Point‑Voxel Consistency Constraint Network for LiDAR Point Cloud Classification Under Urban Scenes

Fig. 4. Overall visualization of the classification results on the Toronto3D dataset. (a) Point cloud (RGB); (b) ground truth; (c) PVCC-Net; (d) RandLA-Net; (e) PVCNN

Download full size

Fig. 5. Local visualization of classification results on the Toronto3D dataset. (a) Point cloud (RGB); (b) ground truth; (c) PVCC-Net; (d) RandLA-Net; (e) PVCNN

Download full size

Fig. 6. OA and mIoU of different methods on the Semantic3D dataset. (a) OA; (b) mIoU

Download full size

Fig. 7. Visualization of classification results on the Semantic3D dataset. (a) Point cloud (RGB); (b) PVCC-Net; (c) RandLA-Net;

Download full size

Fig. 8. Normalized confusion matrix of classification results on the Semantic3D dataset

Download full size

Fig. 9. OA and mIoU of different methods on the SensatUrban dataset. (a) OA; (b) mIoU

Download full size

Fig. 10. Visualization of classification results on the SensatUrban dataset. (a) Point cloud (RGB); (b) PVCC-Net; (c) RandLA-Net;

Download full size

Table 1. Introduction of point cloud classification dataset in urban scenes

View table

Table 1. Introduction of point cloud classification dataset in urban scenes

Dataset	Region	Spatial size	Number of points/10⁶	RGB	Number of classes	Sensor
Toronto3D	Toronto， Canada	1×10³ m	78.3	Yes	8	MLS
Semantic3D	Central Europe	3.84×10⁴ m²	4000	Yes	8	TLS
SensatUrban	Birmingham， Cambridge and York， UK	7.64×10⁶ m²	2847	Yes	13	ALS

Table 2. Comparison of classification results of different methods on Toronto3D dataset

View table

Table 2. Comparison of classification results of different methods on Toronto3D dataset

Method	OA	mIoU	IoUs
Method	OA	mIoU	road	road mark	natural	building	util. line	pole	car	fence
ResDLPS-Net	96.49	80.27	95.82	59.80	96.10	90.96	86.82	79.95	89.41	43.31
BAAF-Net	94.20	81.20	96.80	67.30	96.80	92.20	86.80	82.30	93.10	34.00
BAF-LAC	95.20	82.20	96.60	64.70	96.40	92.80	86.10	83.90	93.70	43.50
MFA	97.00	79.90	96.80	70.00	96.10	92.30	86.30	80.40	91.50	29.40
LACV-Net	97.40	82.70	97.10	66.90	97.30	93.00	87.30	83.40	93.40	43.10
RandLA-Net	94.37	81.77	96.69	64.21	96.92	94.24	88.06	77.84	93.37	42.86
MappingConvSeg	94.72	82.89	97.15	67.87	97.55	93.75	86.88	82.12	93.72	44.11
PVCNN*	97.36	78.56	97.68	71.31	94.81	91.93	81.87	77.42	90.56	22.90
Baseline*	96.44	79.48	95.90	60.59	95.76	91.31	86.05	79.13	89.60	37.53
PVCC-Net*	97.97	82.92	97.87	74.95	97.00	93.91	86.65	81.71	94.86	36.42

Table 3. Comparison of classification results of different methods on Semantic3D dataset

View table

Table 3. Comparison of classification results of different methods on Semantic3D dataset

Method	IoUs /%
Method	man-made.	natural.	high veg.	low veg.	building	hardscape	scanning art.	car
SnapNet	82.00	77.30	79.70	22.90	91.10	18.40	37.30	64.40
DeepVoxNet	82.70	53.10	83.80	28.70	89.90	23.60	29.80	65.00
MSDeepVoxNet	83.00	67.20	83.90	36.70	92.40	31.30	50.00	78.20
OctNet	90.70	82.00	82.40	39.30	90.00	10.90	31.20	46.00
SEGCloud	83.90	66.00	86.00	40.50	91.10	30.90	27.50	64.30
RFMSSF	87.60	80.30	81.80	36.40	92.20	24.10	42.60	56.60
ShellNet	96.30	90.40	83.90	41.00	94.20	34.70	43.90	70.20
GACNet	86.40	77.70	88.50	60.60	94.20	37.30	43.50	77.80
SPGraph	97.40	92.60	87.90	44.00	93.20	31.00	63.50	76.20
KPConv	90.90	82.20	84.20	47.90	94.90	40.00	77.30	79.70
RGNet	97.50	93.00	88.10	48.10	94.60	36.20	72.00	68.00
ResDLPS-Net	95.60	90.70	89.20	53.40	94.70	50.80	58.90	78.60
SCF-Net	97.10	91.80	86.30	51.20	95.30	50.50	67.90	80.70
RandLA-Net	95.60	91.40	86.60	51.50	95.70	51.50	69.80	76.80
PVCNN*	88.90	69.50	78.10	27.90	94.20	29.10	37.40	66.60
Baseline*	97.30	90.30	87.20	47.80	94.00	44.80	57.00	79.60
PVCC-Net*	95.60	87.50	85.40	51.50	94.30	42.80	67.60	80.90

