Researching | Online class incremental learning for multi-pose point cloud targets

Chinese Journal of Liquid Crystals and Displays, Volume. 38, Issue 11, 1542(2023)

Online class incremental learning for multi-pose point cloud targets

Run-jiang ZHANG^1,2, Jie-long GUO^2,3, Hui YU^2,3, Hai LAN², Xi-hao WAGN², and Xian WEI^2,3、*

Author Affiliations

¹College of Electrical Engineering and Automation，Fuzhou University，Fuzhou 350108，China

²Fujian Institute of Research on the Structure of Matter，Chinese Academy of Sciences，Fuzhou 350108，China

³Quanzhou Institute of Equipment Manufacturing，Haixi Institutes，Chinese Academy of Sciences，Quanzhou 362000，China

show less

Abstract Get PDF(in Chinese)

Figures&Tables (10)

References (46)

Paper Information

Figures & Tables(10)

Fig. 1. Model structure diagram

Download full size

View in Article

Fig. 2. Global maximum pooling layer

Download full size

View in Article

Fig. 3. Fixed posture target（a）and multi-posture target（b）

Download full size

View in Article

Fig. 4. Accuracy of each task at different times

Download full size

View in Article

Fig. 5. Results of MNIST with different postures

Download full size

View in Article

Table 0. [in Chinese]

View in Article

Table 0. [in Chinese]

算法2 基于损失变化的记忆重放

输入：数据流 $(x_{D}, y_{D})$ ；

输出： $θ^{n e w}$ .

初始化：记忆存储器 $M$ ，学习率 $α$ ，模型 $θ^{o l d}$ .

1）FOR $t = 1, \dots, T$ ；

2）通过公式（12）计算 $L_{1}$ ；

3）通过公式（13）更新临时模型 $θ^{t e m p} \leftarrow$ $S G D (L_{1},$ $θ^{o l d}, α)$ ；

4） $(x_{m}, y_{m}, L_{m}) \in M$ ；

5）通过公式（14）计算 $L_{2}$ ；

6）根据损失变化选择要回传的损失 ${L_{2}^{i}}_{i = 1}^{n_{1}} \leftarrow$ $s o r t (L_{m} - L_{2})$ ；

7）通过公式（15）更新模型 $θ^{n e w} \leftarrow S G D$ $({L_{2}^{i}}_{i = 1}^{n_{1}},$ $θ^{t e m p}, α)$ ；

8）更新存储器 $M \leftarrow u p d a t a (x_{D}, y_{D}, L_{1}, L_{2})$ ；

9）END.

Table 0. [in Chinese]

View in Article

Table 0. [in Chinese]

算法1 群等变点卷积

输入： ${(u_{i}, f_{i} {)}}_{i = 1}^{N * m}$ ，（ $u_{i} \in T_{g} x_{i}$ ， $m$ 为群元素的个数，第一层群等变点卷积中 $m = 1$ ）

输出： ${h_{i}}_{i = 1}^{m}$

初始化： ${g_{i}}_{i = 1}^{m} ~ μ$

1）FOR $i = 1, \dots, N * m$

2）计算 $u_{i}$ 的逆元 ${u_{i}}^{- 1}$

3）找到 $u_{i}$ 的邻域内点集 $n b h d (i)$

4）计算 $u_{i}$ 邻域内点的个数 $n_{i} = | n b h d (i) |$

5）初始化 $h_{i} = 0$

6）FOR $j = 1, \dots, N * m$

7） $h_{i} + = \sum_{j \in n b h d_{i}} Ψ (u_{i}^{- 1} u_{j}) f (u_{j})$

8）END

9） $h_{i} = \frac{1}{n_{i}} h_{i}$

10）END

Table 1. Experimental environment
View table
View in Article
Table 1. Experimental environment
实验环境环境配置
编程语言 Python3.8.13
开发工具 Pycharm11.0.11
深度学习框架 Pytorch1.11.0
CUDA 10.2
GPU GTX TITAN xp
CPU Intel（R）Xeon（R）Silver 4210 CPU @ 2.20 GHz

Table 2. Experimental results on 2D image data

View in Article

Table 2. Experimental results on 2D image data

固定姿态目标
方法	MNIST			CIFAR-10
方法	AvgACC/%	AvgF/%		AvgACC/%	AvgF/%
fine-tuning	19.8±0.5	98.9±0.4		17.2±0.6	84.3±0.4
idd online	88.6±0.9	N/A		58.3±1.7	N/A
idd offline	99.2±0.2	N/A		80.8±0.6	N/A
ER	82.1±1.5	15.0±2.1		33.1±1.7	35.4±2.0
ER-MIR	87.6±0.7	7.0±0.9		46.3±2.5	34.2±1.5
Ours	88.0±2.8	4.1±1.8		42.6±2.3	19.3±4.6
多姿态目标
方法	RotMNIST			trCIFAR-10
方法	AvgACC/%	AvgF/%		AvgACC/%	AvgF/%
fine-tuning	19.7±0.3	97.9±1.3		17.4±0.5	85.6±1.1
idd online	86.7±0.5	N/A		56.4±2.4	N/A
idd offline	98.6±0.1	N/A		81.9±3.5	N/A
ER	23.1±1.0	39.6±2.3		15.8±2.6	34.2±3.5
ER-MIR	30.7±0.6	30.4±3.4		22.3±1.3	31.4±9.6
Ours	86.1±2.1	4.5±1.4		40.4±0.9	20.3±3.4

Table 3. Experimental results on 3D data

View in Article

Table 3. Experimental results on 3D data

方法	ModelNet40		trModelNet40
方法	AvgACC/%	AvgF/%	AvgACC/%	AvgF/%
fine-tuning	7.3±0.9	67.1±2.2	6.7±0.9	58.8±2.5
idd online	51.0±1.2	N/A	47.9±0.8	N/A
idd offline	90.9±1.1	N/A	85.0±1.6	N/A
Ours	52.8±3.7	22.0±2.9	48.8±3.9	24.0±2.2

Tools

Get Citation

Copy Citation Text

Run-jiang ZHANG, Jie-long GUO, Hui YU, Hai LAN, Xi-hao WAGN, Xian WEI. Online class incremental learning for multi-pose point cloud targets[J]. Chinese Journal of Liquid Crystals and Displays, 2023, 38(11): 1542

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: Dec. 21, 2022

Accepted: --

Published Online: Nov. 29, 2023

The Author Email: Xian WEI (xian.wei@fjirsm.ac.cn)

DOI:10.37188/CJLCD.2022-0419

Topics