Point Cloud Model Recognition and Classification Based on K-Nearest Neighbor Convolutional Neural Network

Ting Yu; Jun Yang

doi:10.3788/LOP57.101510

Journals >Laser & Optoelectronics Progress >Volume 57 >Issue 10 >Page 101510 > Article

Laser & Optoelectronics Progress
Vol. 57, Issue 10, 101510 (2020)

Point Cloud Model Recognition and Classification Based on K-Nearest Neighbor Convolutional Neural Network

Ting Yu and Jun Yang^*

Author Affiliations

School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou, Gansu 730070, China

show less

DOI: 10.3788/LOP57.101510 Cite this Article Set citation alerts

Ting Yu, Jun Yang. Point Cloud Model Recognition and Classification Based on K-Nearest Neighbor Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2020, 57(10): 101510 Copy Citation Text

show less

Fig. 1. Construction process of the local neighborhood

Download full size

Fig. 2. Schematic diagram of feature extraction process

Download full size

Fig. 3. JANet model

Download full size

Fig. 4. KNN-CNN model

Download full size

Fig. 5. Graph of activation function

Download full size

Fig. 6. Three representations of three-dimensional models. (a) Rendered model; (b) mesh model; (c) 3D point cloud model

Download full size

Fig. 7. Examples of misclassified point cloud models. (a) Television stand; (b) chair; (c) plant; (d) flower pot

Download full size

Method	Input	MA /%	OA /%
PointNet^[4]	1024	86.2	89.2
PointNet++^[20]	1024	-	90.7
PointNet++^[20]	normal	-	91.9
KCNet^[22]	1024	-	91.0
Kd-Net(depth 10)^[24]	1024	86.3	90.6
Kd-Net(depth 15)^[24]	32768	88.5	91.8
Ours	1024	89.9	92.0

Table 1. Recognition accuracy of different algorithms on ModelNet40 dataset

Category	PointNet^[4]	Our model
Airplane	100.0	100.0
Bathtub	86.0	94.0
Bed	97.0	99.0
Bench	70.0	75.0
Bookshelf	91.0	99.0
Bottle	94.0	98.0
Bow	90.0	100.0
Car	98.0	99.0
Chair	97.0	98.0
Cone	95.0	100.0
Cup	80.0	75.0
Curtain	90.0	95.0
Desk	80.2	87.2
Door	80.0	95.0
Dresser	72.1	76.7
Flower pot	20.0	20.0
Glass box	98.0	92.0
Guitar	100.0	100.0
Keyboard	100.0	100.0
Lamp	95.0	90.0
Laptop	100.0	100.0
Mantel	95.0	96.0
Monitor	96.0	99.0
Night stand	72.1	82.6
Person	90.0	100.0
Piano	86.0	95.0
Plant	76.0	82.0
Radio	75.0	90.0
Range hood	92.0	95.0
Sink	80.0	80.0
Sofa	95.0	96.0
Stairs	85.0	95.0
Stool	85.0	85.0
Table	81.0	81.0
Tent	95.0	95.0
Toilet	98.0	98.0
Television stand	79.0	91.0
Vase	80.0	83.0
Wardrobe	60.0	70.0
Xbox	85.0	90.0

Table 2. Comparison of recognition accuracy of each category model in ModelNet40 test setunit:%

k₁	k₂	OA /%	Time /h
15	15	88.5	72.0
20	10	91.2	72.1
20	20	91.3	93.3
30	20	91.6	107.3

Table 3. Effect of k value on OA

Activate function	OA
SELU	91.3
ReLU+BN	91.6
ELU+BN	91.8
SELU+BN	92.0

Table 4. Recognition accuracy of different optimization schemesunit:%

Ting Yu, Jun Yang. Point Cloud Model Recognition and Classification Based on K-Nearest Neighbor Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2020, 57(10): 101510

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information