Lightweight Point Cloud Classification Model Based on Offset Attention Mechanism

Leicheng Yang; Yuhong Du; Guangyu Dong

doi:10.3788/LOP242058

Laser & Optoelectronics Progress, Volume. 62, Issue 10, 1015006(2025)

Lightweight Point Cloud Classification Model Based on Offset Attention Mechanism

Leicheng Yang^1,2, Yuhong Du^1,2、*, and Guangyu Dong^1,2

Author Affiliations

¹School of Mechanical Engineering, Tiangong University, Tianjin 300387, China

²Key Laboratory of Advanced Mechatronics Equipment Technology, Tianjin 300387, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Proposed network structure

Download full size

Fig. 2. Local feature aggregation module

Download full size

Fig. 3. Offset attention mechanism

Download full size

Fig. 4. Partial ModelNet40 data. (a) Car; (b) desk; (c) bed; (d) airplane; (e) bookcase; (f) chair

Download full size

Fig. 5. Loss and OA variation curves. (a) Loss variation curves; (b) OA variation curves

Download full size

Table 1. Software configuration
View table
Table 1. Software configuration
Item Configuration
GPU NVIDIA GeForce RTX 2060 SUPER
CUDA CUDA11.3
Programming language Python 3.7.16
Deep learning framework PyTorch 1.12.1

Table 2. Comparison of classification results on ModelNet40 dataset

View table

Table 2. Comparison of classification results on ModelNet40 dataset

Model	OA /%	mAcc /%
PointNet^［9］	89.2	86.2
PointNet++^［10］	91.9	–
PointPN^［14］	93.8	–
PointConT^［19］	93.5	–
PointConv^［20］	92.5	–
DGCNN^［21］	92.9	90.2
LDGCNN^［22］	92.9	90.3
GAPointNet^［23］	93.0	90.3
Point Tranformer^［15］	92.8	–
PCT^［16］	93.2	–
Point-PT（N=2）	92.9	89.7
Point-PT（N=3）	93.9	90.7

Table 3. Comparison of classification accuracy
View table
Table 3. Comparison of classification accuracy
Model OA /%
Cup Flower pot Vase Bookshelf Plant Bed
PCT 65 10 85 99 85 99
Point-PT（N=2） 70 10 85 98 87 100

Table 4. Model complexity comparison

View table

Table 4. Model complexity comparison

Model	Quantity of parameters /Mbit	Time /s	OA /%
PointNet^［9］	3.65	62	89.2
PointNet++^［10］	1.73	64	91.9
PointPN^［14］	0.80	31	93.8
DGCNN^［21］	1.89	94	92.9
PCT^［16］	3.10	73	93.2
PointMLP^［13］	12.60	182	94.1
Point-PT（N=2）	0.36	15	92.9
Point-PT（N=3）	0.87	30	93.9

Table 5. Model correctness for different number of neighborhood points k_N
View table
Table 5. Model correctness for different number of neighborhood points k_N
Index Neighborhood points k_N
20 30 35 40
OA /% 92.5 92.5 92.6 92.9
mAcc /% 89.0 90.0 90.1 89.7
Time /s 12.0 13.0 14.0 15.0

Table 6. Results of ablation experiment
View table
Table 6. Results of ablation experiment
OA /%
√ × ××
N=2 92.9（0.36 Mbit） 92.6（0.27 Mbit） 90.8（0.30 Mbit）
N=3 93.9（0.87 Mbit） 92.6（0.51 Mbit） 91.1（0.65 Mbit）
N=4 93.7（2.77 Mbit） 92.7（1.33 Mbit） 91.4（1.88 Mbit）

Tools

Get Citation

Copy Citation Text

Leicheng Yang, Yuhong Du, Guangyu Dong. Lightweight Point Cloud Classification Model Based on Offset Attention Mechanism[J]. Laser & Optoelectronics Progress, 2025, 62(10): 1015006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites