LiDAR Data Classification Based on Dilated Convolution Capsule Network

Aili Wang; Yuxiao Zhang; Haibin Wu; Kaiyuan Jiang; Yuji Iwahori

doi:10.3788/CJL202148.1110003

Chinese Journal of Lasers, Volume. 48, Issue 11, 1110003(2021)

LiDAR Data Classification Based on Dilated Convolution Capsule Network

Aili Wang¹, Yuxiao Zhang¹, Haibin Wu^1、*, Kaiyuan Jiang¹, and Yuji Iwahori²

¹Heilongjiang Province Key Laboratory of Laser Spectroscopy Technology and Application, Harbin University of Science and Technology, Harbin, Heilongjiang 150080, China

²Department of Computer Science, Chubu University, Aichi 487- 8501, Japan

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. Structure of dilated convolution capsule network

Download full size

Fig. 2. Residual block structure

Download full size

Fig. 3. Dilated convolution with different dilation rates. (a) r=1; (b) r=2; (c) r=3

Download full size

Fig. 4. Odd-even mixed dilation rates. (a) r=5; (b) r=2; (c) r=1

Download full size

Fig. 5. Capsule network structure

Download full size

Fig. 6. Capsule structure

Download full size

Fig. 7. DSM and groundtruth map of Bayview Park dataset. (a) DSM; (b) groundtruth map

Download full size

Fig. 8. DSM and groundtruth map of Recology dataset. (a) DSM; (b) groundtruth map

Download full size

Fig. 9. Dilation rate distribution of different datasets. (a) Bayview Park dataset; (b) Recology dataset

Download full size

Fig. 10. Classification results of Bayview Park dataset. (a) Groundtruth map; (b) SVM; (c) Random Forest; (d) CNN; (e) CapsNet; (f) ResNet; (g) Dilated-ResNet; (h) ResCapsNet; (i) DCCN

Download full size

Fig. 11. Classification results of Recology dataset. (a) Groundtruth map; (b) SVM; (c) Random Forest; (d) CNN; (e) CapsNet; (f) ResNet; (g) Dilated-ResNet; (h) ResCapsNet; (i) DCCN

Download full size

Table 1. ResNet-34 model parameters

View table

Table 1. ResNet-34 model parameters

Layer name	Output size	ResNet-34 parameter setting
Conv 1	18×18	3×3, 16, stride 1
Conv2_x	18×18	3×3 max pooling, stride 2 $(\begin{array}{l} 3 \times 3 \\ 3 \times 3 \end{array} \begin{matrix} , 16 \end{matrix})$ ×3
Conv3_x	18×18	$(\begin{array}{l} 3 \times 3 \\ 3 \times 3 \end{array} \begin{matrix} , 28 \end{matrix})$ ×4
Conv4_x,distribution ofdilation rate is[1,2,5]	18×18	$(\begin{array}{l} 3 \times 3 \\ 3 \times 3 \end{array} \begin{matrix} , 40 \end{matrix})$ ×6
Conv5_x,distribution ofdilation rate is[1,2,5]	18×18	$(\begin{array}{l} 3 \times 3 \\ 3 \times 3 \end{array} \begin{matrix} , 52 \end{matrix})$ ×3
Output	9×9	average pooling, stride 2

Table 2. Classification results of different training samples on Bayview Park dataset

View table

Table 2. Classification results of different training samples on Bayview Park dataset

Training sample	Index	400	500	600	700
	OA /%	72.30±2.07	75.27±1.36	75.36±2.02	76.75±0.76
SVM	AA /%	76.98±1.42	78.50±2.10	78.86±1.12	81.20±2.27
	100K	65.22±1.79	66.36±0.95	67.65±1.72	69.12±2.23
	OA /%	86.56±0.75	87.34±0.62	88.23±0.34	90.56±0.47
Random Forest	AA /%	88.94±1.52	89.12±0.23	89.45±0.38	90.14±0.73
	100K	82.66±0.24	83.61±0.38	84.26±0.39	86.73±0.64
	OA /%	87.23±2.01	88.12±0.96	88.52±0.43	90.72±1.69
CNN	AA /%	88.70±1.13	89.63±2.71	89.96±1.65	90.23±0.68
	100K	83.26±1.45	85.25±1.46	86.23±1.78	86.72±2.34
	OA /%	85.43±1.12	87.25±0.89	90.07±1.08	90.73±0.36
CapsNet	AA /%	84.26±1.78	88.26±1.38	91.14±1.35	91.86±1.52
	100K	81.21±0.81	83.19±0.69	86.81±1.45	86.92±1.22
	OA /%	90.25±1.73	92.16±1.26	93.26±1.21	94.59±1.20
ResNet	AA /%	91.53±1.38	93.23±0.81	94.25±1.06	95.86±1.25
	100K	87.15±1.49	89.46±1.48	91.26±1.77	93.49±1.28
	OA /%	91.32±0.45	93.16±0.75	93.89±0.34	95.84±1.25
Dilated-ResNet	AA /%	92.67±0.76	94.09±1.06	95.27±0.87	96.42±1.08
	100K	88.46±1.05	91.58±0.57	92.65±0.65	94.15±1.34
	OA /%	93.15±0.52	94.79±0.41	94.59±0.73	96.42±0.71
ResCapsNet	AA /%	94.27±0.43	95.42±0.90	96.03±0.75	97.01±1.07
	100K	90.49±1.00	92.48±0.47	93.23±0.75	94.99±1.17
	OA /%	93.48±0.39	94.51±047	95.45±0.60	97.07±0.54
DCCN	AA /%	94.97±0.44	95.39±0.61	95.90±0.74	97.70±0.20
	100K	91.36±0.57	92.79±0.42	94.02±0.72	96.14±0.71

Table 3. Classification results of different training samples on Recology dataset

View table

Table 3. Classification results of different training samples on Recology dataset

Training sample	Index	400	500	600	700
	OA /%	72.68±1.89	76.83±0.26	76.94±2.25	77.25±0.86
SVM	AA /%	77.20±1.22	78.69±1.87	78.73±1.08	81.28±2.33
	100K	67.22±1.79	68.42±0.98	68.82±1.59	69.78±2.24
	OA /%	85.24±1.25	87.33±0.74	89.20±1.88	91.75±1.00
Random Forest	AA /%	88.36±1.92	89.85±3.06	90.27±1.26	91.29±1.41
	100K	82.25±0.88	86.26±1.57	86.59±1.97	89.22±1.34
	OA /%	86.26±1.48	88.24±0.98	90.52±0.68	92.73±1.86
CNN	AA /%	89.16±2.84	90.15±0.29	90.67±1.24	92.44±2.34
	100K	83.18±1.59	86.78±0.67	86.83±0.88	90.15±2.17
	OA /%	80.73±1.07	84.92±1.67	86.75±0.43	90.26±1.24
CapsNet	AA /%	81.93±1.94	86.29±1.08	86.95±1.09	91.26±1.99
	100K	76.79±1.68	81.67±0.85	83.92±0.76	88.37±1.52
	OA /%	90.58±1.92	93.56±1.42	95.57±0.76	95.83±0.99
ResNet	AA /%	88.86±2.14	94.52±1.22	94.33±1.32	95.36±1.88
	100K	88.84±2.34	93.04±1.70	94.98±0.84	95.24±1.27
	OA /%	92.07±0.98	93.75±0.34	94.87±0.54	95.89±0.76
Dilated-ResNet	AA /%	93.76±0.67	94.88±0.96	95.98±1.25	96.34±0.38
	100K	90.67±0.34	92.39±1.07	94.77±1.15	95.09±0.47
	OA /%	93.44±1.21	94.35±1.23	96.07±0.48	96.31±0.73
ResCapsNet	AA /%	94.63±0.17	95.22±060	97.16±1.15	97.30±0.17
	100K	91.29±0.70	93.79±0.99	95.33±0.63	95.43±0.31
	OA /%	94.01±0.36	94.99±0.96	96.42±0.63	96.98±0.76
DCCN	AA /%	94.97±0.58	95.67±0.46	97.49±0.69	97.77±0.78
	100K	93.28±0.34	94.35±0.99	96.06±0.70	96.41±0.90

Table 4. Classification results of each class for 700 samples on Bayview Park dataset unit:%

View table

Table 4. Classification results of each class for 700 samples on Bayview Park dataset unit:%

Classe	SVM	RandomForest	CNN	CapsNet	ResNet	Dilated-ResNet	ResCapsNet	DCCN
Building1	82.20±3.72	95.24±3.75	93.53±1.50	94.28±1.45	98.26±1.55	98.96±1.04	99.36±0.64	99.74±0.26
Building2	84.59±2.58	98.83±1.03	92.72±1.08	95.22±2.03	99.64±0.36	99.78±0.22	99.79±0.21	99.54±0.46
Building3	91.36±4.96	100	92.90±1.50	93.32±1.93	99.57±0.43	99.68±0.32	100	99.69±0.31
Road	81.60±4.43	82.59±6.35	91.36±1.58	94.90±1.20	96.38±2.65	97.09±2.83	98.12±1.22	98.82±1.18
Trees	83.78±1.69	90.43±1.20	86.79±1.82	92.82±1.79	97.65±0.82	98.06±0.76	98.67±0.88	98.39±0.36
Soil	62.23±2.27	86.78±0.52	85.59±1.71	83.56±0.80	86.95±1.62	87.47±1.52	89.32±2.06	90.98±1.39
Seawater	86.79±2.48	84.21±1.25	90.75±2.74	85.53±1.23	91.18±2.87	91.88±2.65	93.47±2.30	94.79±1.17

Table 5. Classification results of each class for 700 samples on Recology dataset unit:%

View table

Table 5. Classification results of each class for 700 samples on Recology dataset unit:%

Classe	SVM	RandomForest	CNN	CapsNet	ResNet	Dilated-ResNet	ResCapsNet	DCCN
Building1	71.78±1.21	92.56±2.78	98.78±1.22	92.29±1.29	98.75±1.25	98.86±1.14	98.06±1.74	98.97±0.32
Building2	64.54±1.77	94.67±3.75	96.52±1.42	94.45±1.36	99.24±0.76	99.02±0.98	99.46±0.54	99.82±0.18
Building3	92.78±1.20	94.24±1.44	94.16±1.25	93.47±1.35	98.42±1.58	98.56±1.44	98.41±1.16	98.04±0.96
Building4	90.39±2.47	97.88±0.24	97.55±1.36	95.16±1.01	95.41±1.72	97.52±2.31	99.67±0.33	98.24±1.76
Building5	86.26±1.73	96.30±2.77	97.48±2.38	98.06±1.94	99.86±0.14	99.07±0.93	98.72±1.28	99.55±0.45
Building6	71.52±1.62	95.27±1.36	94.37±1.07	87.47±1.59	96.85±2.34	97.26±1.09	98.35±1.65	97.06±2.83
Building7	88.37±2.74	97.26±2.74	97.48±1.95	95.86±2.15	92.43±2.53	95.97±1.75	98.79±1.21	99.51±0.49
Trees	86.88±1.24	95.67±0.21	95.64±1.24	90.14±0.35	97.48±1.85	96.47±1.66	95.54±1.31	97.32±1.55
Parking Lot	62.77±1.92	76.21±0.18	77.86±1.87	83.39±0.14	89.45±1.75	89.29±1.06	89.28±1.27	91.31±0.58
Soil	81.83±3.22	73.16±0.32	73.29±1.72	75.41±1.45	88.69±2.47	90.66±2.54	95.68±2.43	96.92±3.08
Grass	97.77±1.37	98.26±1.25	97.26±1.58	98.26±1.42	92.54±2.46	93.48±2.49	95.52±2.61	99.61±0.39

Table 6. Comparison of calculation time for 700 samples on Recology dataset and Bayview Park dataset

View table

Table 6. Comparison of calculation time for 700 samples on Recology dataset and Bayview Park dataset

Dataset	Network	Traintime /s	Testtime /s	OA /%
	ResNet	125.15	1.82	94.51
	Dilated-ResNet	195.34	2.76	95.67
BayviewPark	CapsNet	99.58	2.18	90.73
	ResCapsNet	343.85	3.00	96.42
	DCCN	518.65	3.75	97.07
	ResNet	196.34	2.67	95.83
	Dilated-ResNet	257.98	2.98	96.36
Recology	CapsNet	94.27	1.39	90.26
	ResCapsNet	428.69	3.34	96.31
	DCCN	586.55	3.96	96.98

Tools

Get Citation

Copy Citation Text

Aili Wang, Yuxiao Zhang, Haibin Wu, Kaiyuan Jiang, Yuji Iwahori. LiDAR Data Classification Based on Dilated Convolution Capsule Network[J]. Chinese Journal of Lasers, 2021, 48(11): 1110003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: remote sensing and sensor

Received: Nov. 17, 2020

Accepted: Jan. 4, 2021

Published Online: Jun. 4, 2021

The Author Email: Wu Haibin (woo@hrbust.edu.cn)

DOI:10.3788/CJL202148.1110003

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology