Point Cloud Semantic Segmentation Method Based on Improved Point Transformer v2

Qi Chen; Kangnian Wang; Zhiqiang Jiao; Zhanhua Huang

doi:10.3788/LOP242365

Laser & Optoelectronics Progress, Volume. 62, Issue 14, 1415001(2025)

Point Cloud Semantic Segmentation Method Based on Improved Point Transformer v2

Qi Chen^*, Kangnian Wang, Zhiqiang Jiao, and Zhanhua Huang

Key Laboratory of Opto-Electronics Information Technology, Ministry of Education, School of Precision Instrument and Opto-Electronics Engineering, Tianjin University, Tianjin 300072, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(9)

Fig. 1. Improved Point Transformer v2 model structure

Download full size

Fig. 2. Detailed structure design of each module. (a) Input embedding module; (b) encoder; (c) decoder; (d) semantic segmentation head

Download full size

Fig. 3. Construction of full-scale aggregation feature sequence $F_{D E, 2}^{}$ of the second decoder

Download full size

Fig. 4. Channel attention mechanism

Download full size

Fig. 5. Vector attention mechanisms. (a) Vector attention mechanism encoded by linear layer; (b) grouped vector attention mechanism encoded by grouped linear layer; (c) grouped vector attention mechanism encoded by linear layer

Download full size

Fig. 6. Visualization of semantic segmentation results of Point Transformer v2 and proposed model on S3DIS

Download full size

Table 1. Results of semantic segmentation on S3DIS

View table

Table 1. Results of semantic segmentation on S3DIS

Method	OA/ %	mAcc/ %	mIoU/ %	mIoU /% for each category
Method	OA/ %	mAcc/ %	mIoU/ %	Ceiling	Floor	Wall	Beam	Column	Window	Door	Table	Chair	Sofa	Bookcase	Board	Clutter
PointNet^［12］	-	49.0	41.1	88.8	97.3	69.8	0.1	3.9	46.3	10.8	59.0	52.6	5.9	40.3	26.4	33.2
PointCNN^［16］	85.9	63.9	57.3	92.3	98.2	79.4	0.0	17.6	22.8	62.1	74.4	80.6	31.7	66.7	62.1	56.7
SPGraph^［20］	86.4	66.5	58.0	89.4	69.9	78.1	0.0	42.8	48.9	61.6	84.7	75.4	69.8	52.6	2.1	56.2
PAT^［23］	-	70.8	60.1	93.0	98.5	72.3	1.0	41.5	58.1	38.2	57.7	83.6	48.1	67.0	61.3	33.6
PCT^［39］	-	67.7	61.3	92.5	98.4	80.6	0.0	19.4	61.6	48.0	76.6	85.2	46.2	67.7	67.9	52.3
PTv1^［25］	89.4	74.3	68.1	92.4	98.3	82.5	0.0	34.9	51.4	69.4	79.9	91.2	76.4	76.1	72.5	59.8
Stratified Transformer^［40］	90.0	76.3	70.1	91.9	97.2	84.5	0.1	30.5	58.0	73.2	82.3	92.6	79.2	76.0	84.5	60.5
Fast Point Transformer^［41］	-	74.8	68.8	90.0	96.4	86.2	0.2	51.0	58.3	69.0	81.2	88.6	62.2	74.4	78.8	58.5
Point Transformer v2^［26］	90.1	75.7	69.6	92.6	98.3	84.0	0.0	33.0	56.6	77.9	81.1	92.7	74.4	75.7	76.8	61.4
Proposed	90.5	77.3	71.5	91.4	98.2	85.3	0.0	37.8	58.8	80.8	82.4	95.3	81.6	77.1	81.3	59.7

Table 2. Network learnable parameters and forward time on S3DIS of Point Transformer v2 and the proposed model
View table
Table 2. Network learnable parameters and forward time on S3DIS of Point Transformer v2 and the proposed model
Method Parameters Forward time /s
Point Transformer v2 4442755 1.6335
Proposed 3813445 1.6329

Table 3. Results of module design ablation experiments
View table
Table 3. Results of module design ablation experiments
Model Module Parameters mIoU /%
Ⅰ Grouped linear-GVA 4442755 69.6
Ⅱ Linear-GVA 3908461 69.7
Ⅲ Linear-GVA+FSSC 3742597 71.1
Ⅳ Linear-GVA+FSSC+CA 3813445 71.5

Tools

Get Citation

Copy Citation Text

Qi Chen, Kangnian Wang, Zhiqiang Jiao, Zhanhua Huang. Point Cloud Semantic Segmentation Method Based on Improved Point Transformer v2[J]. Laser & Optoelectronics Progress, 2025, 62(14): 1415001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites