Stereo Matching Algorithm Based on Adaptive Spatial Convolution

Fanna Meng; ZouYongjia; Yang Cao; Jin Lü; Hongfei Yu

doi:10.3788/LOP241894

Laser & Optoelectronics Progress, Volume. 62, Issue 8, 0815012(2025)

Stereo Matching Algorithm Based on Adaptive Spatial Convolution

Fanna Meng^1、*, ZouYongjia¹, Yang Cao¹, Jin Lü², and Hongfei Yu¹

Author Affiliations

¹School of Artificial Intelligence and Software, Liaoning Petrochemical University, Fushun 113000, Liaoning , China

²Neusoft Reach Automotive Technology (Shenyang) Co., Ltd., Shenyang 110179, Liaoning , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Overall framework of network

Download full size

Fig. 2. Adaptive spatial convolution module

Download full size

Fig. 3. Comparison of disparity maps in KITTI reflection areas. (a) Left view; (b) benchmark algorithm; (c) ours

Download full size

Fig. 4. Comparison of disparity maps in KITTI occlusion regions. (a) Left view; (b) benchmark algorithm; (c) ours

Download full size

Fig. 5. Visualization results comparison on Middlebury dataset. (a) Left view; (b) benchmark algorithm; (c) ours

Download full size

Table 1. Pre-training model parameters
View table
Table 1. Pre-training model parameters
Parameter name Parameter value
Step 2×10⁵
Learning rate 2×10^-4
Iteration round 22
Batch size 8

Table 2. Ablation experiment
View table
Table 2. Ablation experiment
Ablation module EPE /pixel D1 /% Param /M
RAFT-Stereo 0.53 6.10 11.23
RAFT-Stereo+ASCT 0.49 5.63 11.47
RAFT-Stereo+multi-scale GRUs 0.50 5.38 11.64
RAFT-Stereo+ASCT+multi-scale GRUs （ours） 0.46 5.30 11.81

Table 3. Only ablation analysis of changing ASCT convolutional kernel number $κ$
View table
Table 3. Only ablation analysis of changing ASCT convolutional kernel number $κ$
Kernel number EPE /pixel D1 /%
2 0.50 6.08
3 0.49 5.96
4 0.49 5.63

Table 4. Only ablation analysis of changing GRU convolutional kernel size
View table
Table 4. Only ablation analysis of changing GRU convolutional kernel size
Kernel size EPE /pixel D1 /%
3×3 0.53 6.10
3×3+5×5 0.52 5.84
3×3+1×1 0.50 5.38

Table 5. Quantitative results with existing mainstream stereo matching methods on KITTI 2015 test

View table

Table 5. Quantitative results with existing mainstream stereo matching methods on KITTI 2015 test

Method	KITTI-2015 （all）			KITTI-2015 （noc）			Runtime/ s
Method	D1-bg/%	D1-fg/%	D1-all/%	D1-bg/%	D1-fg/%	D1-all/%	Runtime/ s
RAFT-Stereo^［12］	1.58	3.05	1.82	1.45	2.94	1.69	0.38
CRE-Stereo^［25］	1.45	2.86	1.69	1.33	2.60	1.54	0.41
IGEV-Stereo^［13］	1.38	2.67	1.59	1.27	2.62	1.49	0.18
Los^［26］	1.42	2.81	1.65	1.29	2.66	1.52	0.19
CroCo-Stereo^［27］	1.38	2.65	1.59	1.30	2.56	1.51	0.93
MDA^［28］	1.37	2.64	1.58	1.26	2.58	1.48	0.32
Ours	1.42	2.69	1.58	1.24	2.64	1.47	0.44

Table 6. Quantitative results with existing mainstream stereo matching methods on KITTI 2012 test

View table

Table 6. Quantitative results with existing mainstream stereo matching methods on KITTI 2012 test

Method	KITTI-2012 （all）						KITTI-2012 （reflective）
Method	2-noc	2-all	3-noc	3-all	EPE-noc	EPE-all	2-noc	2-all	3-noc	3-all
RAFT-Stereo^［12］	1.92	2.42	1.30	1.66	0.4	0.5	8.41	9.87	5.40	6.48
CRE-Stereo^［25］	1.72	2.18	1.14	1.46	0.4	0.5	9.71	11.26	6.27	7.27
IGEV-Stereo^［13］	1.71	2.17	1.12	1.44	0.4	0.4	7.29	8.48	4.11	4.76
MS-ACV^［29］	1.75	2.30	1.11	1.44	0.4	0.5	—	—	—	—
DVANet^［30］	1.78	2.39	1.09	1.52	0.4	0.5	9.63	11.99	5.68	7.48
HCR^［31］	1.69	2.18	1.09	1.42	0.4	0.4	9.78	11.93	6.01	7.68
MDA^［28］	1.76	2.26	1.09	1.43	0.4	0.5	9.79	11.89	5.64	7.22
Ours	1.63	2.06	1.07	1.40	0.4	0.5	7.13	8.01	4.39	4.62

Table 7. Quantitative results with mainstream stereo matching methods on Scene Flow, ETH3D and Middlebury

View table

Table 7. Quantitative results with mainstream stereo matching methods on Scene Flow, ETH3D and Middlebury

Method	Scene Flow	ETH3D	Middlebury
Method	EPE /pixel	D-1 /%	Half error /%	Quarter error /%
GANet^［32］	0.84	6.5	13.5	8.5
RAFT-Stereo^［12］	0.54	3.2	8.7	7.3
IGEV-Stereo^［13］	0.47	3.6	7.1	6.2
DLNR^［33］	0.48	—	9.5	—
IGEV++^［34］	0.50	3.5	7.8	—
Ours	0.46	3.8	6.6	6.0

Tools

Get Citation

Copy Citation Text

Fanna Meng, ZouYongjia, Yang Cao, Jin Lü, Hongfei Yu. Stereo Matching Algorithm Based on Adaptive Spatial Convolution[J]. Laser & Optoelectronics Progress, 2025, 62(8): 0815012

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Aug. 22, 2024

Accepted: Oct. 28, 2024

Published Online: Mar. 24, 2025

The Author Email:

DOI:10.3788/LOP241894

CSTR:32186.14.LOP241894

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Pre-training model parameters

Table 1. Pre-training model parameters

Table 2. Ablation experiment

Table 2. Ablation experiment

Table 3. Only ablation analysis of changing ASCT convolutional kernel number κ

Table 3. Only ablation analysis of changing ASCT convolutional kernel number κ

Table 4. Only ablation analysis of changing GRU convolutional kernel size

Table 4. Only ablation analysis of changing GRU convolutional kernel size

Table 5. Quantitative results with existing mainstream stereo matching methods on KITTI 2015 test

Table 5. Quantitative results with existing mainstream stereo matching methods on KITTI 2015 test

Table 6. Quantitative results with existing mainstream stereo matching methods on KITTI 2012 test

Table 6. Quantitative results with existing mainstream stereo matching methods on KITTI 2012 test

Table 7. Quantitative results with mainstream stereo matching methods on Scene Flow, ETH3D and Middlebury

Table 7. Quantitative results with mainstream stereo matching methods on Scene Flow, ETH3D and Middlebury

Table 3. Only ablation analysis of changing ASCT convolutional kernel number $κ$

Table 3. Only ablation analysis of changing ASCT convolutional kernel number $κ$