Mobile Laser Scanning Point Cloud Classification Based on Data Augmentation and Mask Learning

Xiangda Lei; Haiyan Guan; Ke Chen; Nannan Qin; Yufu Zang

doi:10.3788/CJL231396

Chinese Journal of Lasers, Volume. 51, Issue 13, 1310004(2024)

Mobile Laser Scanning Point Cloud Classification Based on Data Augmentation and Mask Learning

Xiangda Lei¹, Haiyan Guan^1,2、*, Ke Chen¹, Nannan Qin¹, and Yufu Zang¹

¹School of Remote Sensing & Geomatics Engineering, Nanjing University of Information Science & Technology, Nanjing 210044, Jiangsu , China

²Technology Innovation Center for Integrated Applications in Remote Sensing and Navigation, Ministry of Natural Resources, Nanjing 210044, Jiangsu , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Flow chart of point cloud classification

Download full size

Fig. 2. EC-Mix3D data augmentation strategy

Download full size

Fig. 3. RandLA-Net structure

Download full size

Fig. 4. Toronto3D dataset

Download full size

Fig. 5. Paris dataset

Download full size

Fig. 6. Experimental results on Toronto3D dataset. (a) Overhead classification result; (b) side-view classification result; (c) overhead misclassification result; (d) side-view misclassification result

Download full size

Fig. 7. Experimental results on Paris dataset. (a) Soufflot0 classification result; (b) Soufflot3 classification result; (c) Soufflot0 misclassification result; (d) Soufflot3 misclassification result

Download full size

Fig. 8. Comparison of model complexity. (a) Comparison of Parameters; (b) comparison of FLOPs

Download full size

Table 1. Quantitative classification results of different deep learning methods on Toronto3D dataset

View table

Table 1. Quantitative classification results of different deep learning methods on Toronto3D dataset

Method	OA	mIoU	IoU
Method	OA	mIoU	Road	Road marking	Natural	Building	Utility line	Pole	Car	Fence
MappingC	94.7	82.9	97.2	67.9	97.6	93.8	86.9	82.1	93.7	44.1
RandLA-Net	94.4	81.8	96.7	64.2	96.9	94.2	88.0	77.8	93.4	42.9
ResDLPS-Net	96.5	80.3	95.8	59.8	96.1	90.9	86.8	79.9	89.4	43.3
BAF-LAC	95.2	82.2	96.6	64.7	96.4	92.8	86.1	83.9	93.7	43.5
BAAF-Net	94.2	81.2	96.8	67.3	96.8	92.2	86.8	82.3	93.1	34.0
LACV-Net	97.4	82.7	97.1	66.9	97.3	93.0	87.3	83.4	93.4	43.1
RFCR	97.3	81.2	96.8	65.3	97.1	93.6	86.5	81.7	94.0	34.4
Baseline	95.9	77.5	95.0	53.5	96.3	92.0	86.3	77.9	90.9	28.0
Ours	97.7	83.8	97.4	70.5	97.2	93.2	87.4	83.3	93.9	47.7

Table 2. Quantitative classification results of different deep learning methods on Paris dataset

View table

Table 2. Quantitative classification results of different deep learning methods on Paris dataset

Category		Baseline	BAF-LAC	BAAF-Net	RFCR	SCF-Net	Ours
IoU	Building	88.30	86.18	85.92	88.86	88.51	88.64
	Fence	27.43	43.91	41.71	22.58	41.88	32.01
	Other	34.73	29.13	33.22	32.09	36.13	29.52
	Pedestrian	78.15	69.01	70.99	70.47	78.01	83.59
	Pole	59.68	61.66	55.19	63.51	69.46	74.09
	RoadLine	67.10	74.26	69.88	76.05	69.67	74.15
	Road	93.31	94.40	92.83	94.80	94.02	95.55
	SideWalk	82.13	81.61	77.70	81.31	84.33	87.08
	Vegetation	92.64	93.86	92.29	93.89	94.80	95.33
	Vehicles	93.96	91.78	92.89	93.54	93.96	96.37
	TrafficSign	61.35	60.18	48.65	55.96	58.25	59.39
	TrafficLight	44.26	54.91	38.64	35.25	49.44	67.31
	Static	40.26	44.92	39.55	46.92	60.53	46.64
	Dynamic	22.78	21.98	23.84	26.63	27.06	32.67
OA		91.68	91.84	90.88	92.52	92.74	93.20
mIoU		63.29	64.84	61.66	62.99	67.58	68.74

Table 3. Ablation experiment results

View table

Table 3. Ablation experiment results

Method	EC-Mix3D data augmentation	Mask learning framework			OA /%	mIoU /%
Method	EC-Mix3D data augmentation	Mask supervision	Consistency constraint	Error prediction entropy maximization	OA /%	mIoU /%
Baseline					95.85	77.48
Method A	√				97.01	80.86
Method B	√	√			97.29	82.07
Method C	√	√	√		97.38	82.41
Method D	√			√	97.33	82.81
Method E		√	√	√	97.37	80.97
Ours	√	√	√	√	97.65	83.82

Table 4. Experimental results of different mask block sizes
View table
Table 4. Experimental results of different mask block sizes
Mask block size/m OA /% mIoU /%
0.5 97.53 82.40
1.0 97.65 83.82
1.5 97.77 82.91

Table 5. Experimental results of different mask ratios
View table
Table 5. Experimental results of different mask ratios
Mask ratio OA /% mIoU /%
2/3 97.57 83.12
1/2 97.62 83.50
1/3 97.65 83.82
1/4 97.64 83.73

Table 6. Experimental results of different data augmentation strategies
View table
Table 6. Experimental results of different data augmentation strategies
Data augmentation strategy OA /% mIoU /%
Global data augmentation 97.37 81.17
Mix3D 97.58 82.59
PolarMix 97.43 81.62
EC-Mix3D 97.65 83.82

Tools

Get Citation

Copy Citation Text

Xiangda Lei, Haiyan Guan, Ke Chen, Nannan Qin, Yufu Zang. Mobile Laser Scanning Point Cloud Classification Based on Data Augmentation and Mask Learning[J]. Chinese Journal of Lasers, 2024, 51(13): 1310004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites