Stereo Matching Method Based on Gated Recurrent Unit Networks

Hongzhi Du; Teng Zhang; Yanbiao Sun; Linghui Yang; Jigui Zhu

doi:10.3788/LOP202158.1415001

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 14 >Page 1415001 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 14, 1415001 (2021)

Stereo Matching Method Based on Gated Recurrent Unit Networks

Hongzhi Du, Teng Zhang, Yanbiao Sun, Linghui Yang, and Jigui Zhu^*

Author Affiliations

State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University, Tianjin 300072, China

show less

DOI: 10.3788/LOP202158.1415001 Cite this Article Set citation alerts

Hongzhi Du, Teng Zhang, Yanbiao Sun, Linghui Yang, Jigui Zhu. Stereo Matching Method Based on Gated Recurrent Unit Networks[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1415001 Copy Citation Text

show less

Fig. 1. Proposed network structure

Download full size

Fig. 2. GRU structure

Download full size

Fig. 3. Stacked GRU structure

Download full size

Fig. 4. Result of the disparity estimation. (a) Left images; (b) stacked GRU; (c) proposed method; (d) ground truth

Download full size

Fig. 5. Result of the disparity estimation on KITTI2015 test dataset. (a) Left image; (b) PSMNet; (c) proposed method

Download full size

Operation	Layer setting	Output size
input		1/4H×1/4W×64
GRU_1	K=3×3,C=32	1/4H×1/4W×32
GRU_2	K=3×3,C=32	1/4H×1/4W×32
Conv_1	K=3×3,C=48,S=2	1/8H×1/8W×48
Conv_2	K=3×3,C=64,S=2	1/16H×1/16W×64
GRU_3	K=3×3,C=64	1/16H×1/16W×64
Deconv_1	K=4×4,C=48,S=2	1/8H×1/8W×48
add(Conv_1)	K=3×3,C=48,S=1	1/8H×1/8W×48
Deconv_2	K=4×4,C=32,S=2	1/4H×1/4W×32
add(Conv_2)	K=3×3,C=32,S=1	1/4H×1/4W×32
Conv_3	K=3×3,C=8,S=1	1/4H×1/4W×8
Conv_4	K=3×3,C=1,S=1	1/4H×1/4W×1

Table 1. Parameters of recurrent aggregation module

Module	E_ep /pixel	R₁ /%	R₃ /%
Stacked GRU	1.7	16.29	7.79
Proposed method	1.2	12.01	4.99

Table 2. Comparison of different cost aggregation modules

Method	E_ep /pixel	Memory /GB	t_run /ms
DispNetC^[3]	1.68	1.62	18.7
PSMNet^[7]	1.09	4.65	399.3
GANet^[8]	0.84	6.65	2251.1
Proposed method	1.22	2.57	326.8

Table 3. Performance evaluation of different methods on Scene Flow test dataset

Method	All			Noc
Method	E_D1-bg /%	E_D1-fg /%	E_D1-all /%	E_D1-bg /%	E_D1-fg /%	E_D1-all /%
DispNetC^[3]	4.32	4.41	4.34	4.11	3.72	4.05
MADNet^[18]	3.75	9.20	4.66	3.45	8.41	4.27
CRL^[4]	2.48	3.59	2.67	2.32	3.12	2.45
FADNet^[5]	2.68	3.50	2.82	2.49	3.07	2.59
GC-Net^[6]	2.21	6.16	2.87	2.02	5.58	2.61
PSMNet^[7]	1.86	4.62	2.32	1.71	4.31	2.14
Proposed method	2.20	4.85	2.64	1.82	4.09	2.20

Table 4. Evaluation of different method on KITTI2015 test dataset

Method	>2 pixel		>3 pixel		>4 pixel		>5 pixel		Avg.error
Method	Noc	All	Noc	All	Noc	All	Noc	All	Noc	All
DispNetC	7.38	8.11	4.11	4.65	2.77	3.20	2.05	2.39	0.9	1.0
FADNet	3.98	4.63	2.42	2.86	1.73	2.06	1.34	1.62	0.6	0.7
GC-Net	2.71	3.46	1.77	2.30	1.36	1.77	1.12	1.46	0.6	0.7
PSMNet	2.44	3.01	1.49	1.89	1.12	1.42	0.90	1.15	0.5	0.6
GANet	1.89	2.50	1.19	1.60	0.91	1.23	0.76	1.02	0.4	0.5
Proposed method	2.39	3.03	1.48	1.91	1.10	1.43	0.87	1.14	0.5	0.5

Table 5. Evaluation of different methods on KITTI2012 test dataset unit: %

Hongzhi Du, Teng Zhang, Yanbiao Sun, Linghui Yang, Jigui Zhu. Stereo Matching Method Based on Gated Recurrent Unit Networks[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1415001

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information