Semantic Recognition and Segmentation of 3D Point Clouds Using Multistage Hierarchical Fusion Residual MLP

Method	Number of inputs	OA /%
PointNet^［6］	1000	89.2
MKConv^［17］	1024	94.0
PointASNL^［11］	1000	92.9
RepSurf-U^［18］	1000	94.7
FPConv^［19］	1000	92.5
PointNet++^［20］	1000	90.7
DGCNN^［21］	1000	92.9
SpiderCNN^［22］	1000+normal	92.4
Point2vec（+Voting）^［23］		94.8
Recon++（-L）^［24］	1000	94.8
PAConv^［25］	1000	93.2
PTV2^［15］		94.2
GBNet^［14］	1000	93.8
Ours	1000	95.1

Table 2. Comparison of segmentation accuracy

View table

Table 2. Comparison of segmentation accuracy

Method	Cls. mIoU /%	Inst.mIoU /%	Airplane	Bag	Earphone	Guitar	Knife	Laptop	Motor	Mug	Pistol	Rocket
PointNet^［6］	80.4	83.7	83.4	78.7	73.0	91.5	85.9	95.3	65.2	93.0	81.2	57.9
SO-Net^［26］		84.9	82.8	77.8	73.5	90.7	83.9	94.8	69.1	94.2	80.9	53.1
Kd-Net^［27］		82.3	80.1	74.6	73.5	90.2	87.2	94.9	57.4	86.7	78.1	51.8
PCNN^［28］	81.8	85.1	82.4	80.1	73.2	91.3	86.0	95.7	73.2	94.8	83.3	51.0
PointNet++^［20］		85.1	82.4	79.0	71.8	91.0	85.9	95.3	71.6	94.1	81.3	58.7
SpiderCNN^［22］	82.4	85.3	83.5	81.0	76.8	91.1	87.3	95.8	70.2	93.5	82.7	59.7
PointASNL^［11］		86.1	84.1	84.7	73.7	91.0	87.2	95.8	74.4	95.2	81.0	63.0
DGCNN^［21］	82.3	85.2	84.0	83.4	74.7	91.2	87.5	95.7	66.3	94.9	81.1	63.5
Ours	85.1	86.6	84.2	83.0	80.1	92.8	89.0	96.9	77.8	95.9	85.1	66.3

Table 3. Effect of network depth on recognition and segmentation
View table
Table 3. Effect of network depth on recognition and segmentation
Depth ofnetwork OA of recognition /% mIoU of segmentation /%
24 layers 94.2 84.4
40 layers 95.1 86.6
56 layers 93.9 83.2

Table 4. Network depth value

View table

Table 4. Network depth value

Depth of network	［low₁，low₂，low₃，low₄］	［deep₁，deep₂，deep₃，deep₄］
24 layers	［1，1，1，1］	［1，1，1，1］
40 layers	［2，2，2，2］	［2，2，2，2］
56 layers	［3，3，3，3］	［3，3，3，3］

Table 5. Effect of the T-Net module on model recognition and segmentation accuracy at different network depth

View table

Table 5. Effect of the T-Net module on model recognition and segmentation accuracy at different network depth

T-Net module /Embedding layer	Network depth	OA of recognition /%	mIoU of segmentation /%
T-Net module	24 layers	94.2	84.2
Embedding layer	24 layers	92.7	83.8
T-Net module	40 layers	95.1	86.6
Embedding layer	40 layers	93.8	83.5
T-Net module	56 layers	93.9	83.9
Embedding layer	56 layers	92.0	83.6

Table 6. Effect of feature extraction operators $Γ_{l o w}$ and $Γ_{d e e p}$ on recognition and segmentation accuracy
View table
Table 6. Effect of feature extraction operators $Γ_{l o w}$ and $Γ_{d e e p}$ on recognition and segmentation accuracy
$Γ_{l o w}$ $Γ_{d e e p}$ OA of recognition /% mIoU of segmentation /%
√ × 94.7 86.0
× √ 94.1 85.2
√ √ 95.1 86.6

Table 7. Effect of expanding networks on recognition and segmentation accuracy
View table
Table 7. Effect of expanding networks on recognition and segmentation accuracy
RM module OA of recognition /% mIoU of segmentation /%
RM+ 95.3 86.6
RM++ 95.4 86.7

Table 8. Effect of network depth on recognition and segmentation efficiency
View table
Table 8. Effect of network depth on recognition and segmentation efficiency
Network depth Number of parameters /10⁶ Training speed /（sample/s） Testing speed（sample /s）
24 layers 0.79 120 176
40 layers 0.94 78.1 140
56 layers 12.9 49.1 112

Tools

Get Citation

Copy Citation Text

Jun Yang, Jiachen Guo. Semantic Recognition and Segmentation of 3D Point Clouds Using Multistage Hierarchical Fusion Residual MLP[J]. Laser & Optoelectronics Progress, 2025, 62(4): 0415007

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: May. 13, 2024

Accepted: Jul. 29, 2024

Published Online: Feb. 12, 2025

The Author Email:

DOI:10.3788/LOP241270

CSTR:32186.14.LOP241270

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Comparison of model recognition accuracy

Table 1. Comparison of model recognition accuracy

Table 2. Comparison of segmentation accuracy

Table 2. Comparison of segmentation accuracy

Table 3. Effect of network depth on recognition and segmentation

Table 3. Effect of network depth on recognition and segmentation

Table 4. Network depth value

Table 4. Network depth value

Table 5. Effect of the T-Net module on model recognition and segmentation accuracy at different network depth

Table 5. Effect of the T-Net module on model recognition and segmentation accuracy at different network depth

Table 6. Effect of feature extraction operators Γlow and Γdeep on recognition and segmentation accuracy

Table 6. Effect of feature extraction operators Γlow and Γdeep on recognition and segmentation accuracy

Table 7. Effect of expanding networks on recognition and segmentation accuracy

Table 7. Effect of expanding networks on recognition and segmentation accuracy

Table 8. Effect of network depth on recognition and segmentation efficiency

Table 8. Effect of network depth on recognition and segmentation efficiency

Table 6. Effect of feature extraction operators $Γ_{l o w}$ and $Γ_{d e e p}$ on recognition and segmentation accuracy

Table 6. Effect of feature extraction operators $Γ_{l o w}$ and $Γ_{d e e p}$ on recognition and segmentation accuracy