Point Cloud Segmentation Model Integrating Deep Residual Shrinkage and Lightweight Transformer Encoding Technique

Hongxu Li; Yitao Lu; Ronghua Chi; Shiyu Li; Zhenbo Yang

doi:10.3788/LOP241922

Laser & Optoelectronics Progress, Volume. 62, Issue 10, 1015002(2025)

Point Cloud Segmentation Model Integrating Deep Residual Shrinkage and Lightweight Transformer Encoding Technique

Hongxu Li^1,2、*, Yitao Lu¹, Ronghua Chi^1,2, Shiyu Li², and Zhenbo Yang^1、**

¹School of Electronics and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, Jiangsu , China

²Jiangsu Province Engineering Research Center of Integrated Circuit Reliability Technology and Testing System, Wuxi University, Wuxi 214105, Jiangsu , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Flowchart of proposed segmentation method

Download full size

Fig. 2. Structure of DRST-UNet

Download full size

Fig. 3. Structure of SE attention

Download full size

Fig. 4. Comparison between DRSAM and traditional residual block. (a) Structure of traditional residual block; (b) structure of DRSAM

Download full size

Fig. 5. Structure of Transformer encoder

Download full size

Fig. 6. Point clouds of partial scenes in the ScanNetV2 dataset

Download full size

Fig. 7. Schematic diagram for IoU calculation

Download full size

Fig. 8. Visualization of segmentation results for partial scenes in the ScanNetV2 test set

Download full size

Fig. 9. Visualization of segmentation results for partial scenes in the ScanNetV2 validation set

Download full size

Fig. 10. Visualization of large scene segmentation results on the ScanNetV2 and S3DIS datasets

Download full size

Table 1. Performance comparison of different methods on the ScanNetV2 test set

View table

Table 1. Performance comparison of different methods on the ScanNetV2 test set

Method	AP	AP₅₀	AP₂₅
PanopticFusion^［21］	21.4	47.8	69.3
3D-BoNet^［5］	25.3	48.8	68.7
PointGroup^［9］	40.7	63.6	77.8
GICN^［22］	34.1	63.8	78.8
Dyco3D^［14］	39.5	64.1	76.1
Searilized Point Mamba^［23］	40.0	61.4	76.4
Proposed	41.5	66.4	78.6

Table 2. AP₅₀ comparison of different methods on the ScanNetV2 test set

View table

Table 2. AP₅₀ comparison of different methods on the ScanNetV2 test set

Class	SGPN^［6］	3D-BEVIS^［8］	GSPN^［4］	PanopticFusion^［21］	3D-BoNet^［5］	PointGroup^［9］	GICN^［22］	MaskGroup^［24］	Dyco3D^［14］	Proposed
Average	14.4	24.8	30.6	47.8	48.8	63.6	63.8	66.4	64.1	66.4
Cabinet	6.5	3.5	34.8	25.9	30.1	50.5	48.0	61.6	53.1	61.9
Bed	39.0	56.6	40.5	71.2	67.2	76.5	89.5	82.2	84.1	84.3
Chair	27.5	39.4	58.9	55.0	48.4	79.7	67.6	81.5	80.2	81.1
Sofa	35.1	60.4	39.6	59.1	49.9	75.6	83.6	81.9	76.4	78.8
Table	16.8	18.1	27.5	26.7	51.3	55.6	59.9	63.5	65.7	66.7
Door	8.7	9.9	28.3	25.0	34.1	44.1	44.7	45.9	43.8	45.1
Window	13.8	17.1	24.5	35.9	43.9	51.3	47.3	60.3	56.8	57.7
Bookeshe.	16.9	7.6	31.1	59.5	59.0	62.4	80.0	76.4	89.3	86.6
Picture	1.4	2.5	2.8	43.7	12.5	47.6	36.5	59.9	43.0	41.3
Counter	2.9	2.7	5.4	0.0	9.8	11.6	14.4	13.9	11.5	11.1
Desk	0.0	9.8	12.6	17.5	30.6	38.4	35.4	59.7	44.8	45.7
Curtain	6.9	3.5	6.8	61.3	62.0	69.6	73.7	69.4	58.8	64.5
Refrige.	2.7	9.8	21.9	41.1	43.4	59.6	70.0	60.0	55.0	55.3
S.Curtain	0	37.5	21.4	85.7	79.6	100.0	100.0	51.6	85.7	100.0
Toilet	43.8	85.4	82.1	94.4	90.0	99.7	100.0	100.0	98.7	100.0
Sink	11.2	12.6	33.1	48.5	40.2	66.6	56.9	71.5	53.4	58.1
Bathtub	20.8	66.7	50.0	66.7	100.0	100.0	100.0	100.0	100.0	100.0
Otherfu.	4.3	3.0	29.0	43.4	25.9	55.9	40.0	56.6	53.7	56.6

Table 3. Comparison of experimental results by different methods on the S3DIS Area 5

View table

Table 3. Comparison of experimental results by different methods on the S3DIS Area 5

Method	AP	AP₅₀	mPrec₅₀	mRec₅₀
SGPN^［6］	—	—	36.0	28.7
ASIS^［25］	—	—	55.3	42.4
3D-BoNet^［5］	—	—	57.5	40.2
3D-MPA^［11］	—	—	63.1	58.0
PointGroup^［9］	—	57.8	61.9	62.1
SSTNet^［26］	42.7	59.3	65.5	64.2
SoftGroup^［13］	51.6	66.1	73.6	66.6
Proposed	49.8	61.7	63.0	67.3

Table 4. Accuracy ablation experimental results of different modules on the ScanNetV2 validation set
View table
Table 4. Accuracy ablation experimental results of different modules on the ScanNetV2 validation set
Plan Basic model DRSAM Transformer AP AP₅₀ AP₂₅
1 √ — — 35.2 57.1 71.4
2 √ √ — 38.1 61.3 74.1
3 √ — √ 36.6 59.9 72.5
4 √ √ √ 40.9 63.3 75.7

Table 5. Time-cost ablation experimental results of different modules on the ScanNetV2 validation set

View table

Table 5. Time-cost ablation experimental results of different modules on the ScanNetV2 validation set

Plan	Basic model	DRSAM	Transformer	Scene 1	Scene 2	Scene 3	Scene 4	Average
1	√	—	—	355.20	101.80	209.80	97.20	191.00
2	√	√	—	340.60	103.40	207.40	99.20	187.65
3	√	—	√	385.40	103.20	202.80	99.40	197.70
4	√	√	√	365.20	106.00	200.20	99.00	192.60

Table 6. Results of the model fused with Transformer under different clustering radii on the ScanNetV2 validation set
View table
Table 6. Results of the model fused with Transformer under different clustering radii on the ScanNetV2 validation set
Method Radius /cm AP /% AP₅₀ /% AP₂₅ /%
Basic model 2 28.5 50.1 65.1
3 34.8 56.9 71.3
4 33.7 55.2 70.0
5 34.2 55.2 69.9
Basic model+Transformer 2 28.8 51.1 64.8
3 36.3 59.8 72.3
4 34.8 57.1 72.0
5 35.5 58.3 72.5

Tools

Get Citation

Copy Citation Text

Hongxu Li, Yitao Lu, Ronghua Chi, Shiyu Li, Zhenbo Yang. Point Cloud Segmentation Model Integrating Deep Residual Shrinkage and Lightweight Transformer Encoding Technique[J]. Laser & Optoelectronics Progress, 2025, 62(10): 1015002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Aug. 28, 2024

Accepted: Nov. 1, 2024

Published Online: Apr. 22, 2025

The Author Email: Hongxu Li (hongxuli@cwxu.edu.cn), Zhenbo Yang (1040056581@qq.com)

DOI:10.3788/LOP241922

CSTR:32186.14.LOP241922

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Performance comparison of different methods on the ScanNetV2 test set

Table 1. Performance comparison of different methods on the ScanNetV2 test set

Table 2. AP50 comparison of different methods on the ScanNetV2 test set

Table 2. AP50 comparison of different methods on the ScanNetV2 test set

Table 3. Comparison of experimental results by different methods on the S3DIS Area 5

Table 3. Comparison of experimental results by different methods on the S3DIS Area 5

Table 4. Accuracy ablation experimental results of different modules on the ScanNetV2 validation set

Table 4. Accuracy ablation experimental results of different modules on the ScanNetV2 validation set

Table 5. Time-cost ablation experimental results of different modules on the ScanNetV2 validation set

Table 5. Time-cost ablation experimental results of different modules on the ScanNetV2 validation set

Table 6. Results of the model fused with Transformer under different clustering radii on the ScanNetV2 validation set

Table 6. Results of the model fused with Transformer under different clustering radii on the ScanNetV2 validation set

Table 2. AP₅₀ comparison of different methods on the ScanNetV2 test set

Table 2. AP₅₀ comparison of different methods on the ScanNetV2 test set