Point Cloud Analysis Method Based on Spatial Feature Attention Mechanism

Yanlin Qu; Yue Wang; Qian Zhang; Shaokun Han

doi:10.3788/LOP230840

Laser & Optoelectronics Progress, Volume. 60, Issue 24, 2415003(2023)

Point Cloud Analysis Method Based on Spatial Feature Attention Mechanism

Yanlin Qu, Yue Wang, Qian Zhang, and Shaokun Han^*

Beijing Key Lab for Precision Optoelectronic Measurement Instrument and Technology, School of Optics and Photonics, Beijing Institute of Technology, Beijing 100081, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. CSA module structure

Download full size

Fig. 2. CA module structure

Download full size

Fig. 3. SA module structure

Download full size

Fig. 4. Overview of the CSA-PointNet++ structure

Download full size

Fig. 5. Structure of the improved set abstraction module

Download full size

Fig. 6. Kinect V2 camera. (a) Appearance; (b) internal structure

Download full size

Fig. 7. Real-world self-constructed data. (a) Original point cloud data; (b) intercepted point cloud data

Download full size

Fig. 8. Point cloud data after translational transformation

Download full size

Fig. 9. Point cloud data after rotational transformation

Download full size

Fig. 10. Part segmentation results

Download full size

Table 1. Classification results under ModelNet40 dataset (ACC)
View table
Table 1. Classification results under ModelNet40 dataset (ACC)
Method ACC /%
MVCNN 90.10
VoxNet 85.90
PointNet 90.32
PointNet++ 91.89
DGCNN 92.20
PCT 93.07
Ours（r=4，4） 93.20
Ours（r=4，16） 93.16
Ours（r=16，16） 93.10
Ours（r=16，4） 92.55
Ours（without ReLU） 92.56

Table 2. Part segmentation results under ShapeNetPart dataset (IoU)

View table

Table 2. Part segmentation results under ShapeNetPart dataset (IoU)

Method	mIoU	airplane	bag	cap	car	chair	earphone	guitar	knife
PointNet	79.87	81.73	77.28	87.18	75.01	90.24	73.54	90.81	86.05
PointNet++	81.75	82.21	81.91	83.17	78.41	90.62	73.90	91.15	86.38
DGCNN	82.20	82.01	81.45	84.70	78.83	90.56	73.89	91.47	86.58
PCT	82.31
Ours	82.62	82.64	80.67	85.35	78.94	90.52	75.90	91.36	87.18
Method	lamp	laptop	motorbike	mug	pistol	rocket	skateboard	table
PointNet	81.85	95.15	61.27	93.26	79.98	48.17	74.22	82.25
PointNet++	83.06	95.64	70.91	95.87	80.65	55.19	76.84	82.04
DGCNN	83.81	95.70	70.37	94.97	81.15	61.53	76.56	81.65
PCT
Ours	84.08	95.63	71.33	95.43	82.06	61.91	76.61	82.29

Table 3. ACC of the ablation experiment
View table
Table 3. ACC of the ablation experiment
Method ACC /%
PointNet++ 91.89
CA+PointNet++ 92.93
SA+PointNet++ 92.99
CSA+PointNet++ 93.20
SCA+PointNet++ 92.20

Table 4. Part segmentation results of the ablation experiment (mIoU)
View table
Table 4. Part segmentation results of the ablation experiment (mIoU)
Method mIoU /%
PointNet++ 81.75
CA+PointNet++ 82.15
SA+PointNet++ 82.42
CSA+PointNet++ 82.62
SCA+PointNet++ 82.13

Table 5. ACC under the real-world self-constructed dataset
View table
Table 5. ACC under the real-world self-constructed dataset
Type ACC /%
With ground 52.81
Without ground 92.14

Table 6. ACC of the translated real-world self-constructed dataset
View table
Table 6. ACC of the translated real-world self-constructed dataset
Translation ACC /%
（x+1，y，z） 90.96
（x，y+1，z） 91.85
（x，y，z+1） 90.85
（x+1，y+1，z+1） 91.70

Table 7. ACC of the rotated real-world self-constructed dataset
View table
Table 7. ACC of the rotated real-world self-constructed dataset
Rotation ACC /%
Rot_30° 91.75
Rot_60° 89.59
Rot_90° 92.91
Rot_120° 90.99
Rot_150° 91.40
Rot_180° 90.59
Rot_210° 91.00
Rot_240° 90.01
Rot_270° 90.93
Rot_300° 91.37
Rot_330° 92.55

Tools

Get Citation

Copy Citation Text

Yanlin Qu, Yue Wang, Qian Zhang, Shaokun Han. Point Cloud Analysis Method Based on Spatial Feature Attention Mechanism[J]. Laser & Optoelectronics Progress, 2023, 60(24): 2415003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites