Weakly-Supervised Point Cloud Semantic Segmentation with Consistency Constraint and Feature Enhancement

Dong Wei; Yifan Bai; He Sun; Jingtian Zhang

doi:10.3788/LOP241945

Laser & Optoelectronics Progress, Volume. 62, Issue 8, 0837006(2025)

Weakly-Supervised Point Cloud Semantic Segmentation with Consistency Constraint and Feature Enhancement

Dong Wei^*, Yifan Bai, He Sun, and Jingtian Zhang

Author Affiliations

School of Information Science and Engineering, Shenyang University of Technology, Shenyang 110870, Liaoning , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(13)

Fig. 1. Structure of proposed algorithm

Download full size

Fig. 2. Structure of GFA module

Download full size

Fig. 3. Structure of LFE module

Download full size

Fig. 4. Structure of CAM module

Download full size

Fig. 5. Comparison of semantic segmentation results for various classifications

Download full size

Fig. 6. Semantic segmentation results for the S3DIS validation set by proposed method

Download full size

Fig. 7. Semantic segmentation results for Semantic3D validation set by proposed method

Download full size

Fig. 8. Semantic segmentation result for Toronto3D validation set by proposed method

Download full size

Table 1. Quantitative comparison of different methods on the Area-5 of the S3DIS dataset

View table

Table 1. Quantitative comparison of different methods on the Area-5 of the S3DIS dataset

Labeling type	Method	mIoU	OA
Fully	PointNet^［3］	41.1	—
	SPGraph^［22］	58.0	86.4
	PointWeb^［23］	60.3	87.0
	RandLA-Net^［21］	63.0	87.2
	PSD^［8］	65.1	—
1.0%	PSD^［8］	63.5	—
	HybridCR^［19］	65.3	—
	SQN^［9］	63.6	—
	RPSC^［16］	63.1	—
	Proposed	66.4	89.8
0.1%	RandLA-Net^［21］	52.9	82.0
	SQN^［9］	61.4	86.1
	PointMatch^［12］	63.4	—
	Proposed	63.4	87.9

Table 2. Quantitative comparison of different methods on the reduced-8 of the Semantic3D dataset
View table
Table 2. Quantitative comparison of different methods on the reduced-8 of the Semantic3D dataset
Labeling type Method mIoU OA
Fully KPConv^［24］ 74.6 92.9
RandLA-Net^［21］ 77.4 94.8
1.0% Ref.［14］ 72.6 93.7
PSD^［8］ 75.8 94.3
HybridCR^［19］ 76.8 94.9
Proposed 77.9 94.9
0.1% SQN^［9］ 74.7 93.7
Proposed 76.8 94.4

Table 3. Quantitative comparison of different methods on the Toronto3D dataset
View table
Table 3. Quantitative comparison of different methods on the Toronto3D dataset
Labeling type Method mIoU OA
Fully PointNet++^［4］ 41.8 84.9
DGCNN^［25］ 61.8 94.2
KPConv^［24］ 69.1 95.4
RandLA-Net^［21］ 77.7 92.9
1.0% Proposed 80.5 97.2
0.1% SQN^［9］ 77.7 96.7
Proposed 78.9 96.8

Table 4. Ablation experimental results of each module
View table
Table 4. Ablation experimental results of each module
Method In-con. Sub. Ada. CAM mIoU OA
1 × × × × 52.9 82.0
2 √ × × × 62.2 86.5
3 √ √ × × 62.6 86.9
4 √ √ √ × 63.1 87.4
5 √ √ √ √ 63.4 87.9

Table 5. Ablation experimental results of input consistency projection constraint
View table
Table 5. Ablation experimental results of input consistency projection constraint
Method mIoU OA
In-con. 66.4 89.8
Out-con. 65.4 88.6

Tools

Get Citation

Copy Citation Text

Dong Wei, Yifan Bai, He Sun, Jingtian Zhang. Weakly-Supervised Point Cloud Semantic Segmentation with Consistency Constraint and Feature Enhancement[J]. Laser & Optoelectronics Progress, 2025, 62(8): 0837006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Digital Image Processing

Received: Sep. 3, 2024

Accepted: Oct. 8, 2024

Published Online: Mar. 24, 2025

The Author Email:

DOI:10.3788/LOP241945

CSTR:32186.14.LOP241945

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Quantitative comparison of different methods on the Area-5 of the S3DIS dataset

Table 1. Quantitative comparison of different methods on the Area-5 of the S3DIS dataset

Table 2. Quantitative comparison of different methods on the reduced-8 of the Semantic3D dataset

Table 2. Quantitative comparison of different methods on the reduced-8 of the Semantic3D dataset

Table 3. Quantitative comparison of different methods on the Toronto3D dataset

Table 3. Quantitative comparison of different methods on the Toronto3D dataset

Table 4. Ablation experimental results of each module

Table 4. Ablation experimental results of each module

Table 5. Ablation experimental results of input consistency projection constraint

Table 5. Ablation experimental results of input consistency projection constraint