Table 4. Comparison of classification results of different methods on SensatUrban dataset

View table

Table 4. Comparison of classification results of different methods on SensatUrban dataset

Method	IoUs /%
Method	ground	veg.	build.	wall	bridge	park.	rail	traffic.	street.	car	foot.	bike	water
PointNet	67.96	89.52	80.05	0.00	0.00	3.95	0.00	31.55	0.00	35.14	0.00	0.00	0.00
PointNet++	72.46	94.24	84.77	2.72	2.09	25.79	0.00	31.54	11.42	38.84	7.12	0.00	56.93
SPGraph	69.93	94.55	88.87	32.83	12.58	15.77	15.48	30.63	22.96	56.42	0.54	0.00	44.24
KPConv	87.10	98.91	95.33	74.40	28.69	41.38	0.00	55.99	54.43	85.67	40.39	0.00	86.30
BAF-LAC	84.40	98.40	94.10	57.20	27.60	42.50	15.00	51.60	39.50	78.10	40.10	0.00	75.20
SCF-Net	83.20	98.10	93.80	56.00	52.40	47.40	11.10	53.20	38.90	79.60	37.60	0.00	64.70
RandLA-Net	80.11	98.07	91.58	48.88	40.75	51.62	0.00	56.67	33.23	80.14	32.63	0.00	71.31
TagentConv	71.54	91.38	75.90	35.22	0.00	45.34	0.00	26.69	19.24	67.58	0.01	0.00	0.00
SparseConv	74.10	97.90	94.20	63.30	7.50	24.20	0.00	30.10	34.00	74.40	0.00	0.00	54.80
PVCNN*	80.00	97.70	92.70	34.60	0.00	0.00	0.00	40.60	10.00	67.00	13.10	0.00	17.70
Baseline*	82.30	97.60	93.10	56.80	40.60	46.40	0.00	53.10	36.60	79.50	35.40	0.00	68.70
PVCC-Net*	84.90	98.50	95.50	62.40	24.00	55.40	0.00	58.70	46.80	81.30	33.80	0.00	79.20

Table 5. Classification results of different modules on Toronto3D dataset

View table

Table 5. Classification results of different modules on Toronto3D dataset

Method	Voxel	PV-SA	PV-CC	Lovasz	OA	mIoU	IoUs
Method	Voxel	PV-SA	PV-CC	Lovasz	OA	mIoU	road	road m.	natural	build.	util. l.	pole	car	fence
A					97.72	81.06	97.56	70.75	96.73	92.48	84.24	79.61	94.18	32.95
B	√				97.90	81.43	97.97	75.71	96.35	93.74	85.57	84.04	93.98	24.10
C	√	√			97.98	82.37	97.93	75.83	97.12	93.43	85.70	83.94	94.01	31.01
D	√	√	√		97.95	82.67	97.86	74.99	97.04	93.65	86.77	82.03	94.63	34.38
PVCC-Net	√	√	√	√	97.97	82.92	97.87	74.95	97.00	93.91	86.65	81.71	94.86	36.42

Table 6. Model complexity comparison results

View table

Table 6. Model complexity comparison results

Method	Parameters /10⁶	FLOPs /10⁶	Latency /ms
RandLA-Net^［4］	4.99	26.38	637.39
SCF-Net^［14］	12.15	69.25	711.84
BAF-LAC^［38］	11.64	66.17	966.88
MVPNet^［39］	138.47	572.20	1458.29
PVCNN^［22］	15.61	65.91	724.72
PVCC-Net	49.04	206.78	1133.40

Tools

Get Citation

Copy Citation Text

Huchen Li, Haiyan Guan, Xiangda Lei, Nannan Qin, Huan Ni. Point‑Voxel Consistency Constraint Network for LiDAR Point Cloud Classification Under Urban Scenes[J]. Chinese Journal of Lasers, 2024, 51(13): 1310003

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. 16, 2023

Accepted: Jan. 4, 2024

Published Online: Jul. 2, 2024

The Author Email: Haiyan Guan (guanhy.nj@nuist.edu.cn)

DOI:10.3788/CJL231411

CSTR:32183.14.CJL231411

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology