Point Cloud Classification and Segmentation Network Based on Dynamic Graph Convolutional Network

Jiang’an Wang; Jiao He; Dawei Pang

doi:10.3788/LOP202158.1215008

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 12 >Page 1215008 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 12, 1215008 (2021)

Point Cloud Classification and Segmentation Network Based on Dynamic Graph Convolutional Network

Jiang’an Wang^*, Jiao He, and Dawei Pang

Author Affiliations

School of Information Engineering, Chang’an University, Xi’an, Shaanxi 710064, China

show less

DOI: 10.3788/LOP202158.1215008 Cite this Article Set citation alerts

Jiang’an Wang, Jiao He, Dawei Pang. Point Cloud Classification and Segmentation Network Based on Dynamic Graph Convolutional Network[J]. Laser & Optoelectronics Progress, 2021, 58(12): 1215008 Copy Citation Text

show less

Fig. 1. Local digraph of point cloud

Download full size

Fig. 2. Linked-DGCNN network structure

Download full size

Fig. 3. Classification test accuracy curve

Download full size

Fig. 4. Training loss curve

Download full size

Fig. 5. Comparison of the segmentation results of the algorithm in this paper and DGCNN for object components in the ShapeNet dataset. (a)Ground truth; (b)DGCNN; (c)DGCNN_diff; (d)Linked-DGCNN; (e)Linked-DGCNN_diff

Download full size

Fig. 6. Influence of number of EdgeConv convolution layers on accuracy

Download full size

Fig. 7. Influence of K value on accuracy

Download full size

Fig. 8. Effect of number of different sampling points on classification accuracy. (a) DGCNN; (b) Linked-DGCNN

Download full size

Method	Input(size)	Model size /MB	Forward time /ms	Mean class accuracy /%	Accuracy overall /%
3D ShapeNets	Voxels(1)	—	180.2	77.5	84.5
VoxNet	Voxels(12)	—	87.0	82.8	85.5
Subvolumes^[17]	Voxels(20)	190.8	120.3	86.1	89.3
MVCNN	Views(80)	690	163.2	88.8	—
PointNet	Points(1024×3)	40	25.3	84.3	87.8
PointNet++	Points+normal(5000×6)	24	44.3	—	90.5
DGCNN	Points(1024×3)	21.2	98.03	89.0	91.0
Ours	Points(1024×3)	12.4	88.5	89.7	92.6

Table 1. Results of dataset point cloud classification of ModelNet40

Method	mIoU	Bag	Cat	Chair	Guitar	Knife	Laptop	Lamp	Board	Table
PointNet	83.6	78.6	82.4	89.5	91.5	85.8	80.7	95.2	72.7	80.5
PointNet++	85.0	78.9	87.6	90.7	90.9	85.8	83.6	95.2	76.3	82.5
SO-Net	84.6	83.5	84.8	90.8	90.1	83.6	82.3	95.1	72.0	82.5
DGCNN	85.0	83.6	84.3	90.6	91.4	87.2	82.8	96.0	75.4	81.8
Ours	84.7	85.4	88.6	90.8	90.9	88.6	82.4	95.4	74.7	81.9

Table 2. Segmentation accuracy of object components in ShapeNet dataset unit: %

Jiang’an Wang, Jiao He, Dawei Pang. Point Cloud Classification and Segmentation Network Based on Dynamic Graph Convolutional Network[J]. Laser & Optoelectronics Progress, 2021, 58(12): 1215008

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information