Stereo Matching Algorithm Based on Three-Dimensional Convolutional Neural Network

Yufeng Wang; Hongwei Wang; Guang Yu; Mingquan Yang; Yuwei Yuan; Jicheng Quan

doi:10.3788/AOS201939.1115001

Journals >Acta Optica Sinica >Volume 39 >Issue 11 >Page 1115001 > Article

Acta Optica Sinica
Vol. 39, Issue 11, 1115001 (2019)

Stereo Matching Algorithm Based on Three-Dimensional Convolutional Neural Network

Yufeng Wang^1、2, Hongwei Wang^2、3、**, Guang Yu², Mingquan Yang², Yuwei Yuan⁴, and Jicheng Quan^1、2、*

Author Affiliations

¹Naval Aviation University, Yantai, Shandong 264001, China

²Aviation University of Air Force, Changchun, Jilin 130022, China

³Information Engineering University, Zhengzhou, Henan 450001, China

⁴The 91977 Troops, Beijing 102200, China

show less

DOI: 10.3788/AOS201939.1115001 Cite this Article Set citation alerts

Yufeng Wang, Hongwei Wang, Guang Yu, Mingquan Yang, Yuwei Yuan, Jicheng Quan. Stereo Matching Algorithm Based on Three-Dimensional Convolutional Neural Network[J]. Acta Optica Sinica, 2019, 39(11): 1115001 Copy Citation Text

show less

Fig. 1. Architecture overview of proposed method

Download full size

Fig. 2. Graphical depiction of sampling cost in disparity dimension. (a) S=1, C=1; (b) S=2, C=1; (c) S=2, C=4

Download full size

Fig. 3. Disparity predicted by proposed method. (a) Left image; (b) disparity map; (c) error map; (d) local details

Download full size

Method	S	E_EP /pixel	E_D1 /%	t_Run /s	GPU /GB
PSMNet^[14]	1	1.09	-	-	-
	1	1.02	3.41	0.75	2.16
	2	1.10	3.89	0.45	1.51
	3	1.16	4.34	0.35	1.32
Proposed	4	1.22	4.81	0.30	1.20
	5	1.30	5.25	0.25	1.09
	6	1.34	5.62	0.25	1.08
	7	1.41	6.07	0.22	1.01
	8	1.42	6.23	0.22	1.01

Table 1. Performance evaluation of proposed method with different S (C=1)

S	C	E_EP /pixel	E_D1 /%	t_Run /s	GPU /GB
1	1	1.02	3.41	0.75	2.16
	1	1.10	3.89	0.45	1.51
	2	1.07	3.71	0.48	1.68
2	3	1.05	3.63	0.51	1.85
	4	1.08	3.81	0.53	2.02
	5	1.08	3.79	0.54	2.20
	6	1.07	3.69	0.56	2.37
	1	1.16	4.34	0.35	1.32
	2	1.13	4.06	0.37	1.42
3	3	1.11	3.99	0.39	1.55
	4	1.14	4.06	0.41	1.66
	5	1.14	4.03	0.42	1.78
	6	1.09	3.89	0.43	1.89

Table 2. Performance evaluation of proposed method with different C

Setting				Max disparity of 192				Max disparity of 384
S	C	Loss	Dimensionality	E_EP /pixel	E_D1 /%	t_Run /s	GPU /GB	E_EP /pixel	E_D1 /%
1	1	L1	Tri	1.02	3.41	0.75	2.16	1.33	3.64
2	3	L1	Tri	1.05	3.63	0.51	1.85	1.38	3.90
2	3	CE	Bi	1.04	2.71	0.45	1.50	1.29	2.92
2	3	CE+L1	Bi	1.04	2.69	0.45	1.56	1.28	2.87
3	3	L1	Tri	1.11	3.99	0.39	1.55	1.49	4.30
3	3	CE	Bi	1.13	2.73	0.36	1.30	1.39	3.40
3	3	CE+L1	Bi	1.12	2.75	0.36	1.28	1.37	3.00

Table 3. Performance evaluation of proposed method with different settings

Setting	K15-val		K12-val
Setting	E_EP /pixel	E_D1 /%	E_EP /pixel	E_D1 /%
S₁	0.74	2.23	0.62	2.05
S₂	0.75	2.02	0.63	1.78
S₃	0.81	2.23	0.70	1.98

Table 4. Performance evaluation of proposed method with different settings on K-val

Method	All			Noc			t_Run /s
Method	E_D1-bg /%	E_D1-fg /%	E_D1-all /%	E_D1-bg /%	E_D1-fg /%	E_D1-all /%	t_Run /s
MC-CNN-arct^[10]	2.89	8.88	3.89	2.48	7.64	3.33	67.00
DispNetC^[11]	4.32	4.41	4.34	4.11	3.72	4.05	0.06
iResNet-i2^[12]	2.25	3.40	2.44	2.07	2.76	2.19	0.12
GC-net^[13]	2.21	6.16	2.87	2.02	5.58	2.61	0.90
PSMNet^[14]	1.86	4.62	2.32	1.71	4.31	2.14	0.41
Proposed	1.72	4.19	2.13	1.51	3.57	1.85	0.39

Table 5. Performance evaluation of different methods on K15-test

Method	γ₂ /%		γ₃ /%		γ₅ /%		Mean E_EP /pixel
Method	Noc	All	Noc	All	Noc	All	Noc	All
MC-CNN-arct^[10]	3.90	5.45	2.43	3.63	1.64	2.39	0.7	0.9
DispNetC^[11]	7.38	8.11	4.11	4.65	2.05	2.39	0.9	1.0
iResNet-i2^[12]	2.69	3.34	1.71	2.16	1.06	1.32	0.5	0.6
GC-net^[13]	2.71	3.46	1.77	2.30	1.12	1.46	0.6	0.7
PSMNet^[14]	2.44	3.01	1.49	1.89	0.90	1.15	0.5	0.6
Proposed	2.35	3.04	1.42	1.90	0.89	1.19	0.6	0.6

Table 6. Performance evaluation of different methods on K12-test

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information