Author Affiliations
School of Physics & Optoelectronic Engineering, Guangdong University of Technology, Guangzhou 510006, Guangdong , Chinashow less
Fig. 1. Overall structure of multi-scale attention fusion network
Fig. 2. Group-related attention fusion module
Fig. 3. Multi-scale convolution global attention module
Fig. 4. 3D channel attention aggregation module
Fig. 5. Parallax maps obtained by different algorithms on SceneFlow dataset
Fig. 6. Visualization results of ablation experiment on KITTI2015 test set
Fig. 7. Qualitative evaluation results of different networks on KITTI2015 dataset
Fig. 8. Qualitative evaluation results of different networks on KITTI2012 dataset
Fig. 9. Qualitative evaluation results of different networks on Middlebury-v3 dataset
Module | >1 pixel | >2 pixel | >3 pixel | D1-all | EPE /% |
---|
GA | MA | CAA |
---|
| √ | | 0.0809 | 0.0438 | 0.0319 | 0.0260 | 0.757 | √ | √ | | 0.07780 | 0.0429 | 0.0316 | 0.0258 | 0.746 | √ | √ | √ | 0.0702 | 0.0384 | 0.0281 | 0.0226 | 0.662 |
|
Table 1. Ablation study results on SceneFlow dataset
Parameter | MCCNN | GCNet | iResNeti2 | CRL | PSMNet | EdgeStereo | SegStereo | MGNet |
---|
EPE /% | 3.79 | 1.84 | 1.40 | 1.32 | 1.09 | 1.11 | 1.45 | 0.662 |
|
Table 2. Comparison of EPE between MGNet and other methods
GA | MA | Gwc | CAA | >3 pixel /% |
---|
√ | | | | 2.20 | √ | √ | | | 2.18 | √ | √ | √ | | 2.06 | √ | √ | √ | √ | 2.01 |
|
Table 3. Benchmark results of designed module on KITTI2015 dataset
Network | ALL | Noc |
---|
D1-bg | D1-fg | D1-all | D1-bg | D1-fg | D1-all |
---|
DispNetC | 4.32 | 4.41 | 4.34 | 4.11 | 3.72 | 4.05 | CRL | 2.48 | 3.59 | 2.67 | 2.32 | 3.12 | 2.45 | PDSNet | 2.29 | 4.05 | 2.58 | 2.09 | 3.68 | 2.36 | GCNet | 2.21 | 6.16 | 2.87 | 2.02 | 5.58 | 2.61 | PSMNet | 1.86 | 4.62 | 2.32 | 1.71 | 4.31 | 2.14 | AANet | 1.99 | 5.39 | 2.55 | 1.80 | 4.93 | 2.32 | EdgeStereo | 2.27 | 4.18 | 2.59 | 2.12 | 3.85 | 2.40 | Big3D | 1.95 | 3.48 | 2.21 | 1.79 | 3.11 | 2.01 | MGNet | 1.65 | 3.84 | 2.01 | 1.51 | 3.49 | 1.84 |
|
Table 4. Comparison of different networks on KITTI2015 dataset
Network | >2 pixel | >3 pixel | >4 pixel | >5 pixel |
---|
Noc | ALL | Noc | ALL | Noc | ALL | Noc | ALL |
---|
DispNetC | 7.38 | 8.11 | 4.11 | 4.65 | 2.77 | 3.20 | 2.05 | 2.39 | PDSNet | 3.82 | 4.65 | 1.92 | 2.53 | 1.38 | 1.85 | 1.12 | 1.51 | GCNet | 2.71 | 3.46 | 1.77 | 2.30 | 1.36 | 1.77 | 1.12 | 1.46 | PSMNet | 2.44 | 3.01 | 1.49 | 1.89 | 1.12 | 1.42 | 0.90 | 1.15 | Edgestereo | 2.79 | 2.43 | 1.73 | 2.18 | 1.30 | 1.64 | 1.04 | 1.32 | SegStereo | 2.66 | 3.19 | 1.68 | 2.03 | 1.25 | 1.52 | 1.00 | 1.21 | SSPCVNET | 2.47 | 3.09 | 1.47 | 1.90 | 1.08 | 1.41 | 0.87 | 1.14 | EdgestereoV2 | 2.32 | 2.88 | 1.46 | 1.83 | 1.07 | 1.34 | 0.83 | 1.04 | AANet | 2.30 | 2.96 | 1.55 | 2.04 | 1.20 | 1.58 | 0.98 | 1.30 | MGNet | 2.12 | 2.71 | 1.34 | 1.76 | 1.01 | 1.34 | 0.82 | 1.08 |
|
Table 5. Comparison of different networks on KITTI2012 dataset