Point Cloud Analysis Method Based on Feature Negative Feedback Convolution

Lintao Deng; Zhijun Fang

doi:10.3788/LOP202259.1210006

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 12 >Page 1210006 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 12, 1210006 (2022)

Point Cloud Analysis Method Based on Feature Negative Feedback Convolution

Lintao Deng and Zhijun Fang^*

Author Affiliations

School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China

show less

DOI: 10.3788/LOP202259.1210006 Cite this Article Set citation alerts

Lintao Deng, Zhijun Fang. Point Cloud Analysis Method Based on Feature Negative Feedback Convolution[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1210006 Copy Citation Text

show less

Fig. 1. Schematic diagram of overall network structure

Download full size

Fig. 2. Diagram of structure aware KNN and classic KNN

Download full size

Fig. 3. Feature negative feedback convolution module

Download full size

Fig. 4. Global semantic reasoning module

Download full size

Fig. 5. Visualization of part segmentation results

Download full size

Fig. 6. Test results of sparser points

Download full size

Method	Input type	Points	Mean class accuracy /%	Overall accuracy /%
Point-CNN^［40］	coords	1k	88.1	91.7
PointNet^［15］	coords	1k	86.0	89.2
A-SCN^［28］	coords	1k		90.0
Kd-Net^［38］	coords	1k		90.6
PointNet++^［16］	coords	1k		90.7
KCNet^［41］	coords	1k		91.0
Spec-GCN^［42］	coords	1k		91.5
DGCNN^［17］	coords	1k	90.2	92.2
RS-CNN^［18］	coords	1k		93.6
KP-Conv^［43］	coords	1k		92.9
PointASNL^［44］	coords	1k		93.2
Proposed	coords	1k	91.0	93.8
SpiderCNN^［45］	coords + norm	5k		92.4
DensePoint^［46］	coords + norm	1k		93.2
SO-NET^［39］	coords + norm	5k	90.8	93.4
PointNet++^［16］	coords + norm	5k		91.9
DGCNN^［17］	coords	2k	90.7	93.5

Table 1. Classification accuracy on ModelNet40 data set

Method	PointNet^［15］	PointNet++^［16］	SpiderCNN^［45］	SO-NET^［39］	A-SCN^［28］	P2Sequence^［48］	PCNN^［49］	DGCNN^［17］	RS-CNN^［18］	PointASNL^［44］	Proposed
Overall mIou	83.7	85.1	85.3	84.6	84.6	85.2	85.1	85.2	86.2	86.1	86.4
Air plane	83.4	82.4	83.5	81.9	83.8	82.6	82.4	84.0	83.5	84.1	84.3
Bag	78.7	79.0	81.0	83.5	80.8	81.8	80.1	83.4	84.8	84.7	85.1
Cap	82.5	87.7	87.2	84.8	83.5	87.5	85.5	86.7	88.8	87.9	88.6
Car	74.9	77.3	77.5	78.1	79.3	77.3	79.5	77.8	79.6	79.7	79.9
Chair	89.6	90.8	90.7	90.8	90.5	90.8	90.8	90.6	91.2	92.2	91.3
Ear phone	73.0	71.8	76.8	72.2	69.8	77.1	73.2	74.7	81.1	73.7	79.2
Guitar	91.5	91.0	91.1	90.1	91.7	91.1	91.3	91.2	91.6	91.0	91.8
Knife	85.9	85.9	87.3	83.6	86.5	86.9	86.0	87.5	88.4	87.2	89.0
Lamp	80.8	83.7	83.3	82.3	82.9	83.9	85.0	82.8	86.0	84.2	85.2
Laptop	95.3	95.3	95.8	95.2	96.0	95.7	95.7	95.7	96.0	95.8	95.7
Motobike	65.2	71.6	70.2	69.3	69.2	70.8	73.2	66.3	73.7	74.4	72.3
Mug	93.0	94.1	93.5	94.2	93.8	94.6	94.8	94.9	94.1	95.2	94.5
Pistol	81.2	81.3	82.7	80.0	82.5	79.3	83.3	81.1	83.4	81.0	82.0
Rocket	57.9	58.7	59.7	51.6	62.9	58.1	51.0	63.5	60.5	63.0	60.3
Skate board	72.8	76.4	75.8	72.1	74.4	75.2	75.0	74.5	77.7	76.3	76.4
Table	80.6	82.6	82.8	82.6	80.8	82.8	81.8	82.6	83.6	83.2	84.4

Table 2. Part segmentation results [mIou (%)] on ShapeNet Part data set

Model	SAKNN	Attentive pooling	Feature negative feedback convolution	Global context reasoning module	Overall mIou /%
0					84.9
1	√				85.2
2	√	√			85.4
3	√	√	√		85.8
4	√	√	√	√	86.1

Table 3. Ablation experiments about effects of different network components on ShapeNet part data set

Lintao Deng, Zhijun Fang. Point Cloud Analysis Method Based on Feature Negative Feedback Convolution[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1210006

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information