Stereo Matching Based on Convolutional Neural Network

Jinsheng Xiao; Hong Tian; Wentao Zou; Le Tong; Junfeng Lei

doi:10.3788/AOS201838.0815017

Journals >Acta Optica Sinica >Volume 38 >Issue 8 >Page 0815017 > Article

Acta Optica Sinica
Vol. 38, Issue 8, 0815017 (2018)

Stereo Matching Based on Convolutional Neural Network

Jinsheng Xiao^1、2、*, Hong Tian¹, Wentao Zou¹, Le Tong¹, and Junfeng Lei¹

Author Affiliations

¹ Electronic Information School, Wuhan University, Wuhan, Hubei 430072, China

² Collaborative Innovation Center of Geospatial Technology, Wuhan, Hubei 430079, China

show less

DOI: 10.3788/AOS201838.0815017 Cite this Article Set citation alerts

Jinsheng Xiao, Hong Tian, Wentao Zou, Le Tong, Junfeng Lei. Stereo Matching Based on Convolutional Neural Network[J]. Acta Optica Sinica, 2018, 38(8): 0815017 Copy Citation Text

show less

Fig. 1. Matching based on patch. (a) Patch in the left image; (b) patch in the right image

Download full size

Fig. 2. Structure of MC-CNN net before and after modification

Download full size

Fig. 3. Loss function curve before and after modification

Download full size

Fig. 4. Comparison of loss convergence curves before and after network modification

Download full size

Fig. 5. Disparity map of stereo matching obtained by proposed method (Ⅰ). (a) Original left input image; (b) original right input image; (c) disparity map; (d) ground truth; (e) error graph

Download full size

Fig. 6. Disparity map of stereo matching obtained by proposed method (Ⅱ). (a) Original left input image; (b) original right input image; (c) disparity map; (d) ground truth; (e) error graph

Download full size

Method	KITTI2012	KITTI2015	KITTI2012 on KITTI2015	KITTI2015 on KITTI2012
Original loss	2.63	3.28	4.03	3.92
Proposed loss	2.61	3.25	4.02	3.89

Table 1. Error comparison for the improvement of loss function%

Δ	Error
0.02	3.261
0.04	3.266
0.05	3.252
0.06	3.284
0.08	3.267

Table 2. Error comparison between different Δ values%

Training set	KITTI2012		KITTI2015
Training set	MC-CNN-slow	Proposed	MC-CNN-slow	Proposed
KITTI2012	2.63	2.61	4.01	4.02
KITTI2015	4.32	3.89	3.27	3.25

Table 3. [in Chinese]

Algorithm	>2 pixel	>3 pixel	>4 pixel	>5 pixel
Elas	22.72	21.07	20.23	19.66
SGM	6.28	4.98	4.14	3.57
SPSS	4.86	3.79	3.17	2.76
Fast CNN	4.98	3.07	2.39	2.03
MC-CNN-fast	4.88	3.03	2.30	1.93
MC-CNN-slow	4.28	.63	2.02	1.72
Proposed	4.36	2.61	2.00	1.70

Table 4. Error comparison of disparity with different algorithms (KITTI2012)%

Algorithm	>2 pixel	>3 pixel	>4 pixel	>5 pixel
Elas	24.09	19.21	17.59	16.82
SGM	10.03	6.93	5.47	4.48
SPSS	7.15	4.58	3.46	2.93
Fast CNN	6.78	4.38	2.56	2.03
MC-CNN-fast	7.53	4.01	2.84	2.33
MC-CNN-slow	6.38	3.27	2.37	1.97
Proposed	6.56	3.25	2.33	1.92

Table 5. Error comparison of disparity with different algorithms (KITTI2015)%

Jinsheng Xiao, Hong Tian, Wentao Zou, Le Tong, Junfeng Lei. Stereo Matching Based on Convolutional Neural Network[J]. Acta Optica Sinica, 2018, 38(8): 0815017

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information