Lightweight Template Matching Algorithm Based on Rendering Perspective Sampling

Daizhou Wen; Xi Wang; Mingjun Ren

doi:10.3788/LOP240469

Journals >Laser & Optoelectronics Progress >Volume 61 >Issue 18 >Page 1837009 > Article

Laser & Optoelectronics Progress
Vol. 61, Issue 18, 1837009 (2024)

Lightweight Template Matching Algorithm Based on Rendering Perspective Sampling

Daizhou Wen, Xi Wang, and Mingjun Ren^*

Author Affiliations

School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China

show less

DOI: 10.3788/LOP240469 Cite this Article Set citation alerts

Daizhou Wen, Xi Wang, Mingjun Ren. Lightweight Template Matching Algorithm Based on Rendering Perspective Sampling[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1837009 Copy Citation Text

show less

Fig. 1. Algorithm framework

Download full size

Fig. 2. Attention lightweight convolutional submodule

Download full size

Fig. 3. Example of angle sampling

Download full size

Fig. 4. Examples. (a) LINEMOD^[36]; (b) Occlusion-LINEMOD^[37]

Download full size

Fig. 5. Comparison of Blender and PyTorch3D rendering results

Download full size

Fig. 6. Comparison of qualitative experiments before and after optimization

Download full size

算法1：迭代优化过程的渲染视角采样

输入： $n - 1$ 次采样间隔的点集 ${\{P\}}_{n - 1}$ ，边集 ${\{E\}}_{n - 1}$ ，当前最优位姿 $p_{n - 1}$ 对应点 $P_{n - 1}^{k}$

输出： $n$ 次采样间隔的点集 ${\{P\}}_{n}$

For $P_{n - 1}^{j}$ in ${\{P\}}_{n - 1}$ ：

If $(P_{n - 1}^{j}, P_{n - 1}^{k})$ in ${\{E\}}_{n - 1}$ Then

${\{P_{a d j}\}}_{n - 1}^{k}$ .append（ $P_{n - 1}^{j}$ ）

For $P_{a d j}^{i}$ in ${\{P_{a d j}\}}_{n - 1}^{k}$ ：

$P_{n}^{i} = (P_{a d j}^{i} + P_{n - 1}^{k}) / 2$

${\{P\}}_{n} = \{P_{n - 1}^{k}, P_{n}^{1}, P_{n}^{2}, \dots, P_{n}^{i}\}$

End

Table 0. [in Chinese]

Block name	Input channel	Output channel	Kernel size	Expand channel
Conv	3	32	3	–
BaseBlock	32	16	3	–
BaseBlock	16	24	3	96
BaseBlock	24	24	3	144
BaseBlock	24	40	5	144
BaseBlock	40	40	5	240
Conv	40	16	1	–

Table 1. Structure of proposed model

Network	Running time /min	Inference time /ms	Number of parameters /10⁶	Memory size /MB
BaseNet^［25-26］	0.31	5.68	0.024	0.048
ResNet^［23］	1.15	1126.20	24.036	92.053
Proposed	0.74	48.37	0.043	20.464

Table 2. Comparison of model parameter quantity

Split No.	Train	Test	Unseen object name
1	9954	15038	Ape，benchvise，camera，can
2	9928	15064	Cat，driller，duck，eggbox
3	8850	16142	Glue，holepuncher，iron，lamp，phone

Table 3. Details of datasets

Method	Render	Seen LM	Seen O-LM	Unseen LM	Unseen O-LM
Wohlhart’s^［25］	Blender	95.2	19.6	13.3	8.2
Balntas’s^［26］	Blender	96.3	18.3	11.5	7.1
Template-pose^［23］	Blender	99.3	77.3	94.4	71.4
Template-pose^［23］	PyTorch3D	98.5	75.7	95.7	70.3
Proposed	PyTorch3D	97.6	80.0	98.1	74.8

Table 4. Comparison of Acc15 on LM and O-LM datasets

Split No.	Method	Render	Seen LM	Seen O-LM	Unseen LM	Unseen O-LM	Average
2	Template-pose	Blender	99.0	84.1	97.4	72.7	88.3
2	Template-pose	PyTorch3D	97.0	77.9	96.6	70.5	85.5
2	Proposed	PyTorch3D	96.8	81.9	98.8	76.5	88.5
3	Template-pose	Blender	99.2	76.8	88.7	85.3	87.5
3	Template-pose	PyTorch3D	99.0	75.2	85.2	78.3	84.4
3	Proposed	PyTorch3D	99.2	78.4	88.3	88.3	88.6

Table 5. Repeated experiments with LM and O-LM datasets

Condition	Seen LM	Seen O-LM	Unseen LM	Unseen O-LM
Without iteration	4.70	11.45	5.57	20.94
With iteration	4.16	11.18	4.74	20.47

Table 6. Comparison of iterative optimization

Condition	Seen LM	Seen O-LM	Unseen LM	Unseen O-LM	Average
With SE	99.2	78.4	88.3	88.3	88.6
Without SE	99.1	77.2	86.6	86.8	87.4

Table 7. Attention mechanism ablation experiment

Condition

Seen LM /%

Seen O-LM /%

Unseen LM /%

Unseen O-LM /%

Average /%

Inference time /ms

With DW

99.2

78.4

88.3

88.6

48.37

Without

97.4

70.0

62.5

43.6

68.4

669.25

Table 8. Deep separable convolutional ablation experiment

Daizhou Wen, Xi Wang, Mingjun Ren. Lightweight Template Matching Algorithm Based on Rendering Perspective Sampling[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1837009

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享