Mouseover the table cells to see the produced
disparity map. Clicking a cell will blink the ground truth for
comparison. To change the table type, click the links below.
For more information, please see the description of new features.
Upload and evaluation are currently disabled. Published results remain available below.
Set:
test densetest sparsetraining densetraining sparse
Metric:
bad 0.5bad 1.0bad 2.0bad 4.0avgerrrmsA50A90A95A99timetime/MPtime/GD
OpenCV 2.4.8 <a href="http://docs.opencv.org/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html#stereosgbm">StereoSGBM method</a>, full variant (2 passes). Reimplementation of H. Hirschmüller's SGM method (CVPR 2006; PAMI 2008).
OpenCV's "semi-global block matching" method; memory-intensive 2-pass version, which can only handle the quarter-size images. The matching cost is the sum of absolute differences over small windows. Aggregation is performed by dynamic programming along paths in 8 directions. Post filter as implemented in OpenCV. Dense results are created by hole-filling along scanlines.
SAD window: 3x3 pixel
Truncation value for pre-filter: 63
P1/P2: 8*3*3*3/32*3*3*3
Uniqueness ratio: 10
Speckle window size: 100
Speckle range: 32
Full DP: true
C/C++; 1 core, i7@3.3 GHz
07/25/14
1
SGBM2
Q
1
26.4
180
27.9
170
12.1
187
17.8
197
13.7
149
74.5
234
14.0
163
30.3
166
26.3
170
11.0
184
64.4
232
37.9
190
25.8
178
25.3
181
29.3
180
43.7
185
OpenCV 2.4.8 <a href="http://docs.opencv.org/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html#stereosgbm">StereoSGBM method</a>, single-pass variant. Reimplementation and modification of H. Hirschmüller's SGM method (CVPR 2006; PAMI 2008).
OpenCV's "semi-global block matching" method; memory efficient single-pass version. The matching cost is the sum of absolute differences over small windows. Aggregation is performed by dynamic programming along paths in only 5 of 8 directions. Post filter as implemented in OpenCV. Dense results are created by hole-filling along scanlines.
SAD window: 3x3 pixel
Truncation value for pre-filter: 63
P1/P2: 8*3*3*3/32*3*3*3
Uniqueness ratio: 10
Speckle window size: 100
Speckle range: 32
Full DP: false
C/C++; 1 core, i7@3.3 GHz
07/25/14
2
SGBM1
Q
1
27.9
185
28.3
172
17.2
206
19.0
198
14.5
153
57.9
218
15.6
173
31.8
169
31.4
182
13.2
195
58.6
218
38.6
192
27.0
188
25.9
183
31.4
185
59.7
213
Heiko Hirschmüller. <a href="http://www.dlr.de/rmc/rm/en/desktopdefault.aspx/tabid-9389/16104_read-39811/">Stereo processing by semi-global matching and mutual information</a>. CVPR 2006; PAMI 30(2):328-341, 2008.
The images are Census transformed and the Hamming distance is used as pixelwise matching cost. Aggregation is performed by a kind of dynamic programming along 8 paths that go from all directions through the image. Small disparity patches are invalidated. Interpolation is also performed along 8 paths.
OpenCV 2.4.8 <a href="http://docs.opencv.org/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html#stereosgbm">StereoSGBM method</a>, single-pass variant. Reimplementation and modification of H. Hirschmüller's SGM method (CVPR 2006; PAMI 2008).
OpenCV's "semi-global block matching" method; memory efficient single-pass version. The matching cost is the sum of absolute differences over small windows. Aggregation is performed by dynamic programming along paths in only 5 of 8 directions. Post filter as implemented in OpenCV. Dense results are created by hole-filling along scanlines.
SAD window: 3x3 pixel
Truncation value for pre-filter: 63
P1/P2: 8*3*3*3/32*3*3*3
Uniqueness ratio: 10
Speckle window size: 100
Speckle range: 32
Full DP: false
C/C++; 1 core, i7@3.3 GHz
07/25/14
4
SGBM1
F
3
28.4
187
43.5
215
9.09
162
13.6
173
25.9
194
82.0
238
14.4
168
43.4
198
30.3
180
5.98
152
59.3
221
45.8
209
28.5
195
24.9
178
20.1
161
45.9
190
OpenCV 2.4.8 <a href="http://docs.opencv.org/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html#stereosgbm">StereoSGBM method</a>, single-pass variant. Reimplementation and modification of H. Hirschmüller's SGM method (CVPR 2006; PAMI 2008).
OpenCV's "semi-global block matching" method; memory efficient single-pass version. The matching cost is the sum of absolute differences over small windows. Aggregation is performed by dynamic programming along paths in only 5 of 8 directions. Post filter as implemented in OpenCV. Dense results are created by hole-filling along scanlines.
SAD window: 3x3 pixel
Truncation value for pre-filter: 63
P1/P2: 8*3*3*3/32*3*3*3
Uniqueness ratio: 10
Speckle window size: 100
Speckle range: 32
Full DP: false
C/C++; 1 core, i7@3.3 GHz
07/28/14
6
SGBM1
H
2
23.8
169
32.9
180
10.8
179
13.6
172
16.2
156
71.2
229
12.6
145
26.6
152
23.0
161
5.83
147
53.8
211
39.2
194
25.6
177
22.8
160
18.8
152
47.4
194
Heiko Hirschmüller. <a href="http://www.dlr.de/rmc/rm/en/desktopdefault.aspx/tabid-9389/16104_read-39811/">Stereo processing by semi-global matching and mutual information</a>. CVPR 2006; PAMI 30(2):328-341, 2008.
The images are Census transformed and the Hamming distance is used as pixelwise matching cost. Aggregation is performed by a kind of dynamic programming along 8 paths that go from all directions through the image. Small disparity patches are invalidated. Interpolation is also performed along 8 paths.
Heiko Hirschmüller. <a href="http://www.dlr.de/rmc/rm/en/desktopdefault.aspx/tabid-9389/16104_read-39811/">Stereo processing by semi-global matching and mutual information</a>. CVPR 2006; PAMI 30(2):328-341, 2008.
The images are Census transformed and the Hamming distance is used as pixelwise matching cost. Aggregation is performed by a kind of dynamic programming along 8 paths that go from all directions through the image. Small disparity patches are invalidated. Interpolation is also performed along 8 paths.
Heiko Hirschmüller, Peter Innocent, and Jon Garibaldi. <a href="http://scholar.google.com/scholar?q=Real-time+correlation-based+stereo+vision+with+reduced+border+errors">Real-time correlation-based stereo vision with reduced border errors</a>. IJCV 47(1-3):229-246, 2002.
Correlation with five, partly overlapping windows on Census transformed images using Hamming distance as matching cost. A left-right consistency check ensures unique matches and filtering small disparity segments removes outliers. Interpolation is done within image rows with the lowest, valid neighboring disparity.
Census window: 7x7 pixel
Correlation window: 9x9 pixel
LR-check: on
Min. segments: 200 pixel
Interpolation: horizontal, lowest neighbor
C/SSE; 1 core, i7@3.3 GHz
07/28/14
5
Cens5
H
2
26.6
181
47.1
224
8.74
159
11.9
161
25.6
191
45.3
194
19.5
186
40.6
192
29.0
177
9.93
178
36.5
162
38.6
191
31.0
204
25.0
180
25.6
171
44.6
187
Sudipta Sinha, Daniel Scharstein, and Richard Szeliski. <a href="http://research.microsoft.com/en-us/um/redmond/groups/ivm/LPS/"> Efficient high-resolution stereo matching using local plane sweeps</a>. CVPR 2014.
A fast method for high-resolution stereo matching without exploring the full search space. Plane hypotheses are generated from sparse feature matches. Around each plane, a local plane sweep with +/- 3 disparities levels is performed to establish local disparity hypotheses via SGM using NCC matching costs. Finally, each pixel is assigned to one hypothesis using global optimization, again using SGM.
nRounds=3
The full set of parameters is listed in the paper and the supplemental materials on the <a href= "http://research.microsoft.com/en-us/um/redmond/groups/ivm/LPS/">project webpage</a>.
C++; Core2 Duo, 2 cores @ 3 GHz
08/25/14
9
LPS
H
2
19.2
158
6.14
71
5.34
116
9.24
137
7.53
111
96.0
243
12.3
143
9.61
94
9.40
103
5.18
136
92.4
244
27.4
151
24.3
164
23.0
162
10.0
105
25.6
143
Sudipta Sinha, Daniel Scharstein, and Richard Szeliski. <a href="http://research.microsoft.com/en-us/um/redmond/groups/ivm/LPS/"> Efficient high-resolution stereo matching using local plane sweeps</a>. CVPR 2014.
A fast method for high-resolution stereo matching without exploring the full search space. Plane hypotheses are generated from sparse feature matches. Around each plane, a local plane sweep with +/- 3 disparities levels is performed to establish local disparity hypotheses via SGM using NCC matching costs. Finally, each pixel is assigned to one hypothesis using global optimization, again using SGM.
nRounds=3
The full set of parameters is listed in the paper and the supplemental materials on the <a href= "http://research.microsoft.com/en-us/um/redmond/groups/ivm/LPS/">project webpage</a>.
C++; Core2 Duo, 2 cores @ 3 GHz
08/27/14
10
LPS
F
3
20.3
161
6.72
81
6.06
128
9.72
140
9.87
126
94.3
242
14.1
164
11.2
104
11.2
116
5.88
148
89.3
243
36.0
181
20.5
144
23.8
171
16.0
141
25.4
140
Kang Zhang, Jiyang Li, Yijing Li, Weidong Hu, Lifeng Sun, and Shiqiang Yang. <a href="https://github.com/rookiepig/BinaryStereo">Binary stereo matching</a>. ICPR 2012.
no post processing is used
the same with the original paper.
C/C++ single thread Intel(R) Core(TM)2 Duo CPU P7370 @ 2.00GHz
08/31/14
11
BSM
Q
1
41.5
232
59.8
240
25.8
234
27.9
220
38.9
230
60.6
221
33.3
226
46.9
208
37.3
202
26.3
229
64.8
233
51.5
219
42.6
234
45.2
237
42.8
209
66.6
223
Christos Stentoumis, Lazaros Grammatikopoulos, Ilias Kalisperakis, and George Karras. <a href="http://www.sciencedirect.com/science/article/pii/S0924271614000434">On accurate dense stereo-matching using a local adaptive multi-cost approach</a>. ISPRS Journal of Photogrammetry and Remote Sensing 91:29-49, 2014.
This approach is an adaptive local stereo-method. It is integrated into a hierarchical scheme, which exploits adaptive windows. Sub-pix disparities are estimated,but not refined.
L = 10
t = 35
medianK = [3 3]
censusK = [9 7]
lambda = 45;
Matlab;1 Xeon core@2.5 GHz
09/10/14
12
LAMC_DSM
H
2
26.0
178
55.8
236
11.9
185
14.3
178
18.3
165
44.0
188
18.3
181
39.9
190
29.5
179
6.67
155
31.1
151
34.5
175
28.8
196
26.3
184
30.1
183
35.7
162
Nils Einecke and Julian Eggert. <a href="https://www.researchgate.net/profile/Julian_Eggert/publication/224212630_A_Two-Stage_Correlation_Method_for_Stereoscopic_Depth_Estimation/links/562fc0af08ae02b5739a2123.pdf">A two-stage correlation method for stereoscopic depth estimation</a>. DICTA 2010.
Block-matching stereo with Summed Normalized Cross-Correlation (SNCC) measure. Standard post-processed is applied, including a left-right check, error island removal (region growing), hole-filling and median filtering.
SNCC (first stage 3x3, second stage 11x11)
min correlation threshold = 0.3
region growing threshold = 2.5 disparity
min region size = 200 pixel
median filter = 1x5 and 5x1
C+IPP; 1 i5 core @3.1 GHz
09/18/14
13
SNCC
H
2
21.9
164
48.6
228
6.98
141
9.79
141
25.7
192
46.0
197
12.4
144
36.8
184
16.6
143
7.25
159
23.1
128
34.2
174
26.7
185
21.8
154
19.9
159
28.4
150
Jedrzej Kowalczuk, Eric Psota, and Lance Perez. <a href="http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6213098">Real-time stereo Matching on CUDA using an iterative refinement method for adaptive support-weight correspondences</a>. IEEE TCSVT 23(1):94-104, 2013.
Efficient two-pass aggregation with census/gradient cost metric, followed by iterative cost penalization and disparity re-selection to encourage local smoothness of disparities.
census window size = 9 x 7
max census distance = 38.03
max gradient difference = 2.51
census/gradient balance = 0.09
aggregation window size = 33 x 33
aggregation range parameter = 23.39
aggregation spatial parameter = 7.69
refinement window size = 65 x 65
refinement range parameter = 11.30
refinement spatial parameter = 17.20
cost penalty coefficient = 0.0023
median filter window size = 3 x 3
3 iterations of refinement
confidence threshold of 0.1 for sparse maps
CUDA C++, NVIDIA GeForce TITAN Black
10/07/14
14
IDR
H
2
18.1
148
37.5
196
4.08
67
7.49
122
23.3
183
40.6
174
12.8
146
24.5
144
11.3
117
5.46
143
33.1
156
26.0
144
21.5
152
21.7
153
15.3
138
21.2
131
Mikhail Mozerov and Joost van de Weijer. <a href="http://www.cvc.uab.es/~mozerov/Stereo/index.htm">Accurate stereo matching by two-step energy minimization</a>. IEEE TIP 24(3):1153-1163, 2015.
In stereo matching cost filtering methods and energy minimization algorithms are considered as two different techniques. Due to their global extend energy minimization methods obtain good stereo matching results. However, they tend to fail in occluded regions, in which cost filtering approaches obtain better results. In this paper we intend to combine both approaches with the aim to improve overall stereo matching results.
We propose to perform stereo matching as a two-step energy minimization algorithm. We consider two MRF models: a fully connected model defined on the complete set of pixels in an image and a conventional locally connected model. We solve the energy minimization problem for the fully connected model, after which the marginal function of the solution is used as the unary potential in the locally connected MRF model.
Only gradient component (6D vector) of color images is used
Cevahir Cigla and Aydin Alatan. <a href="http://www.sciencedirect.com/science/article/pii/S0923596513000490">Information permeability for stereo matching</a>. Signal Processing: Image Communication 28(9), 2013.
A local matching technique utilizing SAD+Census cost measure and a recursive edge-aware aggregation through Successive Weighted Summation. Occlusion handling is provided via left-right cross check and a background favored filling.
smoothness parameter sigma = 24
5x5 Census window, Census weight=0.7, SAD weight=0.3, occlusion threshold=2
C/C++ one i5 core @3.10 GHz
04/09/15
17
PFS
F
3
32.2
204
65.1
242
29.4
240
12.1
162
50.0
240
70.8
228
28.2
212
44.6
200
23.1
162
7.85
166
37.0
166
37.7
188
27.9
192
36.0
223
19.8
158
35.7
163
Eric Psota, Jedrzej Kowalczuk, Mateusz Mittek, and Lance Perez. <a href="http://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Psota_MAP_Disparity_Estimation_ICCV_2015_paper.pdf">MAP disparity estimation using hidden Markov trees</a>. ICCV 2015.
Efficient message passing on minimum spanning trees to acquire the maximum a posteriori disparity estimates.
Gradient Coefficient 0.289
Gradient Truncation 1.0
Census Coefficient 0.014
Census Truncation 20
Census Window 7x9
Matlab; i7 core @2.7GHz
04/17/15
18
TMAP
H
2
16.9
141
20.2
149
4.94
102
8.13
128
12.8
145
30.0
151
11.7
138
27.9
160
20.4
155
5.09
135
31.5
152
23.1
136
20.9
147
19.0
137
18.8
153
18.0
120
Chi Zhang, Zhiwei Li, Yanhua Cheng, Rui Cai, Hongyang Chao, and Yong Rui. <a href="http://openaccess.thecvf.com/content_iccv_2015/papers/Zhang_MeshStereo_A_Global_ICCV_2015_paper.pdf">MeshStereo: A global stereo model with mesh alignment regularization for view interpolation</a>. ICCV 2015.
This approach triangulates the polygonized SLIC segmentations of the input images and optimizes a lower-layer MRF on the resulting set of triangles defined by photo consistency and normal smoothness. The lower-layer MRF is solved by a quadratic relaxation method which iterates between PatchMatch and Cholesky Decomposition. The lower-layer MRF is assisted by a upper-layer MRF defined on the set of triangle vertices which exploits local 'visual complexity' cues and encourages smoothness of the vertices' splitting properties. The two layers interact through an Alignment energy term which requires triangles sharing a non-split vertex to have their disparities agree on that vertex. Optimization of the whole model is iterated between optimizations of the two layers till convergence where the upper-layer can be solved in closed form.
omega=0.2
tau_grad=15
theta goes from 0 to 100 by smoothstep function in ten iterations
gamma1=30
gamma2=60
gamma3=0.8
i7-2600 3.40GHz 8 cores, C++
04/19/15
19
MeshStereo
H
2
13.2
125
5.90
68
4.88
98
10.8
153
12.9
146
10.6
80
11.0
133
12.2
107
9.01
97
5.39
141
27.4
139
23.5
138
17.7
134
21.0
150
15.4
139
20.9
129
Jure Zbontar and Yann LeCun. <a href="http://jmlr.org/papers/volume17/15-535/15-535.pdf">Stereo matching by training a convolutional neural network to compare image patches</a> (accurate architecture). JMLR 17:1-32, 2016. <a href="http://jzbontar.github.io/mc-cnn/">Code</a>.
Compute the matching cost with a convolutional neural network (<b>accurate architecture</b>). Then apply cross-based cost aggregation, semiglobal matching, left-right consistency check, median filter, and a bilateral filter.
DETAILS:
The network is similar to the one described in our CVPR paper differing only in the values of some hyperparameters. The input to the network are two 11 x 11 image patches. Five convolutional layers with 3 x 3 kernels and 112 feature maps extract feature vectors from the input image patches. The two 112-length feature vectors are concatenated into a 224-length vector which is passed through three fully-connected layers with 384 units each. The final (fourth) fully-connected layer projects the output to a single number---the matching cost. One important addition was the use of data augmentation techniques to increase the size of the training set. We tried to use as much training data as possible. Therefore we combined all of the 2001, 2003, 2005, 2006, and 2014 Middlebury datasets obtaining 60 image pairs. For the newer datasets (2005, 2006, and 2014) we also used several illumination and exposure settings.
See paper.
NVIDIA GTX TITAN Black
08/28/15
20
MC-CNN-acrt
H
2
8.08
80
5.59
62
4.55
89
5.96
94
2.83
49
11.4
88
5.81
84
8.32
86
8.89
93
2.71
79
16.3
101
14.1
82
13.2
100
13.0
71
6.40
72
11.1
80
Andreas Geiger, Martin Roser, and Raquel Urtasun. <a href="https://link.springer.com/chapter/10.1007/978-3-642-19315-6_3">Efficient large-scale stereo matching</a>. ACCV 2010. <a href="http://www.cvlibs.net/software/libelas/">Code</a>.
A prior disparity image is calculated by matching a set of reliable support points and triangulating between them. A maximum a-posterior approach refines the disparities. The disparities for the left and right image are checked for consistency and disparity segments below a size of 50 pixels removed. (Improved results as of 9/14/2015 due to bug fix in color-to-gray conversion.)
Standard parameters of Libelas as provided with the MiddEval3-SDK.
C++/SSE; 1 core, i7@3.6 GHz
09/14/15
21
ELAS
F
3
32.3
205
50.9
229
9.17
163
11.0
156
33.0
217
88.2
240
18.3
180
47.3
210
26.8
171
11.7
189
41.7
173
37.4
186
23.7
161
28.8
194
63.0
234
42.8
180
Yichao Li and Suping Fang. <a href="https://www.sciencedirect.com/science/article/pii/S0030402618316498">Removal-based multi-view stereo using a window-based matching method</a>. Optik 178:1318-1336, 2019.
Removed erroneous corresponding points from stereo pair and only correct corresponding points are kept which are obtained by NCC.
see paper
C++; core i3-2328M @2.20GHZ
09/28/15
22
R-NCC
F
3
48.4
238
26.2
166
14.8
197
30.2
226
30.9
209
72.9
232
41.6
240
77.7
243
64.1
242
27.4
232
59.1
220
71.9
243
50.9
240
33.9
218
78.2
243
80.8
239
Ang Li, Dapeng Chen, Yuanliu Liu, and Zejian Yuan. <a href="http://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Li_Coordinating_Multiple_Disparity_CVPR_2016_paper.pdf">Coordinating multiple disparity proposals for stereo computation</a>. CVPR 2016.
The method generates multiple proposals on absolute and relative disparities from multi-segmentations. The proposals are coordinated by point-wise competition and pairwise collaboration within a MRF model. During inference, a dynamic programming is performed in different directions with various step sizes.
4 i7 cores @3.4 GHz
10/13/15
23
MDP
H
2
12.6
120
14.4
131
4.99
105
10.6
152
10.7
130
27.2
146
8.11
112
12.5
109
8.07
90
4.27
126
30.4
149
20.5
128
12.6
94
17.8
124
13.4
129
17.3
115
Jonathan Barron and Ben Poole. <a href="https://people.eecs.berkeley.edu/~barron/BarronPooleECCV2016.pdf">The fast bilateral solver</a>. ECCV 2016. <a href="https://github.com/poolio/bilateral_solver">Code</a>.
We post-process the depth maps produced by Zbontar & LeCun's MC-CNN technique. We use a domain transform to compute an edge-aware variance measure of our confidence in the depth map, and then run our robust bilateral solver on that depth map and confidence with a Geman-McClure loss function.
The MC-CNN is computed using the publicly-available implementation (https://github.com/jzbontar/mc-cnn) which using the GPU, and the robust bilateral solver is computed using our CPU implementation which does not use the GPU, and is written in vanilla C++.
Intel(R) Xeon(R) CPU E5-1650 0 @ 3.20GHz, 6 cores; 32 GB RAM; NVIDIA GTX TITAN X
11/03/15
24
MC-CNN+RBS
H
2
8.42
84
6.05
69
5.16
109
6.24
101
3.27
60
11.1
84
6.36
92
8.87
90
9.83
106
3.21
97
15.1
93
15.9
95
12.8
96
13.5
78
7.04
82
9.99
67
Min-Gyu Park and Kuk-Jin Yoon. <a href="https://www.researchgate.net/profile/Min_Gyu_Park2/publication/331014567_As-planar-as-possible_depth_map_estimation/links/5c9229ba92851cf0ae89f904/As-planar-as-possible-depth-map-estimation.pdf">As-planar-as-possible depth map estimation</a>. CVIU 181:50-59, 2019.
We exploit local and dominant plane hypotheses to estimate as-planar-as-possible disparity maps.
beta = 1,000, lambda_aps = 28
lambda_per = 10
Matlab+Mex; i7 core 3.5GHz; 4 cores
05/28/16
34
APAP-Stereo
H
2
7.26
67
5.43
58
4.91
100
5.11
74
5.17
87
21.6
129
6.99
97
4.31
52
4.23
60
3.24
98
14.3
90
9.78
49
7.32
56
13.4
75
6.30
69
8.46
51
Xu Huang, Yongjun Zhang, and Zhaoxi Yue. <a href="https://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/III-3/67/2016/isprs-annals-III-3-67-2016.pdf">Image-guided non-local dense matching with three-steps optimization</a>. ISPRS Congress 2016.
This paper proposes a new image-guided non-local dense matching method with a three-step optimization based on the combination of image-guided methods and energy function-guided methods.
Cost Computation:
Window Size: 5
Weighting Coefficient: 0.3
Truncation Threshold (Census): 15
Truncation Threshold (HOG): 1
Image-guided Non-local Matching:
Smooth Term: 6
Penalty Term P1: 0.3
Penalty Term P2: 3
Disparity Interpolation:
Truncation Threshold: 5
Smooth Term: 3
Penalty Term P1: 3
Penalty Term P2: 30
Function Base: 5
C/C++ 1 i7 core @3.2 GHz
12/18/15
25
INTS
H
2
14.5
133
20.2
148
4.52
84
8.62
131
11.6
133
29.5
148
10.7
131
16.4
124
10.3
110
4.69
127
27.6
141
22.5
133
20.7
145
20.5
144
11.5
119
24.9
139
Kyung-Rae Kim and Chang-Su Kim. <a href="http://ieeexplore.ieee.org/abstract/document/7532996/">Adaptive smoothness constraints for efficient stereo matching using texture and edge information</a>. ICIP 2016.
An efficient stereo matching algorithm, which applies adaptive smoothness constraints using texture and edge information, is proposed in this work. First, we determine non-textured regions, on which an input image yields flat pixel values. In the non-textured regions, we penalize depth discontinuity and complement the primary CNN-based matching cost with a color-based cost. Second, by combining two edge maps from the input image and a pre-estimated disparity map, we extract denoised edges that correspond to depth discontinuity with high probabilities. Thus, near the denoised edges, we penalize small differences of neighboring disparities.
The method uses the MC-CNN code for the matching cost computation only.
omega = 0.2
sigma_NT = 2.0
P_1 = 1.0
P_2 = 45
i7 core @2.2 GHz + Geforce GTX TITAN X
01/19/16
26
NTDE
H
2
7.44
70
5.72
65
4.36
78
5.92
91
2.83
49
10.4
74
5.71
80
5.30
65
5.54
73
2.40
67
13.5
84
14.1
80
12.6
95
13.9
83
6.39
71
12.2
93
Jure Zbontar and Yann LeCun. <a href="http://jmlr.org/papers/volume17/15-535/15-535.pdf">Stereo matching by training a convolutional neural network to compare image patches</a> (fast architecture). JMLR 17:1-32, 2016. <a href="http://jzbontar.github.io/mc-cnn/">Code</a>.
Compute the matching cost with a convolutional neural network (<b>fast architecture</b>). Then apply cross-based cost aggregation, semiglobal matching, left-right consistency check, median filter, and a bilateral filter.
See paper.
NVIDIA GTX TITAN X
01/26/16
27
MC-CNN-fst
H
2
9.47
93
7.35
93
5.07
106
7.18
118
4.71
82
16.8
114
8.47
116
7.37
82
6.97
83
2.82
82
20.7
118
17.4
108
15.4
117
15.1
97
7.90
94
12.6
95
Radouane Ait-Jellal, Manuel Lange, Benjamin Wassermann, Andreas Schilling, and Andreas Zell. <a href="https://ieeexplore.ieee.org/abstract/document/7989019">LS-ELAS: line segment based efficient large scale stereo matching</a>. ICRA 2017.
Our approach is an extension of the ELAS (from Geiger et al.) algorithm. We extract edges and sample our candidate support points along them. For every two consecutive valid support points we create a (straight) line segment. We force the triangulation to include the set of line segments (constrained Delaunay) for a better preservation of the disparity discontinuity at the edges.
Parameters as in the original ELAS algorithm.
For sampling candidate support points along the edge segments:
Adaptive sampling activated:
step = ceil(sqrt(img_diag)*0.5);
sampler(sqrt(step) / 2, step / 2, step / 2);
Jean-Charles Bricola, Michel Bilodeau, and Serge Beucher. <a href="https://hal.archives-ouvertes.fr/hal-01330139/file/article_mpsv.pdf">Morphological processing of stereoscopic image superimpositions for disparity map estimation</a>. HAL archives, hal-01330139f, 2016.
The computation of the sparse disparity maps is achieved by means of a 3D diffusion of the costs contained in the disparity space volume. The watershed segmentations of the left and right views control the diffusion process and valid measurements are obtained by cross-checking.
The estimation of the dense disparity maps uses the sparse measurements as control points and is driven by a 3D watershed separating the disparity space volume into foreground and background pixels.
Python; 1 i5 core @2.7Ghz
03/15/16
29
MPSV
Q
1
43.5
234
58.8
238
33.9
243
34.2
235
37.9
229
52.4
208
30.8
221
56.8
232
51.0
230
30.6
235
56.9
214
51.5
220
44.6
236
43.4
234
44.2
212
54.2
204
Mozhdeh Shahbazi, Gunho Sohn, Jerome Theau, and Patrick Menard. <a href="https://pdfs.semanticscholar.org/782d/35a9582123beeeb5395ad5f715a6ff5dea7e.pdf">Revisiting intrinsic curves for efficient dense stereo matching</a>. ISPRS Congress 2016.
No post processing (no filtering, no hole-filling, no interpolation) performed.
The concepts of intrinsic curves were revisited and used for:
- disparity search space reduction, resulting in 83% reduction of the disparity range (individually for each pixel) directly from the original resolution of the image without needing hierarchical search
- reducing the ambiguities due to occluded pixels by integrating occlusion clues explicitly into the global energy function as a soft prior
The final energy minimization was done using semi global approach along eight paths.
Matching (data) cost = census transform 7*9
Occlusion cost= from intrinsic curves curvature
C++; 1 i7-2760QM CPU @ 2.4 GHz
04/03/16
30
ICSG
F
3
45.6
237
69.7
244
19.1
211
21.3
204
43.6
235
77.6
237
36.9
233
65.3
239
40.4
208
20.3
211
53.6
209
58.7
237
46.5
238
47.1
240
60.7
233
79.1
236
Simon Hadfield, Karel Lebeda, and Richard Bowden. <a href="https://www.sciencedirect.com/science/article/pii/S1077314216301096">Stereo reconstruction using top-down cues</a>. CVIU 157:206-222, 2017.
Incorporating cues from top-down (holistic) scene understanding into existing bottom-up stereo reconstruction techniques (<a href="http://ttic.uchicago.edu/~ayanc/consensus/">CoR - Chakrabarti et al. CVPR 2015</a>).
Learned weightings (from 2006 dataset) for High Level Scene Cues. Default parameters for CoR. Images with max disp > 256 were downsampled before the SGM step of CoR.
Matlab and C++. single E5 core at 2.4GHz
04/24/16
31
HLSC_cor
H
2
26.0
179
26.5
167
15.2
200
21.0
202
20.5
175
35.7
165
23.4
197
33.1
174
35.0
195
11.9
190
39.1
167
34.2
173
25.2
175
32.8
212
28.3
178
22.7
134
Hongyang Xue and Deng Cai. <a href="https://arxiv.org/pdf/1601.03890.pdf">Stereo matching by joint energy minimization</a>. arXiv:1601.03890, 2016.
1 i7 Core @3.5GHz
04/27/16
32
JEM
Q
1
37.2
218
35.7
188
27.9
239
30.6
228
33.2
219
43.0
186
31.4
224
49.5
215
47.3
219
26.5
231
49.6
196
46.0
211
35.7
218
30.8
204
37.5
197
55.8
209
Lincheng Li, Shunli Zhang, Xin Yu, and Li Zhang. <a href="https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7744590">PMSC: PatchMatch-based superpixel cut for accurate stereo matching</a>. IEEE TCSVT 28(3):679-692, 2016.
A 3D label based method with global optimization at pixel level. A bilayer matching cost is employed by first matching small square windows then aggregate on large irregular windows. Global optimization is carried out by fusing candidate proposals, which are generated from our specific superpixel structure.
Luis Horna and Robert Fisher. <a href="https://www.research.ed.ac.uk/portal/files/29400733/visapphorna.pdf">3D plane labeling stereo matching with content aware adaptive windows</a>. VISAPP 2017.
Matlab, core i5, 4 cores+ 2xGTX 970
07/03/16
35
LPU
H
2
10.4
103
11.4
118
3.18
46
8.10
127
6.08
97
20.9
127
8.24
114
6.94
80
4.00
55
4.04
118
33.9
159
16.9
105
15.2
115
17.8
123
9.12
99
11.6
89
Dexmont Pena and Alistair Sutherland. <a href="https://www.researchgate.net/profile/Dexmont_Pena_Carrillo2/publication/315267963_Disparity_Estimation_by_Simultaneous_Edge_Drawing/links/59cd5ccaaca272b0ec14fbc7/Disparity-Estimation-by-Simultaneous-Edge-Drawing.pdf">Disparity estimation by simultaneous edge drawing</a>. ACCV 2016 Workshop on 3D modelling and applications.
CRT edge point matching + Edge based disparity propagation
TBD
C++; 1 i7 core @ 2.1Ghz
08/31/16
36
SED
F
3
63.4
244
54.3
234
22.4
227
72.9
245
64.5
245
71.4
230
42.5
241
80.1
244
67.9
243
49.8
244
79.6
238
74.4
244
65.4
245
55.1
244
86.1
245
91.6
244
Shuangli Zhang, Weijian Xie, Guofeng Zhang, Hujun Bao, and Michael Kaess. <a href="http://frc.ri.cmu.edu/~kaess/pub/Zhang17icra.pdf">Robust stereo matching with surface normal prediction</a>. ICRA 2017.
We propose a method to combine the predicted surface normal constraint by deep learning. With the selected reliable disparities from stereo matching method and effective edge fusion strategy, we can faithfully convert the predicted surface normal map to a disparity map by solving a least squares system which maintains discontinuity. We use the raw matching cost of MC-CNN.
See paper.
Matlab; 1 i5-4590 CPU@3.3 GHz
09/13/16
37
SNP-RSM
H
2
8.75
89
5.46
61
4.85
97
6.50
106
3.37
62
10.4
76
7.31
101
8.73
89
9.37
102
3.58
108
14.3
89
14.7
88
14.9
114
12.8
67
10.1
107
10.8
77
Haesol Park and Kyoung Mu Lee. <a href="https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7778222">Look wider to match image patches with convolutional neural network</a>. IEEE Signal Processing Letters 24(12):1788-1792, 2017.
A novel pooling scheme is used to train a matching cost function with a CNN. It widens the size of receptive field effectively without losing the fine details.
The overall post-processing pipeline is kept almost same as the original MC-CNN-acrt, except that the parameter settings are changed as follows:
cbca_num_iterations_1 = 0, cbca_num_iterations_2 = 1, sgm_P1 = 1.3, sgm_P2 = 17.0, sgm_Q1 = 3.6, sgm_Q2 = 36.0, and sgm_V = 1.4.
Torch; the Intel core i7 4790K
CPU and a single Nvidia Geforce GTX Titan X GPU
10/19/16
38
LW-CNN
H
2
7.04
62
4.65
42
3.95
62
5.30
77
2.63
41
11.2
87
5.41
75
4.32
53
4.22
58
2.43
68
12.2
73
13.4
76
13.6
104
14.8
94
4.72
46
12.0
92
Sonam Nahar and Manjunath Joshi. <a href="https://ipsjcva.springeropen.com/articles/10.1186/s41074-016-0013-0">A learned sparseness and IGMRF-based regularization framework for dense disparity estimation using unsupervised feature learning</a>. IPSJ CVA 9:2, 2017.
An energy minimization framework for disparity estimation where energy function consists of intensity matching cost, feature matching cost, IGMRF prior and sparsity priors.
Manually set
MATLAB 2014 @2.22 Ghz
10/23/16
39
SIGMRF
Q
1
64.2
245
60.0
241
33.0
242
67.9
244
63.2
244
99.5
247
39.8
238
84.8
245
82.0
245
35.2
240
95.2
245
91.5
245
58.1
243
65.8
245
55.0
228
88.6
243
Chloe LeGendre, Konstantinos Batsos, and Philippos Mordohai. <a href="https://par.nsf.gov/servlets/purl/10041310">High-resolution stereo matching based on sampled photoconsistency computation</a>. BMVC 2017.
C/C++ 6 cores Intel Core i7-5820K @3.3 GHz
11/06/16
40
SPS
F
3
19.6
159
14.2
130
12.3
189
14.9
182
12.0
139
15.8
106
19.1
183
17.4
129
15.4
135
8.23
168
30.9
150
34.8
176
30.6
202
25.3
181
28.3
177
28.0
146
Stepan Tulyakov, Anton Ivanov, and Francois Fleuret. <a href="http://openaccess.thecvf.com/content_ICCV_2017/papers/Tulyakov_Weakly_Supervised_Learning_ICCV_2017_paper.pdf">Weakly supervised learning of deep metrics for stereo reconstruction</a>. ICCV 2017.
This is a new weakly supervised method that allows to learn deep metric for stereo reconstruction from unlabeled stereo images, given coarse information about the scenes and the optical system. The deep metric architecture is similar to MC-CNN fst.
1 core 2.5 Ghz + K40 NVIDIA, Lua-Torch
11/15/16
41
MC-CNN-WS
H
2
12.1
117
14.8
134
7.20
145
11.1
158
7.62
112
15.9
108
11.8
139
11.5
106
9.01
97
3.89
113
19.7
114
20.5
127
16.3
122
16.3
112
12.1
122
18.3
123
Menglong Yang and Xuebin Lv. <a href="https://www.sciencedirect.com/science/article/pii/S092523121830537X">Learning both matching cost and smoothness constraint for stereo matching</a>. Neurocomputing 314:234-241, 2018.
C/C++; 4 cores + GTX 1080 GPU
11/16/16
42
MCSC
F
3
11.3
113
13.3
126
5.96
126
10.6
149
8.69
118
7.22
46
11.3
135
10.6
101
7.48
85
3.07
89
3.10
17
25.2
142
19.0
140
17.2
117
10.3
109
25.5
142
Ning Ma, Yobo Men, Chaoguang Men, and Xiang Li. <a href="https://www.mdpi.com/2073-8994/8/12/159/htm">Accurate dense stereo matching based on image segmentation using an adaptive multi-cost approach</a>. Symmetry 8(12):159, 2016.
This is a segmentation based stereo matching algorithm using an adaptive multi-cost approach, which is exploited for obtaining accuracy disparity maps.
Rostam Affendi Hamzah, Haidi Ibrahim, and A. H. Abu Hassan. <a href="http://www.sciencedirect.com/science/article/pii/S1047320316302450">Stereo matching algorithm based on per pixel difference adjustment, iterative guided filter and graph segmentation</a>. Journal of Visual Communication and Image Representation, 42:145-160, 2017.
Stereo matching algorithm based on edge-preserving filter at cost aggregation and image segmentation at disparity refinement stage.
iterative n =3, k segmentation =250
C++
01/15/17
44
IGF
Q
1
34.0
209
42.7
212
20.1
214
23.7
212
32.2
213
45.6
196
28.6
216
43.0
196
37.2
201
21.4
215
50.9
201
44.7
207
34.7
215
31.9
209
37.4
195
47.1
193
Lincheng Li, Xin Yu, Shunli Zhang, Xiaolin Zhao, and Li Zhang. <a href="https://doi.org/10.1364/AO.56.003411">3D cost aggregation with multiple minimum spanning trees for stereo matching</a>. Applied Optics 56(12):3411-3420, 2017.
We propose a cost aggregation method that efficiently weave together MST-based support region filtering and PatchMatch-based 3D label search. We use the raw matching cost of MC-CNN.
gamma = 50;
lambda = 10000;
tao_sim = 0.5;
C++; i7-6700K @4.0GHz + GTX TITAN X
01/24/17
45
3DMST
H
2
5.92
52
3.71
27
2.78
35
4.75
64
2.72
45
7.36
49
4.28
62
3.44
40
3.76
51
2.35
63
12.6
75
11.5
65
8.56
62
14.0
84
5.35
57
8.87
53
Daniel Scharstein, Tatsunori Taniai, and Sudipta Sinha. <a href="http://www.cs.middlebury.edu/~schar/papers/sgmp-3dv-2017.pdf">Semi-global stereo matching with surface orientation priors</a>. 3DV 2017.
C++; 3.4 GHz Xeon E5-2643 v4 processor
03/09/17
46
SGMEPi
F
3
13.9
130
6.92
86
6.71
138
9.47
138
9.72
124
11.8
92
13.6
158
10.9
103
10.6
113
5.26
137
32.8
155
26.9
147
22.7
156
22.7
159
12.0
121
21.7
132
Sebastien Drouyer, Serge Beucher, Michel Bilodeau, Maxime Moreaud, and Loic Sorbier. <a href="https://hal.archives-ouvertes.fr/hal-01484143">Sparse stereo disparity map densification using hierarchical image segmentation</a>. 13th International Symposium on Mathematical Morphology, 2017.
We used hierarchical image segmentation to refine and complete a sparse disparity map obtained using the MC-CNN method.
NVIDIA GTX TITAN X for MC-CNN.
Python, Ubuntu, 4 cores @2.7GHz for our method.
03/10/17
47
MC-CNN+TDSR
F
3
6.35
55
5.45
59
4.45
80
6.80
111
3.46
65
10.7
81
6.05
86
5.01
62
5.19
68
2.62
73
10.8
64
9.62
46
6.59
50
11.4
55
6.01
67
7.04
45
Patrick Knöbelreiter, Christian Reinbacher, Alexander Shekhovtsov, and Thomas Pock. <a href="http://openaccess.thecvf.com/content_cvpr_2017/papers/Knobelreiter_End-To-End_Training_of_CVPR_2017_paper.pdf">End-to-end training of hybrid CNN-CRF models for stereo</a>. CVPR 2017. <a href="https://github.com/VLOGroup/cnn-crf-stereo">Code</a>.
We propose a novel method for stereo estimation, combining advantages of convolutional neural networks (CNNs) and optimization-based approaches. The optimization, posed as a conditional random field (CRF), takes local matching costs and consistency-enforcing (smoothness) costs as inputs, both estimated by CNN blocks. To perform the inference in the CRF we use an approach based on linear programming relaxation with a fixed number of iterations. We address the challenging problem of training this hybrid model end-to-end. We show that in the discriminative formulation (structured support vector machine) the training is practically feasible. The trained hybrid model with shallow CNNs is comparable to state-of-the-art deep models in both time and performance. The optimization part efficiently replaces sophisticated and not jointly trainable (but commonly applied) post-processing steps by a trainable, well-understood model.
-
NVidia Titan X
03/22/17
48
JMR
H
2
12.5
119
4.09
35
3.97
63
8.44
130
6.93
106
11.1
83
13.8
159
19.5
135
19.0
152
3.66
109
17.0
105
18.2
114
18.0
136
21.0
151
7.29
86
17.8
119
Williem and In Kyu Park. <a href="https://www.sciencedirect.com/science/article/pii/S0167865518303003">Deep self-guided cost aggregation for stereo matching</a>. Pattern Recognition Letters 112:168-175, 2018.
This method utilizes a deep learning technique to perform a self-guided cost aggregation which does not require any guidance color image.
Matlab; i7-4770 @ 3.40 GHz; GTX 1080 GPU
03/23/17
49
DSGCA
Q
1
33.8
207
42.9
213
20.9
221
23.6
211
30.2
206
45.5
195
27.6
210
42.0
195
36.0
199
21.0
212
50.2
198
44.2
206
33.3
214
34.6
220
38.4
200
46.8
191
Jihao Yin, Hongmei Zhu, Ding Yuan, and Tianfan Xue. <a href="https://www.sciencedirect.com/science/article/pii/S0031320317302364">Sparse representation over discriminative dictionary for stereo matching</a>. Pattern Recognition 71:278-289, 2017. <a href="https://github.com/Homay0118/DDL_demo">Code</a>.
Matlab/C; 4 i7 cores @ 4.0 GHz
04/04/17
50
DDL
Q
1
30.1
194
44.3
217
19.4
212
25.8
216
28.3
201
42.1
180
21.1
192
37.1
185
28.7
176
21.7
216
46.8
188
36.0
180
30.3
201
28.4
191
32.7
188
37.5
168
Leonid Keselman, John Woodfill, Anders Grunnet-Jepsen, and Achintya Bhowmik. <a href="https://arxiv.org/abs/1705.05548">Intel RealSense stereoscopic depth cameras</a>. CVPR workshop CCD 2017.
Software model of Intel RealSense R200 at a "high" preset (see paper for details)
Masamichi Kitagawa, Ikuko Shimizu, and Radim Sara. <a href="https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7986850">High accuracy local stereo matching using DoG scale map</a>. IAPR MVA 2017.
Our method is local matching approach using the Guided Filter for cost aggregation. We give appropriate the Guided Filter size for each pixel in input image by the Filter Size Map computed by using the DoG Kernel.
Parameters for Filter Size Map computation:
DoGparam.scalesize = 25 (index of scale space)
DoGparam.mfsize = 1 (window size for Filter Size Map optimization)
Parameters for Guided Filter:
eps = 0.001
Parameters for cost computation:
gamma = 0.11 (Weight of cost)
Parameters for Bilateral Filter in disparity map optimization:
gamma_c = 1
gamma_d = 11
r_median = 19
Matlab, core-i5 @3.0GHz (2 cores, 4 threads)
06/14/17
52
DoGGuided
Q
1
41.4
230
45.4
220
23.6
230
30.6
229
34.6
222
52.5
209
28.3
213
59.1
237
53.8
234
26.4
230
60.6
225
54.7
231
38.3
224
35.5
222
44.5
213
72.0
232
Tatsunori Taniai, Yasuyuki Matsushita, Yoichi Sato, and Takeshi Naemura. <a href="http://taniai.space/projects/stereo/">Continuous 3D label stereo matching using local expansion moves</a>. PAMI 40(11):2725-2739, 2018. <a href="https://github.com/t-taniai/LocalExpStereo">Code</a>.
We propose local expansion moves for estimating dense 3D labels on a pairwise MRF. The data term uses a PatchMatch-like 3D slanted window formulation, where raw matching costs within a window are computed by MC-CNN-acrt and aggregated using guided image filtering. The smoothness term uses a pairwise curvature regularization term by Olsson et al. 2013.
See our paper.
C++; 4 cores i7-4770K @3.5GHz
06/22/17
53
LocalExp
H
2
5.43
46
3.65
23
2.87
38
2.98
35
1.99
30
5.59
26
3.37
50
3.48
42
3.35
43
2.05
52
10.3
60
9.75
48
8.57
63
14.4
88
5.40
58
9.55
63
Xiaoqing Ye, Jiamao Li, Han Wang, Hexiao Huang, and Xiaolin Zhang. <a href="http://ieeexplore.ieee.org/document/8047938/">Efficient stereo matching leveraging deep local and context information</a>. IEEE Access 5:18745-18755, 2017.
We propose a feature ensemble network leveraging deep convolutional neural network to perform matching cost computation and the disparity refinement. For matching cost computation, patch-based network architecture with multi-size and multi-layer pooling unit is adopted to learn cross-scale feature representations. For disparity refinement, the initial optimal and sub-optimal disparity maps are incorporated and diverse base learners are applied.
See the paper.
lua and C/C++; 6 cores + GTX 1080 GPU
10/12/17
54
FEN-D2DRR
H
2
7.23
66
4.68
43
4.11
68
5.03
72
3.03
57
8.42
55
6.05
86
4.90
61
5.32
70
3.20
96
11.5
69
14.1
80
13.4
103
13.9
82
5.06
50
14.3
106
Konstantinos Batsos, Changjiang Cai, and Philippos Mordohai. <a href="http://openaccess.thecvf.com/content_cvpr_2018/papers/Batsos_CBMV_A_Coalesced_CVPR_2018_paper.pdf">CBMV: A Coalesced bidirectional matching volume for disparity estimation</a>. CVPR 2018. <a href="https://github.com/kbatsos/CBMV">Code</a>.
We propose a robust learning-based method for stereo cost volume computation. We accomplish this by coalescing diverse evidence from a bidirectional matching process via random forest classifiers. We show that our matching volume estimation method achieves similar accuracy to purely data-driven alternatives and that it generalizes to unseen data much better. In fact, we used the same model trained on Middlebury 2014 dataset to submit to the KITTI and ETH3D benchmarks.
Mikhail Mozerov and Joost van de Weijer. <a href="https://ieeexplore.ieee.org/document/8611229">One-view-occlusion detection for stereo matching with a fully connected CRF model</a>. IEEE TIP 28(6):2936-2947, 2019. <a href="http://datasets.cvc.uab.es/OVOD/">Code</a>.
We extend the standard BP sequential technique to the fully connected CRF models with the geodesic distance affinity.
Also a new approach to the BP marginal solution is proposed that we call one-view-occlusion detection (OVOD). In contrast to the standard winner takes all (WTA) estimation, the proposed OVOD solution allows to find occluded regions in the disparity map and simultaneously improve the matching result.
As a result we can perform only one energy minimization process and avoid the cost calculation for the second view and the left-right check procedure.
All parameter settings are given in the C++ MS VS project available at the <a href="http://datasets.cvc.uab.es/OVOD/">project website</a>.
Intel(R) Xeon(R) CPU E5-1620 v4 @3.50 GHz
12/11/17
56
OVOD
H
2
8.87
91
4.74
47
3.64
54
5.51
81
4.82
83
12.8
94
6.51
94
9.91
98
9.96
108
3.13
94
16.6
103
14.8
89
14.1
108
15.4
99
6.92
79
13.2
99
Wendong Mao and Minglun Gong. <a href="https://ieeexplore.ieee.org/abstract/document/8575760">Disparity filtering with 3D convolutional neural networks</a>. CRV 2018.
Matlab 2017
02/07/18
57
DF
Q
1
56.1
242
47.7
226
27.9
238
36.1
237
46.7
237
62.5
223
50.2
245
72.4
242
69.9
244
37.8
241
88.2
242
70.0
241
52.8
242
50.2
243
77.5
242
91.8
245
Tingman Yan, Yangzhou Gan, Zeyang Xia, and Qunfei Zhao. <a href="https://ieeexplore.ieee.org/document/8661596">Segment-based disparity refinement with occlusion handling for stereo matching</a>. IEEE TIP 2019. <a href="https://github.com/TingmanYan/SDR">Code</a>.
We propose a stereo matching algorithm that directly refines the winner-take-all (WTA) disparity map by exploring its statistic significance. WTA disparity maps are obtained from the pre-computed raw matching costs of MC-CNN-acrt.
lambda: 0.3
gamma: 20
k: 30
C/C++;1 core i5-4210M @2.6GHz
02/28/18
58
SDR
H
2
7.69
72
5.41
57
4.22
73
4.20
59
2.73
46
10.2
70
5.40
74
6.40
77
5.76
75
4.72
128
11.2
65
15.4
93
13.4
102
16.5
113
5.22
55
13.0
98
Heiko Hirschmüller. <a href="http://www.dlr.de/rmc/rm/en/desktopdefault.aspx/tabid-9389/16104_read-39811/">Stereo processing by semi-global matching and mutual information</a>. CVPR 2006; PAMI 30(2):328-341, 2008. RVC 2020 baseline.
GPU implementation of SGM using Census as matching cost
p1=18
p2=32
segmentation filter=80 pixel
median filter=5x5
GeForce GTX 980
03/09/18
59
SGM_RVC
H
2
18.4
150
37.4
194
5.31
114
9.03
134
14.2
151
31.7
155
14.3
167
24.7
147
12.6
121
5.27
138
31.8
153
29.7
159
24.9
173
22.0
155
18.6
151
28.2
148
Johannes Schönberger, Sudipta Sinha, and Marc Pollefeys. <a href="http://openaccess.thecvf.com/content_ECCV_2018/papers/Johannes_Schoenberger_Learning_to_Fuse_ECCV_2018_paper.pdf">Learning to fuse proposals from multiple scanline optimizations in semi-global matching</a>. ECCV 2018.
Semi-Global Matching (SGM) uses an aggregation scheme to combine costs from multiple 1D scanline optimizations that tends to hurt its accuracy in difficult scenarios. We propose replacing this aggregation scheme with a new learning-based method that fuses disparity proposals estimated using scanline optimization. Our proposed SGM-Forest algorithm solves this problem using per-pixel classification. SGM-Forest currently ranks 1st on the ETH3D stereo benchmark and is ranked competitively on the Middlebury 2014 and KITTI 2015 benchmarks. It consistently outperforms SGM in challenging settings and under difficult training protocols that demonstrate robust generalization, while adding only a small computational overhead to SGM.
Median disparity over all training images of the <a href="http://www.robustvision.net/rvc2018.php">ROB 2018</a> stereo challenge.
This submission is a baseline for the Robust Vision Challenge <a href="http://www.robustvision.net/rvc2018.php">ROB 2018</a>. Each pixel is set to the median disparity of the pixels at the same location in the training images. No test image information is used.
03/23/18
62
MEDIAN_ROB
H
2
97.8
247
96.1
246
95.6
246
99.0
247
98.4
247
98.4
246
99.2
247
98.4
247
98.1
246
99.0
247
99.0
247
99.6
247
99.9
247
94.7
247
95.1
246
98.3
246
Average disparity over all training images of the <a href="http://www.robustvision.net/rvc2018.php">ROB 2018</a> stereo challenge.
This submission is a baseline for the Robust Vision Challenge <a href="http://www.robustvision.net/rvc2018.php">ROB 2018</a>. Each pixel is set to the average disparity of the pixels at the same location in the training images. No test image information is used.
03/23/18
63
AVERAGE_ROB
H
2
97.6
246
96.2
247
96.5
247
96.8
246
97.8
246
97.8
245
98.2
246
97.9
246
98.2
247
98.9
246
98.9
246
99.1
246
99.7
246
93.1
246
97.9
247
98.9
247
Andreas Geiger, Martin Roser, and Raquel Urtasun. <a href="https://link.springer.com/chapter/10.1007/978-3-642-19315-6_3">Efficient large-scale stereo matching</a>. ACCV 2010. <a href="http://www.cvlibs.net/software/libelas/">Code</a>. RVC 2020 baseline.
A prior disparity image is calculated by matching a set of reliable support points and triangulating between them. A maximum a-posterior approach refines the disparities. The disparities for the left and right image are checked for consistency and disparity segments below a size of 50 pixels removed.
Updated ELAS submission as a baseline for the Robust Vision Challenge (http://robustvision.net), replacing the original ELAS (H) entry.
Standard parameters as provided with the MiddEval3-SDK and the Robust Vision Challenge stereo devkit.
C++/SSE; 1 core, Intel Core i7-6700K
03/26/18
64
ELAS_RVC
H
2
27.3
182
43.4
214
12.4
190
13.9
176
23.8
184
66.4
225
20.4
189
33.0
173
20.7
157
11.0
183
43.9
179
37.5
187
26.3
179
28.7
193
38.4
201
33.3
158
Rostam Affendi Hamzah, Fauzan Kadmin, Saad Hamid, Fakhar Ghani, and Haidi Ibrahim. <a href="https://www.sciencedirect.com/science/article/pii/S0923596518300079">Improvement of stereo matching algorithm for 3D surface reconstruction</a>. Signal Processing: Image Communication 65:165-172, 2018.
Adaptive Support Weight with Guided Filter
Refer to the published article
Windows 10; i7 core @3.2GHz; 8G RAM; C/C++;
04/17/18
65
ISM
Q
1
40.8
227
42.5
211
26.4
235
34.8
236
36.1
225
44.5
191
34.4
228
56.2
229
52.7
233
25.2
225
51.0
203
52.4
223
39.7
226
33.3
214
38.8
204
75.3
234
Jia-Ren Chang and Yong-Sheng Chen. <a href="http://openaccess.thecvf.com/content_cvpr_2018/papers/Chang_Pyramid_Stereo_Matching_CVPR_2018_paper.pdf">Pyramid stereo matching network</a>. CVPR 2018. <a href="https://github.com/JiaRenChang/PSMNet">Code</a>. ROB 2018 entry by Hisao Chien Yang.
GeForce GTX 1080 Ti/PCIe/SSE2
05/01/18
66
PSMNet_ROB
Q
1
42.1
233
33.0
181
23.1
228
30.1
225
31.4
210
54.8
213
30.7
220
48.7
212
48.3
221
28.3
233
80.8
240
53.5
228
36.9
222
38.6
230
63.9
236
71.2
228
Stepan Tulyakov, Anton Ivanov, and Francois Fleuret. <a href="http://papers.nips.cc/paper/7828-practical-deep-stereo-pds-toward-applications-friendly-deep-stereo-matching.pdf">Practical deep stereo (PDS): Toward applications-friendly deep stereo matching</a>. NeurIPS 2018.
Python, GTX 1080 Ti
05/18/18
67
PDS
H
2
14.2
131
14.4
131
5.80
124
10.5
148
10.5
128
22.1
130
14.0
162
14.5
115
8.97
96
5.93
150
24.2
133
21.5
131
18.2
139
18.9
135
11.9
120
33.6
160
Tonmoy Saikia, Eddy Ilg, and Thomas Brox. <a href="https://lmb.informatik.uni-freiburg.de/Publications/2018/ISKB18/ROB18_submission.pdf">DispNet-CSS: Robust Vision submission</a>. ROB 2018 entry.
A modification of the FlowNet 2 architecture [1] for the Robust Vision 2018 Stereo Challenge.
[1] E. Ilg, N. Mayer, T. Saikia, M. Keuper, A. Dosovitskiy, and T. Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. CVPR 2017.
See paper.
GTX1070
05/22/18
68
DN-CSS_ROB
H
2
22.8
167
31.4
177
9.28
165
13.5
171
12.4
141
44.3
190
12.1
140
28.1
162
17.6
147
9.11
174
50.9
202
40.0
197
21.2
149
25.0
179
31.9
187
43.2
182
Penglei Ji, Jie Li, Hanchao Li, and Xinguo Liu. <a href="https://doi.org/10.1016/j.jvcir.2021.103238">Superpixel alpha-expansion and normal adjustment for stereo matching</a>. Journal of Visual Communication and Image Representation, 2021.
c/c++; core i7 7700@3.6GHz
05/26/18
69
NOSS_ROB
H
2
5.01
43
3.57
22
2.84
37
3.99
54
1.93
26
5.15
22
3.34
48
3.32
38
3.15
41
2.32
60
8.55
46
7.45
32
7.06
53
12.5
63
5.20
54
9.06
56
Benedikt Wiberg. Stereo matching with neural networks. Bachelors thesis, TU Munich 2018. ROB 2018 entry.
Neural Network based on a multidimensional similarity metric and Deeplab v3+
lr=1.0
Nvidia GTX 1080 TI
05/31/18
70
FBW_ROB
H
2
32.2
203
36.3
190
9.37
166
14.2
177
19.1
168
69.3
227
13.2
156
51.0
217
39.1
205
10.8
182
43.9
179
41.5
201
33.0
212
29.9
202
51.3
224
71.4
229
Zhengfa Liang, Yiliu Feng, Yulan Guo, Hengzhu Liu, Wei Chen, Linbo Qiao, Li Zhou, and Jianfeng Zhang. <a href="http://openaccess.thecvf.com/content_cvpr_2018/papers/Liang_Learning_for_Disparity_CVPR_2018_paper.pdf">Learning for disparity estimation through feature constancy</a>. ROB 2018 entry based on CVPR 2018 paper.
Using feature constancy to improve initial disparity.
https://github.com/leonzfa/iResNet
Nvidia Titan Xp
05/31/18
71
iResNet_ROB
H
2
24.8
175
23.0
156
10.2
176
14.7
180
12.4
140
25.9
142
12.9
147
28.0
161
24.9
164
11.5
187
46.6
187
38.9
193
21.4
150
27.8
187
45.6
217
66.7
224
Konstantinos Batsos, Changjiang Cai, and Philippos Mordohai. <a href="http://personal.stevens.edu/~kbatsos/">CBMV: A coalesced bidirectional matching volume for disparity estimation</a>. ROB 2018 entry based on CVPR 2018 paper.
CBMV + Local Expansion
see our paper
Intel Core i7-5820K & 3.3 GHz
05/31/18
72
CBMV_ROB
H
2
7.65
71
3.48
20
3.35
48
4.80
66
3.57
68
6.32
35
6.88
96
4.84
58
3.91
52
1.97
50
25.4
135
11.1
59
13.1
97
15.8
107
7.34
88
13.8
103
Yu Chen, Youshen Xia, and Chenwang Wu. <a href="https://www.researchgate.net/publication/330875584_A_Crop-Based_Multi-Branch_Network_for_Matching_Cost_Computation">A crop-based multi-branch network for matching cost computation</a>. CISP-BMEI 2018.
using CBMBNet to compute matching cost, CBCA and SGM to refine disparity
Haihua Lu, Hai Xu, Li Zhang, Yanbo Ma, and Yong Zhao. <a href="https://ieeexplore.ieee.org/abstract/document/8698637">Cascaded multi-scale and multi-dimension convolutional neural network for stereo matching</a>. VCIP 2018.
Numerous CNN algorithms focus on the pixel-wise matching cost computation, which is the important building block for many state-of-the-art algorithms. However, these architectures are limited to small and single scale receptive fields and use traditional methods for cost aggregation or even ignore cost aggregation. In this paper, we propose a novel architecture called cascaded multi-scale and multi-dimension network (MSMD) to take them both into consideration. Firstly, we propose a new multi-scale matching cost computation sub-network, in which two different sizes of receptive fields are implemented parallelly. In this way, the network can make the best use of both variants to balance the trade-off between the increase of receptive field and the loss of details. Furthermore, we show that our multi-dimension aggregation sub-network which contains 2D convolution and 3D convolution operations can provide rich context and semantic information for estimating an accurate initial disparity.
Python;8 cores + Nvidia titan x
06/14/18
74
MSMD_ROB
Q
1
30.9
196
26.9
169
14.6
195
20.0
200
22.6
180
33.7
159
27.8
211
43.9
199
38.4
204
21.1
213
49.5
195
40.8
199
31.8
207
31.6
205
37.5
199
43.6
184
Wendong Mao, Mingjie Wang, Jun Zhou, and Minglun Gong. <a href="https://ieeexplore.ieee.org/abstract/document/8659297">Semi-dense stereo matching using dual CNNs</a>. WACV 2019.
A robust solution for semi-dense stereo matching is presented. It utilizes two CNN models for computing stereo matching cost and performing confidence-based filtering, respectively. Compared to existing CNNs-based matching cost generation approaches, our method feeds additional global information into the network so that the learned model can better handle challenging cases, such as lighting changes and lack of textures. Through utilizing non-parametric transforms, our method is also more self-reliant than most existing semi-dense stereo approaches, which rely highly on the adjustment of parameters.
Matlab, GTX1080Ti, Lua, Python
06/27/18
75
DCNN
H
2
10.9
109
5.66
64
4.98
103
6.49
105
5.73
93
12.5
93
8.51
117
15.6
119
10.9
115
3.08
90
24.1
132
20.2
125
16.8
128
15.5
102
10.3
110
13.8
101
Julien Valentin, Adarsh Kowdle, Jonathan Barron, et al. <a href="https://dl.acm.org/citation.cfm?id=3275041">Depth from motion for smartphone AR</a>. ACM TOG 37(6):193 (Proc. of SIGGRAPH Asia), 2018.
Single core of a Mobile Phone (QualComm Snapdragon 821 Kryo @ 2.15Ghz)
07/31/18
76
MotionStereo
H
2
40.4
225
67.6
243
25.0
232
29.2
223
40.9
232
57.3
217
35.5
230
57.5
234
40.4
207
19.9
208
42.8
177
52.6
225
39.8
227
37.1
225
51.7
225
34.9
161
Kunal Swami, Kaushik Raghavan, Nikhilanj Pelluri, Rituparna Sarkar, and Pankaj Bajpai. <a href="https://arxiv.org/abs/1906.00050">DISCO: Depth inference from stereo using context</a>. ICME 2019.
Python, Tensorflow
10/10/18
77
DISCO
H
2
24.5
172
35.0
186
7.34
147
11.7
159
18.7
167
48.6
200
17.1
175
31.4
167
22.4
159
9.33
177
46.0
185
33.5
170
27.5
191
24.6
175
27.5
175
55.0
207
Zhengfa Liang, Yiliu Feng, Yulan Guo, Hengzhu Liu, Wei Chen, Linbo Qiao, Li Zhou, and Jianfeng Zhang. <a href="http://openaccess.thecvf.com/content_cvpr_2018/papers/Liang_Learning_for_Disparity_CVPR_2018_paper.pdf">Learning for disparity estimation through feature constancy</a>. CVPR 2018.
C/C++,NVIDIA Titan XP
10/29/18
79
iResNet
H
2
22.9
168
28.3
172
9.19
164
15.8
187
19.3
171
35.1
161
11.3
136
27.7
157
16.8
145
15.2
197
54.7
212
27.6
152
19.5
142
21.5
152
31.9
186
51.6
200
Congxuan Zhang, Junjie Wu, Zhen Chen, Wen Liu, Ming Li, and Shaofeng Jiang. <a href="https://www.sciencedirect.com/science/article/pii/S0923596521001211">Dense-CNN: Dense convolutional neural network for stereo matching using multiscale feature connection</a>. Signal Processing: Image Communication, 2021.
Lua/Torch7; 8 cores + GTX 1080 GPU
10/29/18
78
Dense-CNN
H
2
7.98
78
5.59
62
4.54
87
5.83
87
2.79
48
10.4
75
5.78
81
8.26
84
8.84
92
2.66
76
15.6
99
14.2
83
13.2
99
13.2
74
6.30
69
11.1
80
Chao He, Congxuan Zhang, Zhen Chen, and Shaofeng Jiang. <a href="https://doi.org/10.1109/CISP-BMEI.2017.8302008">Minimum spanning tree based stereo matching using image edge and brightness information</a>. CISP-BMEI 2017.
we propose a MST-based stereo
matching method using image edge and brightness
information due to the classical MST based methods were
used to produce the inaccurate matching weight in the
areas of image boundaries and similar color background.
λ=0.8
c++ opencv2.4.9 I7-8750 uhd630
11/07/18
80
IEBIMst
H
2
33.8
208
36.7
192
12.1
188
16.9
193
32.5
215
51.0
204
25.3
203
58.1
235
49.8
225
11.2
185
48.6
192
56.9
235
30.2
200
26.8
186
26.9
173
71.7
230
Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. <a href="http://openaccess.thecvf.com/content_CVPR_2019/papers/Yang_Hierarchical_Deep_Stereo_Matching_on_High-Resolution_Images_CVPR_2019_paper.pdf">Hierarchical deep stereo matching on high-resolution images</a>. CVPR 2019. <a href="https://github.com/gengshan-y/high-res-stereo">Code.</a> Also in RVC 2020.
Python; Tesla V100
11/08/18
81
HSM-Net_RVC
F
3
10.2
99
12.0
122
5.32
115
7.50
123
6.72
105
15.6
104
9.89
125
6.83
79
5.14
64
4.17
124
22.7
124
17.1
106
15.6
118
14.3
86
10.8
115
14.6
108
Qiong Chang and Tsutomu Maruyama. <a href="https://ieeexplore.ieee.org/abstract/document/8419688">Real-time stereo vision system: a multi-block matching on GPU</a>. IEEE Access 6:27971-27983, 2018.
We propose a real-time stereo vision system on GPU for the high-resolution images.
Ncc Size: 3x3
SAD Size: 3x3
Block Size: 3x21,21x3,9x9
Resize Factor: 2 for H-size; 4 for F-size
Bilateral Threshold: 7
GTX TITAN X
11/11/18
82
MBM
H
2
22.8
166
36.4
191
9.95
174
15.3
184
19.3
170
36.5
167
19.9
187
27.5
156
18.1
148
10.1
180
41.5
172
32.7
166
26.3
181
23.3
164
21.4
165
39.9
176
Kyung-Rae Kim, Yeong Jun Koh, and Chang-Su Kim. <a href="http://mcl.korea.ac.kr/krkim_access2018/">Multiscale feature extractors for stereo matching cost computation</a>. IEEE Access 6:27971-27983, 2018.
We propose four efficient feature extractors based on convolutional neural networks for stereo matching cost computation. Two of them generate multiscale features with diverse receptive field sizes. These multiscale features are used to compute the corresponding multiscale matching costs. We then determine an optimal cost by combining the multiscale costs using edge information. On the other hand, the other two feature extractors produce uni-scale features by combining multiscale features directly through fully connected layers. Finally, after obtaining matching costs using one of the four extractors, we determine optimal disparities based on the cross-based cost aggregation and the semiglobal matching.
Run Wang. <a href="https://github.com/rainwang1995/MasterThesis">An end to end network for stereo matching using exploiting hierarchical context information</a>. Master's thesis, HUST, 2019.
We design a full-convolutional network to generate disparity map as a regression problem. Applying pyramid pooling and skip connection to integrate hierarchical context information.
lr=0.001
Python, GTX 1080 ti GPU
01/12/19
85
EHCI_net
H
2
9.47
94
3.75
30
4.27
76
13.1
167
27.6
198
5.30
25
3.23
46
3.47
41
3.18
42
3.90
114
9.20
47
9.58
45
9.26
72
13.9
81
17.3
146
10.6
74
Wenhuan Wu, Hong Zhu, Shunyuan Yu, and Jing Shi. <a href="https://ieeexplore.ieee.org/document/8712528">Stereo matching with fusing adaptive support weights</a>. IEEE Access 7:61960-61974, 2019.
C++;Intel Core i5-6500@3.2GHZ
01/17/19
86
FASW
Q
1
28.6
189
41.7
208
18.1
208
23.1
209
27.2
196
40.6
174
19.1
184
34.9
179
28.1
175
18.5
204
40.8
171
36.4
183
29.3
197
28.4
190
31.1
184
41.0
177
Zhengfa Liang, Yulan Guo, Yiliu Feng, Wei Chen, Linbo Qiao, Li Zhou, Jianfeng Zhang, and Hengzhu Liu. <a href="https://doi.org/10.1109/TPAMI.2019.2928550">Stereo matching using multi-level cost volume and multi-scale feature constancy</a>. PAMI 2019.
Titan xp
12/18/18
84
MCV-MFC
H
2
24.8
174
26.8
168
9.63
172
14.8
181
17.4
161
54.1
212
14.2
165
26.5
151
18.2
149
16.0
199
62.6
229
28.7
157
20.9
146
24.6
176
37.4
196
48.1
197
Xianzhi Du, Mostafa El-Khamy, and Jungwon Lee. <a href="https://arxiv.org/pdf/1904.09099.pdf">AMNet: Deep atrous multiscale stereo disparity estimation networks</a>. arXiv:1904.09099, 2019.
Python; Titan XP
02/05/19
87
AMNet
Q
1
53.3
240
54.3
233
63.7
245
51.2
243
51.3
241
40.6
176
39.9
239
51.6
218
55.9
238
55.4
245
58.9
219
57.5
236
52.7
241
49.5
242
58.1
231
61.6
217
Julia Navarro and Antoni Buades. <a href="https://www.sciencedirect.com/science/article/pii/S0262885619301040">Semi-dense and robust image registration by shift adapted weighted aggregation and variational completion</a>. Image and Vision Computing, 2019.
The method comprises two main steps. First, we use adaptive support weights for local matching. Apart from the color similarity and geometric distance, the adaptive weight distribution favors pixels in the block matching with smaller cost. Besides, we use a multiscale strategy with invalidation criteria to reduce match ambiguity and computational time.
Second, a global interpolation using a variational formulation is carried out. The energy functional penalizes deviations from the local disparity estimation at different scales.
Local approach (DAWA): 23x23 squared window, beta=11, lambda=6, gamma=4, pixel precision 1/4, three scales for multiscale procedure.
Variational model: alpha=1, gamma=5, phi1=30, phi2=15.
C++; 3.33 GHz 6-Core Intel Xeon
02/15/19
88
DAWA-F
H
2
27.4
183
47.2
225
13.6
192
13.1
169
19.2
169
66.4
225
20.4
190
30.3
165
33.9
191
8.73
171
48.9
193
37.8
189
26.7
184
29.9
201
28.0
176
36.5
165
Siti Safwana Abd Razak, Mohd Azlishah Othman, and Ahmad Fauzan Kadmin. <a href="https://www.ijeat.org/wp-content/uploads/papers/v8i3/C5839028319.pdf">The effect of adaptive weighted bilateral filter on stereo matching algorithm</a>. International Journal of Engineering and Advanced Technology(IJEAT) 8(3) 2019, C5839028319.
Stereo matching process is attracted numbers of study in recent years. The process is unique and difficult due to visual discomfort occurred which contributed to effect of accuracy of disparity maps. By using multistage technique implemented most of Stereo Matching Algorithm; taxonomy by D. Scharstein and R. Szeliski, in this paper proposed new improvement algorithm of stereo matching by using the effect of Adaptive Weighted Bilateral Filter as main filter in cost aggregation stage which able contribute edge-preserving factor and robust against plain colour region. With some improvement parameters in matching cost computation stage where windows size of sum of absolute different (SAD) and thresholds adjustment was applied and Median Filter as main filter in refinement disparity map’s stage may overcome the limitation of disparity map accuracy. Evaluation on indoor datasets, latest (2014) Middlebury dataset were used to prove that Adaptive Weighted Bilateral Filter effect applied on proposed algorithm resulted smooth disparity maps and achieved good processing time.
Yuhao Xiao, Dingding Xu, Guijin Wang, Xiaowei Hu, Yongbing Zhang, Xiangyang Ji, and Li Zhang. <a href="https://dl.acm.org/doi/abs/10.1007/978-3-030-41404-7_25">Confidence map based 3D cost aggregation with multiple minimum spanning trees for stereo matching</a>. ACPR 2019.
C++;i7-6700 3.4GHz CPU;32G memory
03/09/19
90
3DMST-CM
H
2
5.47
47
4.10
36
3.37
49
2.99
36
2.95
53
7.63
51
4.55
64
3.26
37
3.95
54
2.16
54
10.2
57
8.28
40
6.37
48
13.2
72
5.86
65
9.35
60
Chunbo Cheng, Hong Li, and Liming Zhang. <a href="https://www.worldscientific.com/doi/abs/10.1142/S0219691320500605">An unsupervised stereo matching cost based on sparse representation</a>. International Journal of Wavelets, Multiresolution and Information Processing, 2021.
This paper presents a novel unsupervised stereo matching cost for stereo matching. Specifically, a novel two-branch convolutional sparse coding (CSC) is used to learn the convolution filter bank without ground truth disparity maps. Then, the sparse representations over the learned convolutional filter bank are utilized to measure the similarity between image patches, namely, the stereo matching cost can be computed by measuring the l1 distance between sparse representations of image patches.
Matlab/C/C++; 1 i5 core @2.9 GHz
04/12/19
91
TCSCSM
H
2
19.1
157
45.2
219
5.76
123
11.0
155
22.1
179
41.1
178
13.4
157
24.8
148
11.4
119
7.17
157
29.5
146
26.6
145
26.6
182
20.5
145
16.5
143
17.4
117
Sonali Patil, Tanmay Prakash, Bharath Comandur, and Avinash Kak. <a href="https://arxiv.org/pdf/1911.09800.pdf">A comparative evaluation of SGM variants (including a new variant, tMGM) for dense stereo matching</a>. arXiv:1911.09800, 2019.
Hierarchical MGM-16 where coarser level results limit per pixel disparity search range. Post-Processing at each level include Joint Bilateral Filter, Peak removal and, consistency check. The final disparity maps are interpolated using Discontinuity preserving interpolation
See Paper
C/C++; Intel(R) Xeon(R) Gold 6154 CPU @ 3.00GHz, 16 Cores
05/10/19
92
tMGM-16
F
3
17.3
142
8.70
109
6.49
134
9.82
142
20.7
176
13.7
97
13.9
161
21.8
138
16.0
138
5.57
144
26.9
137
25.4
143
28.3
193
22.1
156
14.6
135
39.6
174
Patrick Knöbelreiter and Thomas Pock. <a href="https://arxiv.org/pdf/1907.13391.pdf">Learned collaborative stereo refinement</a>. GCPR 2019.
In this work, we propose a learning-based method to denoise and refine disparity maps of a given stereo method. The proposed variational network arises naturally from unrolling the iterates of a proximal gradient method applied to a variational energy defined in a joint disparity, color, and confidence image space. Our method allows to learn a robust collaborative regularizer leveraging the joint statistics of the color image, the confidence map and the disparity map. Due to the variational structure of our method, the individual steps can be easily visualized, thus enabling interpretability of the method. We can therefore provide interesting insights into how our method refines and denoises disparity maps. The efficiency of our method is demonstrated by the publicly available stereo benchmarks Middlebury 2014 and Kitti 2015.
GTX 2080Ti
05/13/19
93
VN
H
2
14.2
132
9.69
111
9.58
170
10.9
154
7.33
109
9.54
64
13.8
160
11.3
105
11.3
118
7.17
157
27.4
140
23.3
137
24.8
170
22.8
161
14.6
136
18.4
124
Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang.
<a href="https://link.springer.com/article/10.1007/s11432-019-2803-x">InStereo2K: a large real dataset for stereo matching in indoor scenes</a>. Science China Information Sciences, 2020.
<a href="https://github.com/YuhuaXu/StereoDataset">Dataset</a>.
We have collected 2000 pairs of stereo images with high accuracy disparity maps to fine-tune the network. Our goal is to improve the generalization performance of networks.
fine-tune num: 90000; the initial learning rate: 1e-3.
We propose "DeepPruner", a real-time stereo matching algorithm, which combines the strength of deep network and search space pruning techniques. Towards this goal, we developed a differentiable PatchMatch module that allows us to discard most disparities and generates a sparse representation of the cost-volume. We then exploit this representation to learn which range to prune for each pixel. Our method achieves competitive results on KITTI / SceneFlow datasets while running in real-time at 62ms. Moreover, we obtain the first place (on overall rankings) on the Robust Vision Challenge. For more details, check out <a href="https://arxiv.org/pdf/1909.05845.pdf">our paper</a> and <a href="https://github.com/uber-research/DeepPruner">source code</a>.
GPU @ 2.5 Ghz (Python)
06/26/19
95
DeepPruner_ROB
Q
1
30.1
191
34.2
184
19.9
213
24.3
213
23.8
184
47.2
198
26.1
204
26.1
150
22.8
160
18.4
203
59.8
222
36.5
184
23.2
158
31.7
207
48.3
222
44.8
188
Xiao Song, Xu Zhao, Liangji Fang, and Hanwen Hu. <a href="https://arxiv.org/pdf/1903.01700.pdf">Edgestereo: An effective multi-task learning network for stereo matching and edge detection</a>. To appear in IJCV 2019.
C/C++, NVIDIA Titan XP
07/26/19
96
EdgeStereo
F
3
18.7
151
25.3
164
6.79
139
10.6
150
25.1
190
22.1
130
8.31
115
24.5
143
16.5
142
6.63
154
9.20
47
32.0
164
24.6
168
20.2
142
19.2
155
54.3
205
Patrick Knöbelreiter, Christian Sormann, Alexander Shekhovtsov, Friedrich Fraundorfer, and Thomas Pock. <a href="https://openaccess.thecvf.com/content_CVPR_2020/papers/Knobelreiter_Belief_Propagation_Reloaded_Learning_BP-Layers_for_Labeling_Problems_CVPR_2020_paper.pdf">Belief propagation reloaded: Learning BP layers for dense prediction tasks</a>. CVPR 2020. <a href="https://github.com/VLOGroup/bp-layers">Code</a>.
It has been proposed by many researchers that combining deep neural networks with graphical models can create more efficient and better regularized composite models. The main difficulties in implementing this in practice are associated with a discrepancy in suitable learning objectives as well as with the necessity of approximations for the inference. In this work we take one of the simplest inference methods, a truncated max-product Belief Propagation, and add what is necessary to make it a proper component of a deep learning model: We connect it to learning formulations with losses on marginals and compute the backprop operation. This BP-Layer can be used as the final or an intermediate block in convolutional neural networks (CNNs), allowing us to design a hierarchical model composing BP inference and CNNs at different scale levels. The model is applicable to a range of dense prediction problems, is well-trainable and provides parameter-efficient and robust solutions in stereo, optical flow and semantic segmentation.
We novelly formulate the scale transformation of cost volume as a Bayes inference and propose the inter-scale subnetwork to reliably and adaptively generate details under the guidance of geometric information.
we fine-tune the model pre-trained on Scene Flow for 300 epochs with the learning rate of 0.001 in the first 100 epochs and 0.0001 in the rest 1000 epochs.
GTX 1080Ti
11/11/19
98
CACA-Net
Q
1
31.7
201
21.8
154
16.0
202
21.5
205
24.5
186
38.0
169
34.6
229
38.6
187
36.3
200
19.0
206
49.4
194
40.2
198
32.1
210
33.6
217
36.0
191
58.2
212
Jialiang Wang, Varun Jampani, Deqing Sun, Charles Loop, Stan Birchfield, and Jan Kautz. <a href="https://arxiv.org/pdf/2008.11098.pdf">Improving deep stereo network generalization with geometric priors</a>. arXiv:2008.11098, 2020.
Nvidia V100 GPU
11/14/19
99
HSM-Smooth-Occ
F
3
10.8
108
11.7
119
5.62
119
8.75
132
8.39
117
15.4
100
9.60
122
8.29
85
6.11
79
4.09
120
23.6
129
20.4
126
17.0
129
13.4
77
10.2
108
16.6
113
Fengting Yang, Qian Sun, Hailin Jin, and Zihan Zhou. <a href="https://openaccess.thecvf.com/content_CVPR_2020/papers/Yang_Superpixel_Segmentation_With_Fully_Convolutional_Networks_CVPR_2020_paper.pdf">Superpixel segmentation with fully convolutional networks</a>. CVPR 2020.
Nvidia Titan-XP GPU
11/15/19
100
SPPSMNet
F
3
41.4
231
32.5
179
20.6
217
34.0
234
33.0
216
55.9
216
36.3
232
53.8
224
48.6
222
35.0
239
71.6
235
52.8
226
37.1
223
38.1
229
46.7
219
56.9
211
Huaiyuan Xu, Xiaodong Chen, Haitao Liang, Siyu Ren, Yi Wang, and Huaiyu Cai. <a href="https://ieeexplore.ieee.org/document/9055027">CrossPatch-based rolling label expansion for dense stereo matching</a>. IEEE Access, 2020.
c++, i5-7400 CPU@3.0GHz
12/19/19
101
CRLE
H
2
5.75
50
3.66
24
3.11
43
5.92
91
2.14
32
6.01
33
3.39
51
3.49
43
3.68
50
2.34
62
10.2
57
9.63
47
8.04
58
14.9
96
5.45
61
9.26
59
Weimin Yuan, Cai Meng, Xiaoyan Tong, and Zhaoxi Li. <a href="https://www.sciencedirect.com/science/article/pii/S0923596521001181">Efficient local stereo matching algorithm based on fast gradient domain guided image filtering</a>. Signal Processing: Image Communication, 2021.
Matlab R2014a;@3.4GHZ
12/30/19
102
F-GDGIF
H
2
31.6
200
37.3
193
7.72
150
16.1
189
34.9
223
35.2
163
20.1
188
55.3
226
46.6
216
9.01
173
47.6
189
52.9
227
29.4
198
29.7
200
29.4
182
61.0
216
Yuli Fu, Kaimin Lai, Weixiang Chen, and Youjun Xiang. <a href="https://ietresearch.onlinelibrary.wiley.com/doi/full/10.1049/ipr2.12071">A pixel pair based encoding pattern for stereo matching via an adaptively weighted cost</a>. IET Image Processing, 2020.
A novel encoding pattern, which is designed for the situation of radiometric distortion, is proposed. The pattern is applied for stereo matching cost function.
m=25
1 i5 core@3.0Hz
01/02/20
103
PPEP-GF
Q
1
34.6
211
42.4
210
21.7
224
24.8
214
30.8
208
44.0
189
25.1
202
45.7
205
42.1
210
20.1
210
44.1
181
43.6
204
35.2
217
32.8
213
39.3
206
55.0
208
Rafael Brandt, Nicola Strisciuglio, and Nicolai Petkov. <a href="https://arxiv.org/abs/2006.15373">MTStereo 2.0: Improved accuracy of stereo depth estimation</a>. arXiv:2006.15373, 2020.
The method is based on a Max-tree hierarchical representation of image pairs, which we use to identify matching regions along image scan-lines.
The number of color quantization levels was set to 16. α was set to 0.8. The minimum (or maximum) width of nodes to be matched was set to 0 (or 1/2 of the input image width). Matched node levels S was set to {1, 0}. The maximum neighborhood size ω_γ was set to 10. The size of the Gaussian kernel used to aggregate the cost volume was 21. The minimum confidence percentage parameter ω_Π was set to 12. In guided pixel refinement, ω_ω was set to 12% when sparse disparity maps were generated.
i7 8565U (4 cores)
01/05/20
104
MTS2
F
3
53.8
241
51.7
231
21.5
223
38.8
240
52.7
242
97.5
244
43.0
242
66.4
240
60.8
240
32.0
236
85.7
241
69.0
240
46.3
237
45.1
236
71.2
241
85.2
241
Lingyin Kong, Jiangping Zhu, and Sancong Ying. <a href="https://www.researching.cn/articles/OJ9b5c9e339a4dc8d2">Stereo matching based on guidance image and adaptive support region</a>. Acta Optica Sinica, 2020.
Matlab R2018b;Intel Core i7-3770 CPU
01/07/20
105
ADSR_GIF
Q
1
37.1
217
43.6
216
18.6
210
36.7
239
24.6
187
58.6
219
22.8
196
56.3
230
49.7
224
18.7
205
56.0
213
48.5
216
32.2
211
24.5
174
36.3
192
79.1
237
Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. <a href="https://openaccess.thecvf.com/content_CVPR_2020/papers/Gu_Cascade_Cost_Volume_for_High-Resolution_Multi-View_Stereo_and_Stereo_Matching_CVPR_2020_paper.pdf">Cascade cost volume for high-resolution multi-view stereo and stereo matching</a>. CVPR 2020.
GTX 1080Ti
02/07/20
106
CasStereo
H
2
18.8
153
23.9
161
9.01
161
10.5
147
11.7
135
74.0
233
13.1
155
10.1
99
7.86
88
4.09
120
45.4
184
25.2
141
24.4
167
17.3
120
20.5
163
44.3
186
Linghua Zeng and Xinmei Tian. <a href="https://dl.acm.org/doi/10.1145/3488719">CRAR: Accelerating stereo matching with cascaded regression and adaptive refinement</a>. ACM TOMM, 2022.
Python;GTX1080Ti
02/20/20
107
CRAR
H
2
22.0
165
23.2
159
13.5
191
16.4
191
16.3
157
21.0
128
21.8
193
28.5
163
26.9
172
10.5
181
32.5
154
32.9
167
23.3
159
23.6
169
20.3
162
37.4
167
Yaoyu Hu, Weikun Zhen, and Sebastian Scherer. <a href="https://arxiv.org/abs/1912.05012">Deep-learning assisted high-resolution binocular stereo depth reconstruction</a>. ICRA 2020.
A deep-learning model PSMNU, modified based on PSMNet, produces initial disparity and uncertainty on the down-sampled image. SGBMP performs full resolution prediction based on the initial disparity and uncertainty.
PSMNU: max disparity 256, trained on Scene Flow dataset (Flyingthings3D & Monkaa) only, without data augmentation. SGBMP: \lambda_b = 3, \lambda_s = 0.1, \lambda_d = 0.1. For the initial prediction of PSMNU, images are down-sampled to 768x1024.
Rafael Brandt, Nicola Strisciuglio, Nicolai Petkov, and Michael Wilkinson. <a href="https://www.sciencedirect.com/science/article/pii/S0167865520300581">Efficient binocular stereo correspondence matching with 1-D max-trees</a>. Pattern Recognition Letters 2020.
The algorithm is based on a hierarchical representation of image pairs which is used to restrict disparity search range. We propose a cost function that takes into account region contextual information and a cost aggregation method that preserves disparity borders.
Please see paper.
C++;4 i7 core @3.40 GHz
03/13/20
109
MTS
F
3
59.7
243
58.8
239
25.2
233
51.1
242
60.4
243
91.3
241
48.3
244
70.3
241
63.4
241
44.4
243
79.3
237
71.7
242
60.9
244
47.4
241
79.6
244
88.4
242
James Okae, Juan Du, and Yueming Hu. <a href="https://link.springer.com/article/10.1007/s11768-020-00014-y">Robust statistical approach to stereo disparity map denoising and refinement</a>. Control Theory and Technology, 2020.
Using robust statistics and probability to detect and refine outliers in disparity maps by leveraging the joint statistics of the given disparity map and its reference image.
lamda=1,r1=5,r2=25, sigma=10,tho_d=1, tho_s=4
Matlab Intel® Core™ i7-4600U CPU
05/14/20
110
SRM
H
2
13.1
124
8.50
105
7.04
143
7.86
124
7.73
115
16.1
109
7.90
110
18.4
133
18.5
151
5.03
134
22.3
123
20.0
122
18.1
137
18.5
130
11.3
118
19.3
127
Haoyu Ren, Mostafa El-Khamy, and Jungwon Lee. Stereo disparity estimation via joint supervised, unsupervised, and weakly supervised learning. ICIP 2020.
Python Tesla V40 Single GPU
05/19/20
111
SUWNet
Q
1
30.1
193
24.5
162
13.9
193
20.3
201
20.0
173
35.7
164
26.1
205
40.9
193
43.2
213
17.9
200
49.8
197
28.6
156
24.4
165
28.5
192
52.5
227
37.9
170
Haofei Xu and Juyong Zhang. <a href="https://github.com/haofeixu/aanet">AANet: Adaptive aggregation network for efficient stereo matching</a>. CVPR 2020.
Improved AANet
Please refer to the paper
PyTorch; NVIDIA V100 GPU
05/20/20
112
AANet++
H
2
15.4
136
17.5
142
8.37
156
10.2
145
9.86
125
23.9
136
9.82
124
17.7
131
15.9
137
3.25
99
18.1
110
27.1
149
16.2
121
18.4
129
20.0
160
37.7
169
Yun Xie, Shaowu Zheng, and Weihua Li. <a href="https://ieeexplore.ieee.org/document/9224149">Feature-guided spatial attention upsampling for real-time stereo matching</a>. IEEE MultiMedia, 2021.
Python, 2080TI
05/28/20
114
RTSMNet
H
2
45.6
236
47.0
223
21.9
225
31.9
231
36.4
226
75.1
235
43.9
243
58.9
236
55.3
236
32.7
237
62.2
228
56.4
234
42.2
232
39.1
231
58.0
230
59.9
214
Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, Zongyuan Ge. <a href="https://proceedings.neurips.cc/paper/2020/hash/fc146be0b230d7e0a92e66a6114b840d-Abstract.html">Hierarchical neural architecture search for deep stereo matching</a>. NeurIPS 2020.
GPU @ 2.5 Ghz (Python)
05/28/20
113
LEAStereo
H
2
7.15
64
7.56
97
4.52
84
4.62
62
4.64
79
8.83
57
5.66
78
5.86
72
6.03
78
3.30
102
13.1
80
11.3
64
10.3
77
12.1
61
7.06
83
9.90
66
Hector Vazquez, Madain Perez, Abiel Aguilar, Miguel Arias, Marco Palacios, Antonio Perez, Jose Camas, and Sabino Trujillo. <a href="https://ietresearch.onlinelibrary.wiley.com/doi/full/10.1049/cvi2.12031">Real-time multi-window stereo matching algorithm with fuzzy logic</a>. IET Computer Vision, 2021.
The propose a novel stereo matching algorithm with fuzzy logic and also implement it on a FPGA embedded system. We try to select the best window size of SAD for each pixel by leveraging fuzzy logic.
We used block the diferents size. (ex. 5, 15, 21)
Cyclone V 5CSEBA6U23I7 FPGA
06/08/20
115
MANE
H
2
30.9
197
54.7
235
11.5
183
14.6
179
29.4
205
52.6
210
26.4
207
45.1
204
31.5
185
11.5
187
42.5
174
41.8
202
33.1
213
31.6
206
34.2
189
43.5
183
Xianjing Cheng and Yong Zhao. HLocalExp-CM: Confidence map by hierarchical local expansion moves for stereo matching. To appear in Journal of Electronic Imaging, 2022.
core i5-8250@1.60GHz,C++, 8 cores
07/16/20
116
HLocalExp-CM
H
2
5.68
49
3.68
25
2.95
40
3.92
49
2.45
38
8.12
54
3.41
52
3.74
50
3.53
45
2.17
55
10.2
57
10.0
50
8.75
66
14.1
85
5.12
52
9.61
64
GA-Net reference entry as baseline for <a href="http://www.robustvision.net/">RVC 2020</a>. Original reference: Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip Torr. <a href="https://github.com/feihuzhang/GANet">GA-Net: Guided Aggregation Net for End-to-end Stereo Matching</a>. CVPR 2019.
GA-Net reference submission as baseline for the stereo benchmark of the robust vision challenge 2020.
All method credits go to the original author (Zhang et al.)
Submission by Nicolas Jourdan, TU Darmstadt, RVC 2020 team.
Trained on Middleburry, KITTI, ETH3D from the KITTI checkpoint made available in the GANet repository on Github by the original authors.
Frequency of sampling was adapted to the dataset size. Test images scaled to next multiple of 48.
We proposed a robust disparity estimation network. Our major novelty compared to existing work is a novel usage of attention, which can handle scenes with different scenarios.
The RVC submission trained by quarter-resolution Middlebury + KITTI + ETH. After validation, we will go with quarter resolution instead of half-resolution
Python 1 Telsa V40 GPU
08/10/20
119
CVANet_RVC
Q
1
31.8
202
25.6
165
14.6
196
21.7
206
22.1
178
39.8
170
28.4
214
44.7
201
47.0
218
19.1
207
50.6
200
30.4
161
24.7
169
29.3
198
52.1
226
41.6
179
Zhibo Rao, Mingyi He, Yuchao Dai, Zhidong Zhu, Bo Li, and Renjie He. <a href="https://doi.org/10.1017/ATSIP.2020.16">NLCA-Net: A non-local context attention network for stereo matching</a>. APSIPA Transactions on Signal and Information Processing, 2020. RVC 2020 entry.
Accurate disparity prediction is a hot spot in computer vision, and how to efficiently exploit contextual information is the key to improve the performance. In this paper, we propose a simple yet effective non-local context attention network (NLCANet) to exploit the global context information by using attention mechanisms and semantic information for stereo matching. First, we develop a 2D geometry feature learning (GFL) module to get a more discriminative representation by taking advantage of multi-scale features and form them into the variance-based cost volume. Then, we construct a non-local attention matching (NLAM) module by using the non-local block and hierarchical 3D convolutions, which can effectively regularize the cost volume and capture the global contextual information. Finally, we adopt a geometry refinement (GR) module to refine the disparity map to further improve the performance. Moreover, we add the warping loss function to help the model learn the matching rule of the non-occluded region. Our experiments show that (1), our approach achieves competitive results on KITTI and SceneFlow datasets in the end-point error (EPE) and the fraction of erroneous pixels (D 1 ); (2), our proposed method particularly has superior performance in the reflective regions and occluded areas.
600 * 10^-3;
200 * 10^-4;
100 * 10^-5
Nvidia v100
08/11/20
120
NLCA_NET_v2_RVC
H
2
10.4
102
11.8
120
4.12
69
6.39
103
6.44
101
19.7
124
10.9
132
14.5
115
13.2
123
3.26
101
21.2
119
14.7
87
10.1
75
14.5
92
7.17
85
11.5
88
Zhelun Shen, Yuchao Dai, and Zhibo Rao. <a href="https://openaccess.thecvf.com/content/CVPR2021/papers/Shen_CFNet_Cascade_and_Fused_Cost_Volume_for_Robust_Stereo_Matching_CVPR_2021_paper.pdf">CFNet: Cascade and fused cost volume for robust stereo matching</a>. CVPR 2021. <a href="https://github.com/gallenszl/CFNet">Code</a>.
we construct multi-scale cost volumes and fuse lower scale cost volumes and cascade higher scale ones to realize efficient and robust stereo matching
we first pre-train our model on sceneflow dataset and then finetune it jointly on Middlebury + KITTI + ETH3D
tesla V100
08/12/20
121
CFNet_RVC
H
2
10.1
98
14.4
133
7.81
151
7.12
117
6.61
103
15.5
102
7.53
103
12.3
108
11.5
120
3.02
87
10.7
63
16.6
101
10.7
79
15.4
101
10.9
116
9.01
54
Chenglong Xu, Chengdong Wu, Daokui Qu, Fang Xu, Haibo Sun, and Jilai Song. <a href="https://ieeexplore.ieee.org/document/9294069">Accurate and efficient stereo matching by log-angle and pyramid-tree</a>. IEEE TCSVT, 2021.
Combined bearings-only cost metric and Cross-regional connection based aggregation.
C++;Intel Core i7-4558U@2.8 GHz
09/03/20
122
LPSM
Q
1
39.5
223
40.0
202
20.7
219
28.3
222
34.0
221
34.3
160
23.8
199
56.7
231
52.4
232
24.9
224
36.9
165
66.3
239
40.6
230
37.5
227
46.6
218
79.3
238
Xiao Song, Guorun Yang, Xinge Zhu, Hui Zhou, Zhe Wang, and Jianping Shi. <a href="https://arxiv.org/pdf/2004.04627.pdf">AdaStereo: A simple and efficient approach for adaptive stereo matching</a>. CVPR 2021.
Pytorch; NVIDIA Titan-XP GPU
09/09/20
123
AdaStereo
F
3
13.7
129
19.6
146
7.41
148
10.6
151
14.5
152
15.7
105
7.85
108
22.6
140
9.32
101
7.00
156
9.20
47
22.4
132
14.5
111
17.8
125
14.8
137
24.2
137
Vladimir Tankovich, Christian Häne, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. <a href="https://arxiv.org/abs/2007.12140">HITNet: Hierarchical iterative tile refinement network for real-time stereo matching</a>. CVPR 2021.
The approach relies on a fast multi-resolution initialization step, differentiable 2D geometric propagation and warping mechanisms to infer slanted plane hypotheses at multiple resolutions.
see paper
Titan V
10/28/20
124
HITNet
H
2
6.46
56
6.25
75
4.67
92
4.51
60
2.17
33
6.52
38
5.18
70
2.92
34
2.66
36
2.37
64
36.7
163
9.28
44
6.27
47
11.2
52
4.61
44
9.54
62
Dominik Hirner. <a href="https://github.com/thedodo/fc-dcnn2">FC-DCNN: A densely connected neural network for stereo estimation</a>. ICPR 2020.
We propose a novel lightweight network for stereo estimation. The method uses densely connected layer structures to learn expressive features without the need of fully-connected layers or 3D convolutions. This leads to a network structure with only 0.37M parameters while still having competitive results. The post-processing consists of filtering, a consistency check and hole filling.
\eta = 6 \times 10^{-6}
python 3.6; pytorch 1.2.0; GPU RTX 2080 TI
11/10/20
125
FC-DCNN
H
2
17.9
146
21.2
152
6.52
137
9.56
139
14.1
150
31.9
157
23.4
198
23.4
141
19.7
153
5.93
150
26.9
136
22.8
134
20.0
143
19.3
138
18.2
149
23.9
136
Menglong Yang, Fangrui Wu, and Wei Li. <a href="https://ieeexplore.ieee.org/document/9614986">RLStereo: Real-time stereo matching based on reinforcement learning</a>. IEEE TIP, 2021.
Tensorflow 2.0; Nvidia GeForce Titan RTX GPU
11/12/20
126
RLStereo
H
2
27.9
186
20.5
151
15.0
199
23.5
210
26.3
195
51.5
207
35.8
231
27.1
155
23.4
163
15.6
198
63.6
231
32.3
165
21.5
151
23.2
163
44.7
214
17.4
118
Hengli Wang, Rui Fan, Peide Cai, Ming Liu, and Lujia Wang. <a href="https://ieeexplore.ieee.org/document/9811811">UnDAF: A general unsupervised domain adaptation framework for disparity or optical flow estimation</a>. ICRA 2022.
Pytorch
11/12/20
127
UnDAF-GANet
H
2
16.2
139
3.74
29
2.94
39
16.7
192
18.3
163
24.1
137
26.3
206
19.2
134
15.7
136
1.86
47
36.8
164
26.8
146
11.1
84
24.8
177
6.54
74
28.0
147
Ju He, Enyu Zhou, Liusheng Sun, Fei Lei, Chenyang Liu, and Wenxiu Sun. <a href="https://arxiv.org/pdf/2101.10811">Semi-synthesis: a fast way to produce effective datasets for stereo matching</a>. arXiv:2101.10811, 2021.
We propose a novel method namely semi-synthesis for producing large-scale on demand stereo datasets which doesn't require further fine-tuning on real datasets, i,e, we haven't fine-tuned the submission model on Middlebury training data.
Python 1 Nvidia 1080Ti GPU
11/16/20
128
SSCasStereo
H
2
15.2
135
33.6
182
5.73
121
8.13
128
12.6
143
51.1
205
8.19
113
16.7
127
5.02
63
5.70
146
48.5
190
17.3
107
16.0
119
20.1
141
12.3
124
9.25
58
Chengtang Yao, Yunde Jia, Huijun Di, Pengxiang Li, and Yuwei Wu. <a href="https://openaccess.thecvf.com/content/CVPR2021/papers/Yao_A_Decomposition_Model_for_Stereo_Matching_CVPR_2021_paper.pdf">A decomposition model for stereo matching</a>. CVPR 2021.
GTX 1080Ti GPU
11/21/20
129
DecStereo
F
3
20.2
160
19.4
145
11.9
186
15.6
185
13.5
148
23.0
134
26.7
208
13.3
111
15.1
134
7.60
162
28.3
143
30.2
160
23.4
160
17.6
121
38.9
205
38.4
171
Xianjing Cheng, Yong Zhao, Raja Soosaimarian Peter Raj, Zhijun Hu, Xiaomin Yu, and Wenbang Yang. <a href="https://www.scielo.br/j/babt/a/58BqF94vzv9z5vH6WFyCYCs/">Local PatchMatch based on superpixel cut for efficient high-resolution stereo matching</a>. BABT (Brazilian Archives of Biology and Technology), 2022.
we propose an efficient method,i.e, local PatchMatch based on superpixel cut for high-resolution stereo matching.
the number of superpixels N is 500, two iterative parameters: k_fea is set to 9 and k_SP is set to 7. The parameter γ to measure the similarity weight is set to 50 and k=8000.
i5-9400 CPU@2.90GHz, C++;
11/25/20
130
LPSC
H
2
10.7
106
5.15
53
4.23
74
5.48
80
6.38
100
16.5
111
7.84
107
9.56
93
10.3
111
4.02
117
20.2
116
19.0
118
17.7
133
18.5
131
9.73
103
18.0
121
Peng Yao and Jieqing Feng. <a href="https://www.sciencedirect.com/science/article/abs/pii/S1047320321001097">Stacking learning with coalesced cost filtering for accurate stereo matching</a>. Journal of Visual Communication and Image Representation 2021.
By leveraging Stacking Learning with Coalesced Cost Filtering to make the conventional algorithms achieve more accurate disparity estimations.
For the Random Forest, we set 10 Decision Trees, maximum depth is 25 and minimum number of samples in each node to split equal to 12.
C++, Intel Core-i7 Octa-Core CPUs;
12/22/20
131
SLCCF
H
2
8.83
90
6.97
88
4.90
99
6.05
97
4.35
76
8.89
58
5.33
72
6.29
75
5.15
65
4.80
130
13.0
78
18.1
113
17.8
135
17.7
122
6.93
80
15.4
110
Lingyin Kong, Jiangping Zhu, and Sancong Ying. <a href="https://pdfs.semanticscholar.org/99e7/4a98a0de35d5939eebcca970166ccd9becdb.pdf">Local stereo matching using adaptive cross-region based guided image filtering with orthogonal weights</a>. Mathematical Problems in Engineering, 2021.
we propose an improved cost aggregation method, in which the matching cost volume is filtered by ACR-GIF-OW
C1=15/255,C2=12/255,N=40,P=0.5
Matlab R2018b;Intel Core i7-8750H CPU
12/24/20
132
ACR-GIF-OW
Q
1
24.5
171
37.5
195
10.8
178
16.3
190
17.4
162
44.9
193
17.2
176
33.5
177
25.2
166
11.4
186
45.4
183
35.7
179
26.6
183
23.3
166
23.6
168
38.4
172
Yaoyu Hu, Wenshan Wang, Huai Yu, Weikun Zhen, and Sebastian Scherer. <a href="http://theairlab.org/orstereo">ORStereo: Occlusion-aware recurrent stereo matching for 4K-resolution images</a>. arXiv:2103.07798, 2021.
This model is trained on low-resolution data but aims at high-resolution images. It uses a recurrent module to iteratively update a coarse disparity prediction. Then a special refinement module makes a final adjustment. The recurrent update and final refine are applied in a patch-wise manner across the initial disparity.
Trained on Scene Flow, Middlebury 1/4 size, and TartanAir (sampled) datasets. Training disparity range 256 pixels, testing range over 1000 pixels.
Trained on 4 Tesla V100 GPUs. Inference on 1 Tesla V100 GPU.
03/05/21
133
ORStereo
F
3
19.1
156
38.9
201
9.97
175
9.21
135
23.3
182
42.6
184
13.0
152
18.2
132
6.63
82
4.93
131
35.4
160
33.1
168
24.1
163
23.6
168
18.2
148
26.0
144
Haiwei Sang, Zuliu Yang, Yong Xu, Xianjing Cheng, Weicheng Xiong, and Yong Zhao. <a href="https://doi.org/10.1117/1.JEI.31.1.013020">Local expansion moves for stereo matching based on random sample consensus confidence</a>. Journal of Electronic Imaging, 2022.
A stereo matching algorithm based on collaborative optimization among pixels is proposed. Based on local expansion, the matching energy function of pixels is defined by using the color and gradient features of adjacent pixels, and the cooperative competition mechanism between pixels is introduced.
iterations = 5; pmIterations = 2;
C/C++,i7-4790 CPU@3.60GHz.
03/05/21
134
LocalExp-RC
H
2
5.54
48
3.78
31
3.02
41
3.85
48
2.08
31
5.95
32
3.48
53
3.61
45
3.65
48
2.52
70
10.3
61
6.85
29
7.25
55
16.1
110
5.12
52
10.2
71
Xianjing Cheng, Yong Zhao, Wenbang Yang, Zhijun Hu, Xiaomin Yu, Haiwei Sang, and Guiying Zhang. <a href="https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/ipr2.12361">Superpixel cut-based local expansion for accurate stereo matching</a>. IET Image Processing, 2021.
i7 CPU @2.2GH,C++, 8 cores
04/22/21
135
LESC
H
2
6.78
61
4.07
34
3.46
51
3.26
41
3.36
61
9.15
62
4.08
60
4.76
57
5.21
69
2.80
81
11.7
70
13.0
72
10.2
76
17.0
115
5.52
63
12.5
94
Hao Liu, Hanlong Zhang, Xiaoxi Nie, Wei He, Dong Luo, Guohua Jiao and Wei Chen.
<a href="https://ieeexplore.ieee.org/document/9517511">Stereo matching algorithm based on two-phase adaptive optimization of AD-census and gradient fusion</a>. IEEE RCAR 2021.
In this paper, an improved AD-Census algorithm is proposed to improve the matching ratio in some special regions. The proposed algorithm contains an optimization method and three similarity metrics.
C++; 1 i5-9500 CPU @ 3.00GHz 3.00 GHz
05/09/21
136
ADSG
Q
1
24.7
173
36.3
189
11.3
180
15.6
186
20.0
172
35.9
166
18.2
179
35.2
181
27.2
173
12.2
193
42.7
176
33.8
171
26.3
180
23.3
165
27.0
174
36.6
166
Qiang Wang, Shaohuai Shi, Shizhen Zheng, Kaiyong Zhao, and Xiaowen Chu. <a href="https://ieeexplore.ieee.org/document/9197031">FADNet: A fast and accurate network for disparity estimation</a>. ICRA 2020. <a href="https://github.com/HKBU-HPML/FADNet">Code</a>.
Apply FADNet as backbone, pre-train on SceneFlow for 90 epochs, finetuned on <a href="http://www.robustvision.net/">RVC 2020</a> samples (kitti2015, eth3d, middlebury)
lr=1e-4, finetuned for 1200 epochs, lr decayed by half every 200 epochs
Pre-train FADNet++ on Scene Flow, and finetune on the mixture of KITTI, ETH3D, and Middlebury (quarter samples).
Python; GTX 2070
06/07/21
138
FADNet++
Q
1
40.2
224
25.2
163
22.1
226
33.4
233
28.8
204
54.0
211
33.6
227
46.8
207
46.8
217
23.5
221
73.4
236
47.1
215
28.5
194
37.7
228
65.4
239
71.9
231
Boitumelo Ruf, Jonas Mohrs, Martin Weinmann, Stefan Hinz, and Jürgen Beyerer. <a href="https://www.mdpi.com/1424-8220/21/11/3938">ReS2tAC - UAV-borne real-time SGM stereo optimized for embedded ARM and CUDA devices</a>. MDPI Sensors 21(11), 2021.
We propose an approach for real-time embedded stereo processing on ARM and CUDA-enabled devices, which is based on the popular and widely used Semi-Global Matching algorithm. In this, we propose an optimization of the algorithm for embedded CUDA GPUs, by using massively parallel computing, as well as using the NEON intrinsics to optimize the algorithm for vectorized SIMD processing on embedded ARM CPUs.
Rongcheng Wu, Changming Sun, Zhaoying Liu, and Arcot Sowmya. <a href="https://ieeexplore.ieee.org/document/9647104">Deep learning based stereo cost aggregation on a small dataset</a>. DICTA 2021.
4*RTX 3090
06/11/21
140
R3DCNN
H
2
33.0
206
34.2
185
15.8
201
13.4
170
41.7
233
47.9
199
22.0
194
60.1
238
57.4
239
12.6
194
40.3
170
46.4
212
26.8
186
37.0
224
19.3
156
45.2
189
Tingman Yan, Xilian Yang, Genke Yang, and Qunfei Zhao. <a href="https://ieeexplore.ieee.org/document/10198437">Hierarchical belief propagation on image segmentation pyramid</a>. IEEE TIP 2023. <a href="https://github.com/TingmanYan/HBP-ISP">Code</a>.
This paper presents an accurate and efficient hierarchical BP framework using the representation of the image segmentation pyramid (ISP). We design a hierarchy of MRF networks using the graph of superpixels at each ISP level.
[γ, ϵ, τ] = [20, 0.01, 0.6], T = 10, k = 12.
Intel Xeon E5-2620 v4 CPU and NVIDIA Titan Xp GPU
07/23/21
142
HBP_ISP
H
2
5.20
44
3.70
26
3.05
42
3.57
44
2.34
37
7.80
52
3.79
56
3.34
39
3.09
40
1.87
48
9.85
54
10.1
51
7.82
57
11.2
54
5.26
56
7.86
47
Lahav Lipson, Zachary Teed, and Jia Deng. <a href="https://arxiv.org/abs/2109.07547">RAFT-Stereo: Multilevel recurrent field transforms for stereo matching</a>. 3DV 2021. <a href="https://github.com/princeton-vl/RAFT-Stereo">Code</a>.
We introduce RAFT-Stereo, a new deep architecture for rectified stereo based on the optical flow network RAFT. We introduce multi-level convolutional GRUs, which more efficiently propagate information across the image. A modified version of RAFT-Stereo can perform accurate real-time inference.
Python RTX 6000 GPU
07/26/21
143
RAFT-Stereo
F
3
4.74
40
4.19
38
3.44
50
3.11
38
1.51
21
7.30
48
2.79
44
2.67
32
2.59
35
1.39
36
7.46
40
10.2
52
5.86
44
13.0
70
3.59
36
9.38
61
Zhengyu Huang, Theodore Norris, and Panqu Wang. <a href="https://arxiv.org/pdf/2103.03922.pdf">ES-Net: An efficient stereo matching network</a>. arXiv:2103.03922, 2021.
Python
07/14/21
141
MFN_USFDSRVC
H
2
36.7
216
31.4
178
16.1
203
22.0
207
28.5
202
42.3
181
21.1
191
52.2
220
50.7
228
21.4
214
53.6
210
53.8
229
30.7
203
31.7
208
63.7
235
60.7
215
Krishna Shankar, Mark Tjersland, Jeremy Ma, Kevin Stone, and Max Bajracharya. <a href="https://ieeexplore.ieee.org/document/9684954">A learned stereo depth system for robotic manipulation in homes</a>. IEEE Robotics and Automation Letters, 2022.
A lightweight network with dilated ResNet feature extractor, a correlation cost volume run at a low resolution, and a refinement network to get a full resolution disparity output. Sparse disparity is processed from the dense disparity using a threshold on the network confidence output and a region grower to remove suspected bad disparities.
Max disparity 512
Cost volume downsample 8x
PyTorch on Nvidia Titan RTX
08/24/21
144
MMStereo
F
3
12.7
122
27.9
171
8.71
158
8.81
133
11.7
135
26.9
145
5.82
85
20.9
137
14.6
131
4.10
122
15.4
97
16.0
98
14.2
109
13.6
80
9.71
102
7.35
46
Junda Cheng, Xin Yang, Yuechuan Pu, and Peng Guo. <a href="https://ieeexplore.ieee.org/document/9797840">Region separable stereo matching</a>. IEEE Transactions on Multimedia, 2023.
GeForce 3090
09/20/21
145
GANet-RSSM
H
2
10.6
105
11.9
121
8.54
157
6.60
107
6.26
99
16.3
110
7.10
99
15.5
118
14.5
130
2.93
85
11.3
66
16.1
100
10.8
80
16.0
109
10.7
112
11.3
85
Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. <a href="https://openaccess.thecvf.com/content/CVPR2022/papers/Xu_Attention_Concatenation_Volume_for_Accurate_and_Efficient_Stereo_Matching_CVPR_2022_paper.pdf">Attention concatenation volume for accurate and efficient stereo matching</a>. CVPR 2022.
RTX 3090
10/17/21
146
ACVNet
H
2
13.6
128
9.69
111
3.65
55
4.82
67
7.48
110
22.9
133
12.9
149
15.7
120
14.4
129
3.82
112
21.8
120
17.7
110
14.3
110
15.6
103
25.6
172
32.1
156
Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. <a href="https://openaccess.thecvf.com/content/CVPR2022/papers/Li_Practical_Stereo_Matching_via_Cascaded_Recurrent_Network_With_Adaptive_Correlation_CVPR_2022_paper.pdf">Practical stereo matching via cascaded recurrent network with adaptive correlation</a>. CVPR 2022. <a href="https://github.com/megvii-research/CREStereo">Code</a>.
Python; 8cores + GeForce RTX 2080Ti GPU
11/10/21
147
CREStereo
F
3
3.71
28
4.73
46
3.94
61
5.07
73
1.96
29
3.02
7
1.42
26
2.28
28
2.05
31
1.51
41
6.86
37
6.35
22
4.25
25
6.01
21
4.60
43
5.49
37
Shenglun Chen, Hong Zhang, Baoli Sun, Haojie Li, Xinchen Ye, and Zhihui Wang. <a href="https://doi.org/10.1016/j.imavis.2022.104614">Feature enhancement network for stereo matching</a>. Image and Vision Computing, 2023.
In stereo matching, there are two cases of poor performance: (1) the interior of large objects, and (2) object boundaries and small objects. In this work, we present feature enhancement stereo matching network to solve the problems.
None
2080ti
11/21/21
148
FENet
H
2
11.3
115
7.70
98
3.91
60
3.97
53
6.24
98
16.7
112
5.78
81
32.1
170
32.4
189
2.57
72
11.8
71
10.8
54
6.90
52
13.4
75
5.41
59
11.2
84
Junda Cheng, Gangwei Xu, Peng Guo, and Xin Yang. <a href="https://link.springer.com/article/10.1007/s11263-023-01872-0">Coatrsnet: Fully exploiting convolution and attention for stereo matching by region separation</a>. IJCV, 2023.
RTX 3090
11/21/21
149
Gwc_CoAtRS
H
2
6.50
57
6.92
86
6.82
140
4.55
61
3.48
67
5.12
21
5.80
83
4.88
59
4.96
62
2.69
78
15.3
96
12.8
70
6.40
49
10.2
45
7.13
84
8.48
52
Qibo Chen, Baozhen Ge, and Jianing Quan. <a href="https://ieeexplore.ieee.org/document/10172044">Unambiguous pyramid cost volumes fusion for stereo matching</a>. IEEE TCSVT, 2024.
rtx3090
01/27/22
150
UPFNet
H
2
10.3
101
9.74
113
4.67
92
6.28
102
5.54
92
20.1
125
8.78
118
9.42
92
7.51
86
3.78
111
22.9
126
16.0
99
16.7
127
15.6
104
8.70
96
16.0
111
Weiyu Guo, Zhaoshuo Li, Yongkui Yang, Zheng Wang, Russell H. Taylor, Mathias Unberath, Alan Yuille, and Yingwei Li. <a href="https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136920263.pdf">Context-enhanced stereo transformer</a>. ECCV 2022. <a href="https://github.com/guoweiyu/Context-Enhanced-Stereo-Transformer">Code</a>.
Add context enhanced path to the transformer.
RTX 3090
02/27/22
151
CSTR
H
2
8.72
87
6.28
76
6.00
127
4.13
57
5.00
84
8.03
53
7.81
106
5.33
66
5.80
76
3.25
99
20.3
117
14.4
85
11.2
85
12.9
68
7.29
86
31.5
155
Qiong Chang et al. <a href="https://www.sciencedirect.com/science/article/abs/pii/S138376212100254X">Efficient stereo matching on embedded GPUs with zero-means cross correlation</a>. Journal of Systems Architecture, 2022.
Introduce an effective GPU-based calculation method for ZNCC.
NCC: 5x5
Jetson Tx2
04/11/22
152
Z2ZNCC
Q
1
34.4
210
40.9
206
20.4
216
29.9
224
32.0
212
42.5
183
27.0
209
41.4
194
38.0
203
26.0
226
48.6
191
43.7
205
36.5
219
33.3
215
37.5
198
41.3
178
Chao He, Ming Li, Congxuan Zhang, Peilong Zhong, Zhongkai Zhou, and Junjie Wu. <a href="https://doi.org/10.1016/j.ijleo.2023.171145">Local stereo matching with side window</a>. Optik, 2023.
Xiaowei Yang, Zhiguo Feng, Yong Zhao, Guiying Zhang, and Lin He. <a href="https://www.sciencedirect.com/science/article/pii/S0262885621002419">Edge supervision and multi-scale cost volume for stereo matching</a>. Image and Vision Computing, Volume 117, January 2022, 104336.
we propose a novel and effective network architecture
RDNet that utilizes edge detection and multi-scale cost volume for robust stereo matching.
β1 = 0.9, β2 = 0.999
PowerEdge T640 GPUs
08/05/22
155
RDNet
H
2
11.3
114
11.2
116
5.24
112
5.45
79
6.51
102
16.7
113
8.89
119
16.7
127
14.9
133
4.75
129
16.5
102
19.3
120
15.4
116
12.5
64
12.1
122
14.2
105
Zhelun Shen, Xibin Song, Yuchao Dai, Dingfu Zhou, Zhibo Rao, and Liangjun Zhang. <a href="https://ieeexplore.ieee.org/document/10223424">Digging into uncertainty-based pseudo-label for robust stereo matching</a>. TPAMI, 2023.
Tesla V100
08/08/22
156
UCFNet_RVC
H
2
10.7
107
12.2
124
6.48
133
5.83
87
5.90
95
16.9
115
6.61
95
15.8
122
14.6
132
2.73
80
11.4
67
18.8
116
11.0
82
18.9
136
10.7
112
11.4
86
Pengxiang Li, Chengtang Yao, Yunde Jia, and Yuwei Wu. <a href="https://ieeexplore.ieee.org/document/10663688">Inter-scale similarity guided cost aggregation for stereo matching</a>. IEEE Transactions on Circuits and Systems for Video Technology, 2025.
python; 2 cores + RTX 3090 GPU
08/09/22
157
issga
H
2
18.9
154
12.0
123
11.6
184
11.1
157
18.3
164
14.3
98
14.6
170
28.6
164
26.2
169
5.90
149
13.5
82
41.4
200
21.9
153
22.2
157
19.4
157
30.7
151
A. F. Kadmin, R. A. Hamzah, M. N. Abd Manap, M. S. Hamid, and T. F. Tg. Wook.
<a href="https://ijeecs.iaescore.com/index.php/IJEECS/article/view/24618">Local stereo matching algorithm using modified dynamic cost computation</a>. IJEECS 22(3), 2022.
Stereo matching algorithm based on multi-cost computation with hybrid aggregation using random walk and image segmentation with filtering in refinement stage.
Windows 10; Xeon(R) 8core @2.6GHz; 64G RAM; Matlab; C/C++;
08/29/22
158
MCP-HA-VQ
Q
1
30.6
195
47.8
227
17.8
207
23.0
208
25.9
193
41.9
179
24.6
201
38.2
186
31.5
184
18.4
202
43.0
178
35.7
178
29.6
199
29.6
199
35.8
190
47.9
196
Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. <a href="https://haofeixu.github.io/unimatch/">Unifying flow, stereo and depth estimation</a>. TPAMI, 2023. <a href="https://github.com/autonomousvision/unimatch">Code</a>.
A unified global matching formulation and framework for optical flow and stereo depth estimation
Please refer to the paper
V100 GPU
09/01/22
159
GMStereo
F
3
7.14
63
6.30
77
6.20
130
6.22
100
6.62
104
9.79
67
2.76
43
5.69
71
5.17
67
4.04
118
14.0
88
11.2
60
6.81
51
11.8
59
6.90
78
12.8
96
Xiaowei Yang, Yong Zhao, Zhiguo Feng, Haiwei Sang, Zhenbo Zhang, Guiying Zhang, and Lin He. <a href="https://doi.org/10.1049/ipr2.12756">A light-weight stereo matching network based on multi-scale features fusion and robust disparity refinement</a>. IET Image Processing, 2023.
In recent years, convolutional-neural-network based stereo matching methods have achieved significant
gains compared to conventional methods in terms of both speed and accuracy. Current state-of-the-art disparity
estimation algorithms require many parameters and large amounts of computational resources and are not suited to
applications on edge devices. In this paper, we propose an end-to-end light-weight network (LWNet) for fast stereo
matching, which consists of an efficient backbone with multi-scale feature fusion for feature extraction, a 3D U-Net
aggregation architecture for disparity computation and a color guidance in 2D CNN for disparity refinement.
(β1= 0.9, β2 = 0.999)
GeForce RTX 3090
09/20/22
160
LWNet
H
2
40.9
228
38.1
198
18.4
209
30.5
227
33.3
220
43.2
187
30.9
222
49.2
214
50.6
227
22.8
218
58.1
217
54.2
230
41.8
231
37.5
226
58.8
232
81.5
240
Dominik Hirner and Friedrich Fraundorfer. <a href="https://github.com/thedodo/FCDSN-DC">FCDSN-DC: An accurate and lightweight convolutional neural network for stereo estimation with depth completion</a>. ICPR 2022.
We propose an accurate and lightweight convolutional neural network for stereo estimation with depth completion. The whole method consists of three parts. The first part consists of fully-convolutional densely connected layers that computes expressive features of rectified image pairs.
learning rate: 0.00006 for feature extraction and similarity and learning rate: 0.000006 for depth completion
python3; RTX 3090
09/27/22
161
FCDSN-DC
H
2
18.8
152
23.0
157
7.01
142
10.2
146
20.1
174
37.7
168
17.3
177
27.8
159
20.8
158
7.81
165
23.9
130
24.5
140
22.4
154
20.7
147
16.0
142
19.9
128
Junpeng Jing and Jiankun Li. CREStereo++_RVC: Cascaded recurrent network for robust stereo matching. <a href="http://www.robustvision.net/leaderboard.php?benchmark=stereo">RVC 2022</a> entry.
A single model for RVC2022 Challenge on MiddleBury, ETH3D, and KITTI2015.
Philippe Weinzaepfel, Vincent Leroy, Thomas Lucas, Romain Bregier, Yohann Cabon, Vaibhav Arora, Leonid Antsfeld, Boris Chidlovskii, Gabriela Csurka, and Jerome Revaud. <a href="https://openreview.net/pdf?id=wZEfHUM5ri">CroCo: Self-supervised pretraining for 3D vision tasks by cross-view completion</a>. NeurIPS 2022; <a href="http://www.robustvision.net/leaderboard.php?benchmark=stereo">RVC 2022</a> entry. <a href="https://github.com/naver/croco">Code</a>.
pretraining self-supervised model
masking rate 0.9
python Nvidia
10/03/22
165
CroCo_RVC
F
3
15.1
134
7.43
94
5.85
125
6.71
110
11.7
134
15.4
100
3.94
58
36.2
182
35.8
198
3.41
105
18.1
109
29.3
158
10.9
81
18.0
126
10.6
111
21.0
130
Hualie Jiang. An improved RaftStereo trained with multiple mixed datasets for Robust Vision Challenge. <a href="http://www.robustvision.net/leaderboard.php?benchmark=stereo">RVC 2022</a> entry.
GTX 2080ti
10/03/22
167
iRaftStereo_RVC
H
2
8.07
79
9.13
110
8.25
154
5.55
84
4.68
80
6.92
41
6.41
93
6.29
75
6.19
80
3.96
115
17.9
107
13.0
71
9.58
74
11.4
56
9.24
101
11.8
90
Han Li, Guohua Gou, Hao Zhang, Xuanhao Wang, Yang Ye, and Haigang Sui. <a href="https://doi.org/10.1117/1.JEI.33.6.063051">Adaptive slice stereo matching network</a>. Journal of Electronic Imaging, 2024.
pytorch
10/06/22
168
GwcSlice
H
2
12.7
121
13.4
127
4.76
95
5.33
78
7.69
114
17.0
116
11.1
134
13.7
112
9.88
107
4.22
125
20.1
115
20.1
124
17.4
130
16.9
114
14.0
131
36.5
164
Xiaowei Yang, Lin He, Yong Zhao, Haiwei Sang, Zu Liu Yang, and Xian Jing Cheng. <a href="https://ieeexplore.ieee.org/document/9120049">Multi-attention network for stereo matching</a>. IEEE Access, vol. 8, 2020.
attention
0.9
GTX 1080
10/08/22
169
MANet
H
2
17.5
144
23.0
158
5.25
113
9.82
142
11.4
132
31.1
154
12.9
148
22.5
139
16.1
140
7.95
167
24.9
134
28.1
154
24.4
166
20.5
145
18.9
154
31.1
153
Ziyu Zhong, Xiuze Yang, Xiubian Pan, Wei Guan, Ke Liang, Jing Li, Xiaolan Liao, and Shuo Wang. <a href="https://doi.org/10.1038/s41598-024-57321-6">An efficient and accurate multi-level cascaded recurrent network for stereo matching</a>. Scientific Reports, 2024.
Python;32cores+NVIDIA GeForce RTX3090 GPU
10/16/22
170
LMCR-Stereo
F
3
6.27
53
6.20
72
4.59
90
3.92
49
2.66
43
4.52
15
4.88
66
3.65
46
3.41
44
2.08
53
16.8
104
11.2
62
8.58
64
13.2
72
6.89
77
10.5
73
Yun Wang, Longguang Wang, Kunhong Li, Yongjian Zhang, Dapeng Oliver Wu, and Yulan Guo. <a href="https://ieeexplore.ieee.org/document/10750250">Cost volume aggregation in stereo matching revisited: A disparity classification perspective</a>. IEEE TIP, 2024.
Cost aggregation plays a critical role in existing stereo
matching methods. Generally, aggregating matching costs
in homogeneous regions with similar disparities is benefi-
cial to matching accuracy. However, previous approaches
commonly use 3D convolutions for cost aggregation with-
out considering the homogeneity of different regions. In
this paper, we revisit cost aggregation in stereo match-
ing from a perspective of disparity classification and pro-
pose a generic yet efficient Disparity Context Aggregation
(DCA) module to improve the performance of CNN-based
methods.
Parameters:4.96 M;
Only using half-resolution Middlebury training images for validation.
Oh-Hun Kwon and Eduard Zell. <a href="https://ieeexplore.ieee.org/document/10222247">Image-coupled volume propagation for stereo matching</a>. ICIP 2023.
PyTorch; RTX 3090
11/11/22
173
ICVP
H
2
7.97
76
11.3
117
3.97
63
5.02
71
8.79
119
17.1
117
5.62
77
7.51
83
6.97
83
3.09
91
13.7
85
12.7
67
9.23
71
10.9
48
6.28
68
9.73
65
Peng Xu, Zhiyu Xiang, Chengyu Qiao, Jingyun Fu, and Tianyu Pu. <a href="https://openaccess.thecvf.com/content/CVPR2024/papers/Xu_Adaptive_Multi-Modal_Cross-Entropy_Loss_for_Stereo_Matching_CVPR_2024_paper.pdf">Adaptive multi-modal cross-entropy loss for stereo matching</a>. CVPR 2024.
Python + Nvidia 2080Ti
12/02/22
174
GANet+ADL
H
2
17.7
145
21.3
153
3.97
63
6.61
108
11.7
135
25.9
143
6.07
88
40.5
191
35.0
196
3.68
110
24.1
131
19.3
119
13.9
106
14.4
90
23.0
167
33.5
159
Yuanwei Bi, Chuanbiao Li, Xiangrong Tong, Guohui Wang, and Haiwei Sun. <a href="https://www.nature.com/articles/s41598-023-39964-z">An application of stereo matching algorithm based on transfer learning on robots in multiple scenes</a>. Scientific Reports, 2023.
Zihua Liu, Yizhou Li, and Masatoshi Okutomi. <a href="https://openaccess.thecvf.com/content/WACV2024/papers/Liu_Global_Occlusion-Aware_Transformer_for_Robust_Stereo_Matching_WACV_2024_paper.pdf">Global occlusion-aware transformer for robust stereo matching</a>. WACV 2024. <a href="https://github.com/Magicboomliu/GOAT">Code</a>.
Occlusion-Aware Global Aggregation for robust stereo matching using vision transformer.
Iter=18, resolution=1/8
NVIDIA RTX TITAN X
03/07/23
177
GOAT18
H
2
8.73
88
7.26
91
7.32
146
6.80
111
3.47
66
10.3
72
10.4
130
5.14
64
5.16
66
4.95
133
15.9
100
13.9
79
11.2
86
9.62
42
13.1
127
16.4
112
Kai Zeng, Hui Zhang, Wei Wang, Yaonan Wang, and Jianxu Mao. <a href="https://ieeexplore.ieee.org/document/10239529">Deep stereo network with MRF-based cost aggregation</a>. IEEE TCSVT, 2024.
Tesla V100
04/18/23
178
DMCANet
H
2
7.79
73
7.91
99
4.12
69
3.79
46
4.26
74
11.2
86
10.1
128
6.76
78
4.85
61
3.32
103
12.9
77
13.3
75
10.5
78
12.9
69
9.11
98
10.1
69
Yun Wang, Kunhong Li, Longguang Wang, Junjie Hu, Dapeng Oliver Wu, and Yulan Guo. <a href="https://ieeexplore.ieee.org/document/10890914">ADStereo: Efficient stereo matching with adaptive downsampling and disparity alignment</a>. IEEE TIP, 2025. <a href="https://github.com/cocowy1/ADStereo">Code</a>.
NVIDIA 3090TI
04/28/23
179
ADStereo
H
2
18.0
147
16.4
139
14.9
198
12.6
165
21.3
177
20.6
126
16.6
174
15.8
121
16.0
139
7.43
160
19.1
112
52.0
221
24.8
172
18.1
127
17.7
147
11.2
83
Peng Yao, Haiwei Sang, and Xu Cheng. <a href="https://link.springer.com/article/10.1007/s00371-024-03406-2">Structured support vector machine with coarse-to-fine PatchMatch filtering for stereo matching</a>. The Visual Computer, 2024.
Stereo Matching Using Structured Supported Vector Machine and Coarse to Fine Features
The proposed IGEV-Stereo builds a combined
geometry encoding volume that encodes geometry and context information as well as local matching details, and iteratively indexes it to update the disparity map.
Details in code.
Python RTX 3090
06/22/23
181
IGEV-Stereo
F
3
4.83
41
3.17
15
2.46
27
1.97
21
2.19
35
5.63
27
1.22
18
16.2
123
9.20
99
1.17
27
3.77
19
4.93
15
5.35
40
6.99
29
2.31
20
5.00
32
Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Bregier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jerome Revaud. <a href="https://github.com/naver/croco">CroCo v2: Improved cross-view completion pre-training for stereo matching and optical Flow</a>. ICCV 2023. <a href="https://github.com/naver/croco">Code</a>.
improved cross-view completion pre-training before finetuning for stereo matching
tiling with crops 704x352
NVIDIA A100
08/10/23
182
CroCo-Stereo
F
3
7.29
68
4.90
48
3.62
53
1.74
15
7.01
108
9.90
68
1.78
32
16.4
125
17.4
146
1.45
37
6.20
34
15.3
92
4.95
34
8.62
37
5.00
49
10.0
68
Zhaohuai Liang and Changhe Li. <a href="https://ojs.aaai.org/index.php/AAAI/article/view/28119">Any-Stereo: Arbitrary scale disparity estimation for iterative stereo matching</a>. AAAI 2024. <a href="https://github.com/Zhaohuai-L/Any-Stereo">Code</a>.
8 cores + NVIDIA A40
08/13/23
183
Any-RAFT
F
3
5.22
45
5.19
54
4.20
72
4.00
55
2.23
36
5.88
31
4.06
59
3.05
36
2.91
39
2.04
51
9.76
52
10.7
53
8.77
67
9.90
44
4.94
47
6.72
42
Madiha Zahari, Rostam Affendi Hamzah, Nurulfajar Abd Manap, and Adi Irwan Herman.
<a href="https://doi.org/10.11591/eei.v12i1.4122">Stereo matching algorithm for autonomous vehicle navigation using integrated matching cost and non-local aggregation</a>.
Bulletin of Electrical Engineering and Informatics, 2023.
Visual Studio c++, Intel(R) Core(TM) i7-8565U CPU @ 1.80GHz 1.99 GHz
09/27/23
184
FM-DT
Q
1
40.6
226
45.4
221
20.8
220
32.3
232
40.5
231
51.3
206
30.6
219
52.1
219
48.0
220
23.2
220
53.4
208
47.0
213
39.7
225
29.0
195
55.3
229
76.2
235
Zhien Dai, Zhaohui Tang, Hu Zhang, Can Tian, Mingjun Pan, and Yongfang Xie. <a href="https://dl.acm.org/doi/10.1145/3664647.3681372">Eglcr: Edge structure guidance and scale adaptive attention for iterative stereo matching</a>. ACMMM 2024. <a href="https://github.com/kangarooCV/Eglcr">Code</a>.
RTX3090s
10/09/23
185
EGLCR-Stereo
F
3
4.03
32
4.69
44
2.46
27
3.70
45
2.99
56
10.7
81
2.48
42
1.95
21
1.63
24
0.94
18
5.76
31
8.17
38
3.84
22
10.3
47
2.99
30
4.87
31
Kunhong Li, Longguang Wang, Ye Zhang, Kaiwen Xue, Shunbo Zhou, and Yulan Guo. <a href="https://openaccess.thecvf.com/content/CVPR2024/papers/Li_LoS_Local_Structure-Guided_Stereo_Matching_CVPR_2024_paper.pdf">LoS: Local structure-guided stereo matching</a>. CVPR 2024.
RTX 4090
10/30/23
186
LoS
F
3
4.20
33
5.85
66
4.92
101
4.64
63
2.77
47
3.92
10
1.32
22
2.36
30
2.17
32
1.81
46
8.18
44
6.58
26
4.55
29
8.57
36
4.57
41
5.06
34
Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. <a href="https://openaccess.thecvf.com/content/CVPR2024/papers/Wang_Selective-Stereo_Adaptive_Frequency_Information_Selection_for_Stereo_Matching_CVPR_2024_paper.pdf">Selective-Stereo: Adaptive frequency information selection for stereo matching</a>.
CVPR 2024. <a href="https://github.com/Windsrain/Selective-Stereo">Code</a>.
Selective-Stereo (IGEV-version)
Python; RTX 3090
11/13/23
187
Selective-IGEV
F
3
2.51
11
2.54
8
1.86
13
2.51
31
1.12
16
7.22
46
1.23
20
1.36
8
1.17
15
1.16
25
4.48
23
4.83
14
2.99
17
3.79
9
2.26
18
4.72
30
GongBen Han, Shuo Shan, YinHui Xu, KanJian Zhang, and HaiKun Wei. <a href="https://iopscience.iop.org/article/10.1088/1361-6501/adc623">4D-IteraStereo: stereo matching via 4D cost volume aggregation and iterative optimization</a>. Measurement Science and Technology, 2025.
NVIDIA RTX 3090 (PyTorch)
12/07/23
188
4D-IteraStereo
H
2
10.9
110
5.87
67
5.59
118
6.15
98
6.07
96
9.15
62
7.46
102
27.0
153
32.4
188
2.46
69
12.2
72
11.2
60
7.18
54
12.2
62
7.37
89
6.77
43
Muhammad Nazmi Zainal Azali, Rostam Affendi Hamzah, and Zarina Mohd Noh. <a href="https://ieeexplore.ieee.org/abstract/document/10106746">Disparity map algorithm using census transform and hierarchical segment-tree from stereo image</a>. ETIC 2022.
This article presents a disparity map algorithm to improve the depth map estimation based on Census Transform and hierarchical segment-tree on each block.The stereo matching algorithm presented in this study comprises of four steps: Cost Computation, Cost
Aggregation, Optimization, and Post-Processing, all of which will refine the final disparity map.
CostAlpha = 0.3;
CEN-WND = 9x11;
k = 1600;
LR checking = Yes
PY_LVL = 3.
C++, a personal PC with a CPU i7 8700@3.2 GHz, an RTX 2070 SUPER, and 16GB RAM.
Unsupervised Stereo Matching methods have made significant strides recently. However, these approaches have predominantly relied on the assumption of photometric consistency, leading to potential limitations: sensitivity to illuminance changes and difficulty in dealing with problematic areas like occluded or textureless regions.
To mitigate these limitations, this paper introduces a novel self-supervised dual-level framework named \textbf{\textit{Dual-Net}}.
This framework mainly consists of two key components: self-supervised teacher training and student training based on knowledge distillation.
Specifically, the teacher model is first trained in a self-supervised fashion with a focus on feature space and data augmentation consistency.
On the one hand, pixels from feature space are robust to noise and luminance changes, which are discriminative even in textureless regions.
On the other hand, a data augmentation consistency loss is presented to guide the model toward enhanced contextual awareness, thus leading to a completed depth estimation in problematic regions.
Then, the knowledge learned by the teacher model is distilled and transferred probabilistically to the student model. By leveraging this distilled knowledge, the student model is guided by validated insights, enabling it to outperform its teacher model by a large margin.
700 M
nvidia a100 GPU
01/08/24
190
DualNet
H
2
16.4
140
19.7
147
7.99
152
10.1
144
18.3
165
24.1
137
10.0
126
23.9
142
20.4
154
7.79
164
23.0
127
23.1
135
16.3
124
18.8
134
17.0
144
18.5
125
Aixin Chong, Hui Yin, Qianqian Du, Yanting Liu, and Ming Han. <a href="https://dl.acm.org/doi/10.1016/j.patcog.2025.111397">Gradual interaction network for stereo matching</a>. Pattern Recognition, 2025.
RTX 3090
01/08/24
191
GINet
H
2
15.6
137
16.1
137
7.15
144
7.37
119
9.39
122
25.1
140
7.88
109
35.2
180
32.6
190
3.19
95
15.5
98
16.7
104
11.6
87
14.7
93
15.8
140
26.4
145
Ziyang Chen, Wenting Li, Yongjun Zhang, Bingshu Wang, Yabo Wu, Yong Zhao, and C. L. Philip Chen. <a href="https://arxiv.org/pdf/2501.01023">Hadamard attention recurrent transformer: A strong baseline for stereo matching transformer</a>. arXiv:2501.01023, 2025. <a href="https://github.com/ZYangChen/HART">Code</a>.
Junhong Min and Youngpil Jeon. <a href="https://ieeexplore.ieee.org/document/10647686">Confidence aware stereo matching for realistic cluttered scenario</a>. ICIP 2024.
Our approach estimates disparities using implicitly inferred confidence levels with Unet transformer.
3x unet transformer for feature extraction, runs with 1984*2816px resolution
Python, Nvidia 4090
06/27/24
199
CAS++
F
3
3.33
21
4.27
40
3.72
57
3.17
39
2.17
33
2.44
4
1.33
23
2.24
27
2.01
29
1.47
38
4.04
21
8.15
37
4.97
35
5.80
18
3.73
38
3.04
24
Yansong Zhu, Songwei Pei, BingFeng Liu, and Jun Gao. <a href="https://doi.org/10.1016/j.neucom.2024.128685">AP-Net: Attention-fused volume and progressive aggregation for accurate stereo matching</a>. Neurocomputing, 2025.
Python; 16 cores + 2 * A800
07/22/24
200
apnet
Q
1
30.9
198
18.3
144
9.59
171
17.1
195
24.8
189
49.1
201
19.5
185
32.3
171
29.2
178
22.2
217
60.7
226
33.2
169
27.0
188
28.0
189
64.4
238
63.7
220
Jingyi Zhou, Haoyu Zhang, Jiakang Yuan, Peng Ye, Tao Chen, Hao Jiang, Meiya Chen, and Yangyang Zhang. <a href="https://doi.org/10.1609/aaai.v39i10.33173">All-in-One: Transferring vision foundation models into stereo matching</a>. AAAI 2025.
Pytorch; A100
08/07/24
202
AIO-Stereo
F
3
2.36
8
2.38
5
1.71
10
3.22
40
0.85
8
5.83
28
1.24
21
1.42
13
1.32
18
1.03
22
4.49
24
4.81
13
2.43
11
3.61
7
2.12
13
3.63
26
Soomin Kim, Hyesong Choi, Jihye Ahn, and Dongbo Min. <a href="https://arxiv.org/abs/2409.02545">UniTT-Stereo: Unified training of transformer for enhanced stereo matching</a>. arXiv:2409.02545, 2024.
Pytorch; NVIDIA A100
08/13/24
203
UniTT-Stereo
F
3
6.34
54
3.96
32
2.69
33
1.82
17
7.92
116
11.7
91
1.81
33
14.2
114
13.8
126
1.22
30
5.07
27
16.6
103
4.09
24
5.89
19
2.91
27
8.44
50
Jihye Ahn, Hyesong Choi, Soomin Kim, and Dongbo Min. <a href="https://ieeexplore.ieee.org/document/10836678">MaDis-Stereo: Enhanced stereo matching via distilled masked image modeling</a>. IEEE Access, 2025. <a href="https://github.com/ja053199/madis/">Code</a>.
This paper focuses on effectively capturing local patterns from images during the fine-tuning of Transformer-based models with limited labeled training data in dense downstream tasks, particularly in the context of stereo matching. For that, we propose MaDis-stereo, a novel stereo depth estimation framework that enhances locality inductive biases during fine-tuning via Masked Image Modeling (MIM).
a100 / 1 GPU
08/15/24
204
MaDis-Stereo
F
3
9.49
95
3.73
28
3.14
44
1.76
16
9.05
120
10.5
77
1.74
31
27.8
158
27.9
174
1.50
39
7.47
41
19.8
121
4.80
33
11.8
58
3.40
34
10.2
72
Shimeng Fan, Wei Sun, Jin Zheng, Qiang Fu, Min Xue, and Wei Wu. <a href="https://doi.org/10.1016/j.image.2023.116945">Accurate edge-preserving stereo matching by enhancing anisotropy</a>. Signal Processing: Image Communication, 2023.
Intel Core i5-9300H CPU @2.40 GHz (C++, OpenCV)
07/27/24
201
esmea
H
2
30.1
192
29.4
175
9.48
167
17.0
194
31.7
211
49.7
202
15.2
171
52.6
221
45.9
215
11.9
191
46.5
186
52.1
222
27.2
190
23.7
170
25.2
170
54.5
206
Yun Wang, Longguang Wang, Chenghao Zhang, Yongjian Zhang, Zhanjie Zhang, Ao Ma, Chenyou Fan, Tin Lun Lam, and Junjie Hu. <a href="https://openaccess.thecvf.com/content/ICCV2025/papers/Wang_Learning_Robust_Stereo_Matching_in_the_Wild_with_Selective_Mixture-of-Experts_ICCV_2025_paper.pdf">Learning robust stereo matching in the wild with selective mixture-of-experts</a>. ICCV 2025. <a href="https://github.com/cocowy1/SMoE-Stereo">Code</a>.
We propose S-MoEStereo, which adapts pre-trained VFMs for stereo matching by integrating Low-Rank Adaptation (LoRA) with Mixture-of-Experts (MoE) modules.
This approach balances parameter efficiency and discriminative feature learning by dynamically selecting the optimal expert within each MoE module.
Additionally, we introduce CNN-based adapter layers to incorporate inductive bias, enhancing geometric feature extraction.
Furthermore, we propose a lightweight decision network to reduce computational costs by selectively activating MoE modules based on input complexity.
113M
A5000ada
10/26/24
205
SMoEStereo_RVC
H
2
5.83
51
6.58
78
5.15
108
3.96
52
3.82
71
5.84
29
5.41
75
4.43
55
4.21
57
2.31
59
10.2
56
8.41
41
6.20
45
11.1
51
6.59
75
8.19
48
Hualie Jiang, Zhiqiang Lou, Laiyan Ding, Rui Xu, Minglang Tan, Wenjie Jiang, and Rui Huang. <a href="https://openaccess.thecvf.com/content/CVPR2025/papers/Jiang_DEFOM-Stereo_Depth_Foundation_Model_Based_Stereo_Matching_CVPR_2025_paper.pdf">DEFOM-Stereo: Depth foundation model based stereo matching</a>. CVPR 2025. <a href="https://github.com/Insta360-Research-Team/DEFOM-Stereo">Code</a>.
Pytorch; RTX 4090
11/03/24
206
DEFOM-Stereo
F
3
2.39
10
2.82
11
2.21
17
1.53
9
1.01
11
5.24
24
0.88
10
1.40
12
1.14
14
0.85
14
2.64
12
9.10
43
2.18
10
5.50
16
2.49
22
1.67
8
Hualie Jiang, Zhiqiang Lou, Laiyan Ding, Rui Xu, Minglang Tan, Wenjie Jiang, and Rui Huang. <a href="https://arxiv.org/abs/2501.09466">DEFOM-Stereo: Depth foundation model based stereo matching</a> (RVC version). arXiv:2501.09466, 2025.
Robust Vision Challenge Version of DEFOM-Stereo
Pytorch; RTX 4090
11/28/24
207
DEFOM-Stereo_RVC
F
3
3.28
20
3.50
21
2.61
31
2.41
29
0.87
9
2.51
5
0.89
11
1.38
9
1.26
17
0.97
20
6.35
36
10.8
56
2.43
11
11.0
50
3.03
31
5.00
32
Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. <a href="https://nvlabs.github.io/FoundationStereo/">FoundationStereo: Zero-shot stereo matching</a>. CVPR 2025. <a href="https://github.com/NVlabs/FoundationStereo/">Code</a>.
A100
02/03/25
208
FoundationStereo
F
3
1.84
6
2.46
6
1.71
10
1.36
7
0.79
7
5.19
23
0.53
1
0.93
3
0.84
5
0.93
16
2.41
9
3.39
4
3.45
20
3.28
4
1.82
7
1.17
6
Jiahao Li, Xinhong Chen, Zhengmin Jiang, Qian Zhou, Yung-Hui Li, and Jianping Wang. <a href="https://arxiv.org/pdf/2509.15891">Global regulation and excited attention tuning for stereo matching</a>. ICCV 2025. <a href="https://github.com/JarvisLee0423/GREAT-Stereo">Code</a>.
Apply the global context information to handle the ill-posed regions in stereo matching.
14.44 M
PyTorch RTX 4090 GPUs
02/10/25
209
GREAT-IGEV
F
3
2.81
16
3.30
17
2.44
25
2.31
28
0.96
10
7.12
44
1.17
17
1.38
9
1.36
19
1.04
23
3.89
20
3.82
6
4.66
31
6.24
23
2.17
14
4.65
28
Anze Xu, Lin Yang, Jingzhong Li, Zhen Shi, Yue Jin, YuXuan Chen, Yanxin Ni, and Guifei Li. <a href="https://doi.org/10.1016/j.neunet.2025.107591">Real-time and accurate stereo matching via tri-fusion volume for stereo vision</a>. Neural Networks, 2025.
RTX 3090
02/16/25
210
TCM
H
2
12.8
123
13.5
129
5.64
120
7.01
116
12.0
138
30.2
152
10.0
127
9.66
96
7.53
87
5.58
145
27.0
138
21.0
130
16.0
120
18.8
133
12.5
125
18.1
122
Jiakang Yuan. LG-stereo: Local global cost volume for stereo matching. Submitted to TIP, 2025.
NVIDIA A100
03/04/25
211
LG-Stereo
F
3
1.76
5
2.57
9
1.86
13
2.02
22
0.65
5
3.23
9
0.68
6
0.98
5
0.81
4
0.55
4
2.15
8
4.26
9
2.03
8
3.85
11
1.27
3
2.42
15
Jie Tang, Gaofeng Peng, Jialu Liu, and Bo Yu. <a href="https://ieeexplore.ieee.org/document/11045114">G2L-Stereo: Global to local two-stage real-time stereo matching network</a>. Transactions on Computational Imaging, 2025.
python;Nvidia A800 GPU
03/19/25
212
G2L-Stereo
F
3
13.3
126
16.8
141
4.45
80
5.64
86
16.8
159
23.0
135
5.05
68
20.4
136
16.2
141
2.83
83
15.3
95
24.0
139
18.2
138
19.5
139
5.48
62
25.4
141
Luca Bartolomei, Fabio Tosi, Matteo Poggi, and Stefano Mattoccia. <a href="https://stereoanywhere.github.io/">Stereo Anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail</a>. CVPR 2025. <a href="https://github.com/bartn8/stereoanywhere/">Code</a>.
We introduce Stereo Anywhere, a novel stereo-matching framework that combines geometric constraints with robust priors from monocular depth Vision Foundation Models (VFMs). By elegantly coupling these complementary worlds through a dual-branch architecture, we seamlessly integrate stereo matching with learned contextual cues.
Datasets mixture (FSD, Tartanair, CREStereo)
Python; i7 8 cores + H100 GPU
04/24/25
214
StereoAnywhere
F
3
3.69
27
7.34
92
2.23
18
2.23
26
5.12
86
18.1
120
0.90
12
2.16
25
1.43
21
1.25
31
5.73
30
4.95
16
2.66
14
6.89
27
2.28
19
1.86
10
Jie Lin. M2-Stereo: Multi scale information fusion and multi branch iteration for stereo matching. Submitted to IEEE Robotics and Automation Letters, 2025.
M2-Stereo embedded three Multi scale Feature Fusion Attention Blocks in the feature extraction stage to fuse deep and shallow information, and used a Multi scale Cost Aggregation Module in the cost aggregation stage to achieve sharing of cost information at different scales. Finally, the Multi branch Iterative Strategy was used for efficient iteration.
Iters=12;lr=0.0002;batch size=8
python;8 cores + NVIDIA TITAN Xp GPU
04/24/25
213
M2-Stereo
F
3
3.90
31
8.13
100
2.35
23
1.61
13
3.58
69
15.9
107
1.43
27
1.83
19
1.13
13
0.79
13
3.09
16
7.36
31
6.23
46
7.99
35
2.63
24
4.03
27
Jie Tang, Gaofeng Peng, Jialu Liu, and Bo Yu. <a href="https://ieeexplore.ieee.org/document/11045114">G2L-Stereo: Global to local two-stage real-time stereo matching network</a>. Transactions on Computational Imaging, 2025.
python;Nvidia A800 GPU
05/07/25
215
G2L-ROB
F
3
11.6
116
15.0
135
5.19
110
4.91
69
13.1
147
19.4
123
6.22
89
17.5
130
14.0
127
2.55
71
13.9
87
20.8
129
14.0
107
18.2
128
6.96
81
14.6
109
Jie Lin, Juan Du, and Huake Wang. <a href="https://ieeexplore.ieee.org/document/11192657">DS-Stereo: Deep-shallow information interaction for stereo matching</a>. IEEE Robotics and Automation Letters, 2025.
DS-Stereo utilizes our proposed Adjacent Feature Hybrid Attention Block and Hierarchical Cost Aggregation Module to achieve deep to shallow information interaction in stereo matching. Simultaneously replacing the traditional ConvGRU iterative operator with an Inception like iterative operator to achieve high convergence updates.
lr=0.0002;iters=12;batch size=8
python; 8 cores + NVIDIA TITAN Xp GPU
05/07/25
216
DS-Stereo
F
3
3.13
17
4.92
49
2.14
16
1.40
8
1.23
18
10.1
69
1.16
16
1.55
16
1.25
16
0.93
16
3.41
18
6.18
19
5.10
38
8.66
38
2.21
16
2.43
16
Xiaoyang Zhao. Monocular prior-based water surface stereo matching algorithm. Submitted to Neurocomputing, 2025.
GTX 4090D
05/21/25
217
waterstereo
H
2
8.48
85
6.74
82
6.51
135
9.22
136
5.11
85
6.44
36
4.71
65
5.50
69
5.36
71
3.44
106
18.2
111
14.2
83
9.13
70
14.4
89
13.3
128
12.9
97
Zhien Dai, Zhaohui Tang, Hu Zhang, and Yongfang Xie. <a href="https://ieeexplore.ieee.org/document/11182608">Mgs-Stereo: Multi-scale geometric-structure-enhanced stereo matching for complex real-world scenes</a>. IEEE TIP, 2025.
Shamsul Fakhar Abd Gani, Muhammad Fahmi Miskon, Rostam Affendi Hamzah, Mohd Saad Hamid, Ahmad Fauzan Kadmin, and Adi Irwan Herman. <a href="https://beei.org/index.php/EEI/article/view/6480">Refining disparity maps using deep learning and edge-aware smoothing filter</a>. Bulletin of Electrical Engineering and Informatics, 2024.
This paper proposes a robust stereo matching algorithm that combines a CNN for initial cost computation, bilateral filtering with cross-based cost aggregation (CBCA) for refinement, and a winner-take-all (WTA) strategy for disparity selection, followed by an edge-aware smoothing filter (EASF) to reduce noise
Chuang-Wei Liu, Mingjian Sun, Cairong Zhao, Hanli Wang, Alexander Dvorkovich, and Rui Fan. <a href="https://ieeexplore.ieee.org/document/11121379">Integrating disparity confidence estimation into relative depth prior-guided unsupervised stereo matching</a>. TCSVT, 2025.
RTX4090
06/17/25
220
UnViTAStereo
H
2
24.3
170
28.5
174
9.72
173
12.5
164
12.7
144
29.6
149
12.1
141
45.0
203
39.5
206
9.19
175
42.7
175
27.0
148
21.1
148
22.3
158
36.7
193
38.8
173
Junhong Min, Youngpil Jeon, Jimin Kim, and Minyong Choi. <a href="https://junhong-3dv.github.io/s2m2-project/">S<sup>2</sup>M<sup>2</sup>: Scalable stereo matching model for reliable depth estimation</a>. ICCV 2025. <a href="https://github.com/junhong-3dv/s2m2">Code</a>.
transformer based global stereo matching using stacked multi-resolution vision transformer blocks
ch:384, ntr: 3, num_refine: 3
Intel i9-13900K, Nvidia 4090
06/27/25
222
S2M2
F
3
1.15
2
1.29
1
1.23
2
1.27
5
0.40
1
0.45
1
0.59
4
0.67
1
0.62
1
0.45
1
1.28
5
2.80
3
1.37
4
3.60
6
1.12
1
0.25
1
Yongjian Zhang, Longguang Wang, Kunhong Li, Ye Zhang, Yun Wang, Liang Lin, and Yulan Guo.
<a href="https://arxiv.org/abs/2507.08400">PanMatch: Unleashing the potential of large vision models for unified matching models</a>. arXiv:2507.08400, 2025. Submitted to TPAMI, 2025.
A100
06/17/25
221
PanMatch
F
3
7.18
65
5.21
55
5.34
116
3.34
42
5.43
90
4.52
15
2.47
41
13.2
110
13.3
124
1.51
41
8.34
45
16.6
101
8.07
59
10.3
46
4.96
48
9.23
57
Peng Yao, Haiwei Sang, and Linlin Ge. Stacked learning with exhaustive disparity characteristics for accurate stereo matching. Submitted to Signal Processing: Image Communication, 2025
Using stacked learning with exhaustive disparity characteristics for accurate stereo matching to archieve more accurate disparity estimation.
For the ERF, we set 10 Decision Trees, maximum depth is 25 and minimum number of samples in each node to split equal to 12.
Dodeca-Core Intel Core i7 CPUs，32GB RAM，C++，64-bits OS
07/11/25
223
SLEDC_v1
H
2
6.67
59
4.22
39
2.72
34
3.49
43
3.38
63
13.3
95
5.11
69
4.36
54
3.92
53
2.19
56
13.5
83
10.8
57
11.9
89
14.5
91
6.83
76
8.27
49
Tongfan Guan, Jiaxin Guo, Chen Wang, and Yun-Hui Liu. <a href="https://openaccess.thecvf.com/content/ICCV2025/papers/Guan_BridgeDepth_Bridging_Monocular_and_Stereo_Reasoning_with_Latent_Alignment_ICCV_2025_paper.pdf">BridgeDepth: Bridging monocular and stereo reasoning with latent alignment</a>. ICCV 2025. <a href="https://github.com/aeolusguan/BridgeDepth">Code</a>.
A unified framework that bridges monocular reasoning and stereo matching through iterative bidirectional alignment of their latent representations.
lr=5e-4,iters=100000
Python; PyTorch + GTX 4090D
07/24/25
225
BridgeDepth
H
2
3.78
29
13.0
125
2.45
26
1.58
11
1.54
22
9.56
65
2.27
37
3.67
48
1.65
25
1.29
33
7.63
42
6.44
25
2.72
15
7.79
33
2.70
25
2.68
20
Anonymous. GEAStereo: Geometry-aware stereo-matching via monocular disparity prior and gradients. AAAI 2026 submission 7299.
python with pytorch; Nvidia RTX 3090
07/19/25
224
GEAStereo
F
3
3.80
30
2.93
13
2.29
21
2.08
23
2.52
39
6.53
39
2.14
36
2.11
24
2.32
33
1.36
35
6.97
38
6.42
23
5.55
42
10.9
48
2.33
21
5.06
34
Junda Chen and Wenjing Liao. MonSter++: A unified geometric foundation model for multi-view depth estimation via harnessing monodepth priors. Submitted to TPAMI, 2025.
pytorch, RTX 4090
09/02/25
226
MonSter++
F
3
2.60
15
7.04
90
1.61
8
1.91
20
1.04
12
8.92
59
0.85
9
2.08
23
1.02
9
0.75
11
3.06
15
8.01
33
2.73
16
3.84
10
2.11
12
2.00
11
Anonymous. VMStereo: Accurate real-time stereo matching via Vision Mamba. CVPR 2026 submission 43434 (updated ID).
VMStereo is a real-time stereo matching model, comprising feature extraction, cross-view feature enhancement, two-stage disparity prediction, and multi-scale disparity correction.
Pretrain on Sceneflow and finetune on SceneFlow, Middlebury, ETH3D, KITTI12, KITTI15, Tartan Air, CREStereo, Falling Things, InStereo2k, CARLA HR-VS, Sintel, Booster, and Virtual KITTI 2;
Max disparity=384;LR=0.0006;
Python; Intel i9-14900KF CPU + Nvidia RTX 4090 GPU
10/23/25
227
VMStereo-Base
H
2
4.52
37
5.00
50
4.18
71
3.81
47
2.90
51
3.92
10
3.34
48
3.71
49
3.67
49
1.69
45
9.38
51
6.61
27
5.79
43
6.95
28
3.53
35
9.02
55
Anonymous. BLMT-Stereo: Breaking the local minima trap of iterative stereo matching. CVPR 2026 submission 31980 (updated ID).
A100s
11/11/25
228
BLMT-Stereo
F
3
1.57
4
2.24
4
1.58
6
3.95
51
0.59
4
4.58
18
1.22
18
1.03
6
1.01
8
0.54
3
1.18
2
1.81
2
1.16
2
1.18
2
1.91
9
1.15
5
Anonymous. DispViT: Direct stereo disparity regression with a single-stream vision transformer. ICLR 2026 submission 16420.
PyTorch@GTX 4090D
12/02/25
229
DispViT+
F
3
4.92
42
8.62
107
2.27
20
1.58
11
3.45
64
17.3
118
1.63
28
9.31
91
8.95
94
0.94
18
5.85
32
6.27
20
3.02
18
7.84
34
2.18
15
2.61
19
Chuang-Wei Liu and Rui Fan. Self-improvement of stereo matching via decisive disparity diffusion. Submitted to TCSVT, 2025.
The successor work of our pre-published method UnViTAStereo. We select confident disparities with a plug-and-play confidence estimation method to generate pseudo labels in this work.
BatchSize = 1, learning rate = 1e-5, all codes will be released upon publication
NVIDIA RTX 4090
12/22/25
230
SelfViTAS
H
2
17.4
143
23.2
160
8.83
160
12.2
163
11.4
131
24.7
139
9.67
123
16.5
126
10.7
114
9.96
179
35.6
161
27.3
150
19.3
141
20.9
149
21.6
166
42.9
181
Anonymous. WAFT-Stereo: Warping-alone field transforms for stereo matching. ECCV 2026 submission 1252.
Junhong Min, Jimin Kim, Cheol-Hui Min, Minwook Kim, Youngpil Jeon, and Minyong Choi. <a href="https://arxiv.org/abs/2511.16993">DepthFocus: Controllable depth estimation for see-through scenes</a>. CVPR 2026.
DepthFocus estimates depth in see-through scenes based on user intent, selecting the desired layer from multiple candidates. The submitted results assume focus on the nearest surface.
CH:384, NTR:2, Refine_iter:4
Nvidia GPU 4090
03/04/26
234
DepthFocus
F
3
1.00
1
1.80
2
1.45
4
0.93
2
0.45
2
0.62
2
0.54
2
0.83
2
0.76
2
0.50
2
1.19
3
3.61
5
1.33
3
0.67
1
1.26
2
0.25
1
Anonymous. Laplacian-guided adaptive cost aggregation for stereo matching. ACM MM 2026 submission 4249.
RTX 3090
03/23/26
235
LACA
H
2
4.71
39
6.81
84
2.54
30
1.54
10
2.96
54
10.2
71
1.70
30
9.62
95
8.96
95
1.32
34
5.29
28
6.33
21
5.01
37
6.69
25
2.98
29
2.58
17
Anonymous. MatchAttention: Embedding explicit matching constraints into attention for efficient stereo matching. AAAI 2027 submission 31915.
MatchAttention embeds an explicit matching constraint into attention by treating the relative position between a query and its matched key as a learnable component of attention sampling. Centering a small contiguous sampling window on this learnable relative position enforces the matching constraint and supports long-range correspondence at strictly linear attention complexity. Two variants are provided: MatchAttentionXL for high accuracy and MatchAttentionRT for real-time edge inference at 9.3 ms on RTX 4060 Ti and 79.1 ms on Jetson Orin NX.
M. Y. A. Mohammed Al-Asbahi, Ahmad Fauzan Kadmin, Rostam Affendi Hamzah, Khairul Azha A. Aziz, Nasharuddin Zainal, and Nabil Jazli Jazlan. <a href="https://jtec.utem.edu.my/jtec/article/view/6433">Enhanced local disparity map algorithm segment-side window-based cost aggregation and refinement</a>. Journal of Telecommunication, Electronic and Computer Engineering, 2025.
star census transform, hamming distance, Census Gradient, gradient magnitude
MSI-GPU-i7 core
04/16/26
237
CTHDGM
Q
1
51.2
239
46.6
222
36.1
244
44.9
241
49.1
239
62.1
222
39.1
237
57.2
233
55.0
235
43.3
242
80.6
239
55.9
232
47.0
239
43.5
235
66.8
240
72.2
233
Anonymous. NBS: No bias stereo. NeurIPS 2026 submission 21560.
Anonymous. GCAP-Stereo: Grouped correlation aggregation with propagation for zero-shot stereo matching. ECCV 2026 submission 11793.
8 RTX4090
05/11/26
240
GCAP-stereo
H
2
4.31
35
5.27
56
4.75
94
3.08
37
1.79
25
4.21
14
3.06
45
2.60
31
2.55
34
1.60
43
9.32
50
5.49
18
5.27
39
5.24
14
9.12
99
5.38
36
Anonymous. Bringing geometry foundation model priors into iterative stereo matching. AAAI 2027 submission 6797.
python; A40
07/11/26
241
GeoAtt
F
3
2.57
13
6.24
74
1.33
3
1.64
14
2.71
44
17.3
119
0.62
5
2.93
35
1.09
12
0.61
5
1.32
6
3.96
7
1.53
7
3.43
5
1.59
5
0.38
3
Ahmad Fauzan Kadmin, Rostam Affendi Hamzah, Nadzrie Mohamood, Ken Prameswari Caesarella Aryaputri, Karmarul Hawari Ghazali, and Nabil Jazli. <a href="https://ieeexplore.ieee.org/document/11375024">Segment-side window-based aggregation for local disparity with edge and texture preservation</a>. TENCON 2025.
The proposed stereo matching method employs a multi-cost matching strategy by combining Truncated Absolute Difference (TAD), Gradient Magnitude (GM), and Census Transform (CT) to improve matching robustness. Cost aggregation is performed using a Segment-Side Window-Based (SWB) approach that integrates SLIC superpixel segmentation with Side Window Filtering (SWF) to preserve object boundaries and texture details. Disparity selection is carried out using the Winner-Takes-All (WTA) strategy, followed by left-right consistency checking and median filtering for disparity refinement. The method is designed to reduce edge fattening and texture inconsistency while improving disparity estimation accuracy on the Middlebury benchmark.
Matlab; Intel Core i7-8700K @ 3.2 GHz
07/16/26
242
Q-Census
Q
1
43.7
235
51.3
230
27.6
237
36.4
238
48.5
238
60.2
220
38.3
236
47.1
209
42.4
211
28.4
234
63.1
230
61.1
238
42.3
233
46.5
239
43.0
210
56.7
210
Anonymous. WiStereo: Weaving context through propagation for stereo matching. AAAI 2027 submission 11717.
L20 GPUs
07/17/26
243
WiStereo
H
2
3.40
23
9.74
113
2.29
21
1.33
6
1.69
24
9.72
66
1.69
29
2.84
33
1.43
21
1.19
28
5.93
33
6.42
23
4.97
35
5.39
15
2.76
26
2.04
13
Anonymous. GOStereo: Bridging geometric prior and occlusion awareness for stereo matching. AAAI 2027 submission 9076.
Shamsul Fakhar Abd Gani, Muhammad Fahmi Miskon, Rostam Affendi Hamzah, Mohd Saad
Hamid, Ahmad Fauzan Kadmin, and Adi Irwan Herman. <a href="https://www.scopus.com/pages/publications/85196516068">Stereo matching algorithm using deep learning and edge-preserving filter for machine vision</a>. BEEI 2024.
