Deep Learning Point Cloud Classification Method Based on Fusion Graph Convolution

Tianye Xu; Haiyong Ding

doi:10.3788/LOP202259.0228005

Laser & Optoelectronics Progress, Volume. 59, Issue 2, 0228005(2022)

Deep Learning Point Cloud Classification Method Based on Fusion Graph Convolution

Tianye Xu and Haiyong Ding^*

School of Remote Sensing & Geomatics Engineering, Nanjing University of Information Science & Technology, Nanjing , Jiangsu 210044, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(15)

Fig. 1. Structure of the deep learning classification network

Download full size

Fig. 2. Principle of the spatial domain graph convolution

Download full size

Fig. 3. Schematic diagram of the KNN

Download full size

Fig. 4. Central node information aggregated under different orders. (a) q=1; (b) q=2

Download full size

Fig. 5. Training data and multispectral aerial image of corresponding region. (a) Point cloud; (b) aerial image

Download full size

Fig. 6. Testing data and multispectral aerial images of corresponding region. (a) Point cloud; (b) aerial image

Download full size

Fig. 7. Fusion result of point cloud data and spectral images. (a) Training data; (b) testing data

Download full size

Fig. 8. Schematic diagram of multiscale sampling

Download full size

Fig. 9. Testing data labels and classification results. (a) True label; (b) classification result of our method

Download full size

Fig. 10. Error map of the classification result

Download full size

Fig. 11. Error maps of classification results of different methods. (a) PointNet; (b) DGCNN; (c) PointNet++; (d) our method

Download full size

Table 1. Number of various classes points in Vaihingen data set
View table
Table 1. Number of various classes points in Vaihingen data set
Class p_l l_v i_s car f_h r_f f_e s_b t_e
Training data 546 180850 193723 4614 12070 152045 27250 47605 135173
Testing data 600 98690 101986 3708 7422 109048 11224 24818 54226

Table 2. Confusion matrix and evaluation indexes of the classification result

View table

Table 2. Confusion matrix and evaluation indexes of the classification result

Class	p_l	l_v	i_s	car	f_h	r_f	f_e	s_b	t_e
p_l	60.6	0.2	0.0	0.0	0.0	27.2	3.6	0.0	8.5
l_v	0.0	79.3	10.5	0.4	0.4	0.9	0.7	2.9	4.9
i_s	0.0	6.4	92.0	0.3	0.0	0.8	0.1	0.1	0.2
car	0.0	9.9	11.3	55.0	3.4	5.4	5.4	5.5	4.0
f_h	0.0	25.7	3.8	0.2	19.3	9.2	1.1	11.3	29.3
r_f	0.1	1.2	0.9	0.0	0.0	93.9	1.7	0.2	1.9
f_e	0.2	6.4	2.4	1.3	0.1	26.2	48.6	3.2	11.6
s_b	0.0	18.7	2.3	0.6	1.0	3.4	1.5	19.1	53.3
t_e	0.0	3.1	0.5	0.1	0.1	3.3	1.3	0.9	90.6
Precision	80.9	80.9	90.1	66.5	64.1	93.5	53.4	41.5	70.1
Recall	60.6	79.3	92.0	55.0	19.3	93.9	48.6	19.1	90.6
F1 score	69.3	80.1	91.0	60.2	29.6	93.7	50.9	26.1	79.1

Table 3. Quantitative evaluation index of different methods

View table

Table 3. Quantitative evaluation index of different methods

Method	p_l	l_v	i_s	car	f_h	r_f	f_e	s_b	t_e	OA	F1 score
PointNet	1.7	68.9	77.6	30.6	16.7	69.1	6.8	33.3	46.5	60.9	33.6
DGCNN	80.3	81.2	78.9	26.8	52.0	67.1	18.9	31.1	79.5	70.6	41.6
PointNet++	71.5	72.0	90.0	68.3	16.7	74.0	38.4	39.2	52.7	73.6	50.1
Ours	80.9	80.9	90.1	66.5	64.1	93.5	53.4	41.5	70.1	84.3	64.4

Table 4. Classification results of our method and existing methods

View table

Table 4. Classification results of our method and existing methods

Method	p_l	l_v	i_s	car	f_h	r_f	f_e	s_b	t_e	OA	F1 score
UM	46.1	79.0	89.1	47.7	5.2	92.0	52.7	40.9	77.9	80.8	59.0
WhuY3	37.1	81.4	90.1	63.4	23.9	93.4	47.5	39.9	78.0	82.3	61.6
BIJ_W	13.8	78.5	90.5	56.4	36.3	92.2	53.2	43.3	78.4	81.5	60.3
RIT_1	37.5	77.9	91.5	73.4	18.0	94.0	49.3	45.9	82.5	81.6	63.3
Ref.［21］	68.4	80.2	91.4	78.1	37.0	93.0	60.5	46.0	79.4	82.2	70.7
Ours	69.3	80.1	91.0	60.2	29.6	93.7	50.9	26.1	79.1	84.3	64.4

Tools

Get Citation

Copy Citation Text

Tianye Xu, Haiyong Ding. Deep Learning Point Cloud Classification Method Based on Fusion Graph Convolution[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0228005

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites