8 71.8 91.0 85.9 PointConv［ 21 ］85.7 — — — — — — — — PointASNL［ 22 ］86.1 84.1 84.7 87.9 79.7 92.2 73.7 91.0 87.2 GCN3D［ 24 ］85.2 83.3 78.8 83.2 77.5 90.8 75.8 90.8 86.0 DTNet［ 26 ］85.6 83.0 81.4 84.3 78.4 90.9 74.3 91.0 87.3 3D-GCN［ 27 ］85.1 83.1 84.0 86.6 77.5 90.3 74.1 90.9 86.4 DGCNN［ 7 ］85.2 84.0 83.4 86.7 77.8 90.6 74.7 91.2 87.5 Linked-DGCNN［ 12 ］84.7 — 85.4 88.6 — 90.8 — 90.9 88.6 Ours 85.96 86.5 86.6 89.4 80.4 91.7 81.3 91.9 88.4
Table 4. Comparison of 3d point cloud segmentation results on shapenet part dataset
View tableView in ArticleTable 4. Comparison of 3d point cloud segmentation results on shapenet part dataset
Methods Lamp Laptop Motorbike Mug Pistol Rocket Skateboard Table PointNet［ 5 ］80.8 95.3 65.2 93.0 81.2 57.9 72.8 80.6 PointNet++［ 6 ］83.7 95.3 71.6 94.1 81.3 58.7 76.4 82.6 PointConv［ 21 ］— — — — — — — — PointASNL［ 22 ］84.2 95.8 74.4 95.2 81.0 63.0 76.3 83.2 GCN3D［ 24 ］83.8 95.2 66.4 93.4 81.3 51.2 67.1 83.0 DTNet［ 26 ］84.7 95.6 69.0 94.4 82.5 59.0 76.4 83.5 3D-GCN［ 27 ］83.8 95.6 66.8 94.8 81.3 59.6 75.7 82.8 DGCNN［ 7 ］82.8 95.7 66.3 94.9 81.1 63.5 74.5 82.6 Linked-DGCNN［ 12 ］82.4 95.4 — — — — 74.7 81.9 Ours 84.5 96.4 72.9 96.0 85.4 63.6 78.8 82.9
Table 5. Validation of the effectiveness of each module in point cloud classification task
View tableView in ArticleTable 5. Validation of the effectiveness of each module in point cloud classification task
Nets DAM ECC/
KNN/WKNN
TopK OA/ % mAcc/ % A × KNN × 91.45 87.68 B √ KNN × 92.10 88.82 C √ ECC+KNN × 92.18 88.83 D √ ECC+KNN √ 92.30 89.31 E √ WKNN × 92.42 89.25 F √ WKNN √ 92.63 88.67 G √ ECC+WKNN × 92.91 89.89 H √ ECC+WKNN √ 93.36 90.74
Table 6. Validation of the effectiveness of each module in point cloud segmentation task
View tableView in ArticleTable 6. Validation of the effectiveness of each module in point cloud segmentation task
DAM ECC+WKNN/KNN OA/% mIoU/% × KNN 94.03 84.65 √ KNN 94.13 84.94 √ ECC+KNN 94.28 85.29 √ WKNN 94.25 85.32 √ ECC+WKNN 94.56 85.96
Tools
Get Citation
Copy Citation Text
Jian XIAO, Xiaohong WANG, Wei LI, Yifei YANG, Ji LUO. 3D point cloud classification and segmentation based on dual attention and weighted dynamic graph convolution[J]. Optics and Precision Engineering, 2024, 32(18): 2823
Paper Information
Category:
Received: Jan. 25, 2024
Accepted: --
Published Online: Nov. 18, 2024
The Author Email: WANG Xiaohong (xhwang@gzu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20