• Acta Optica Sinica
  • Vol. 40, Issue 14, 1415001 (2020)
Mingyang Cheng1、2, Shaoyan Gai1、2, and Feipeng Da1、2、3、*
Author Affiliations
  • 1School of Automation, Southeast University, Nanjing, Jiangsu 210096, China
  • 2Key Laboratory of Measurement and Control of Complex Systems of Engineering, Ministry of Education, Southeast University, Nanjing, Jiangsu 210096, China
  • 3Shenzhen Research Institute, Southeast University, Shenzhen, Guangdong 518063, China
  • show less
    DOI: 10.3788/AOS202040.1415001 Cite this Article Set citation alerts
    Mingyang Cheng, Shaoyan Gai, Feipeng Da. A Stereo-Matching Neural Network Based on Attention Mechanism[J]. Acta Optica Sinica, 2020, 40(14): 1415001 Copy Citation Text show less
    Algorithm block diagram
    Fig. 1. Algorithm block diagram
    Structure of our proposed network
    Fig. 2. Structure of our proposed network
    Structure of spatial attention mechanism
    Fig. 3. Structure of spatial attention mechanism
    Structure of channel attention mechanism
    Fig. 4. Structure of channel attention mechanism
    Results on KITTI2015, from the top to the bottom: the left input, predicted disparity map, actual disparity map, error map
    Fig. 5. Results on KITTI2015, from the top to the bottom: the left input, predicted disparity map, actual disparity map, error map
    Results on KITTI2012, from the top to the bottom: the left input, predicted disparity map, actual disparity map, error map
    Fig. 6. Results on KITTI2012, from the top to the bottom: the left input, predicted disparity map, actual disparity map, error map
    Results on Sceneflow, from the top to bottom: the left input, actual disparity map, predicted disparity map
    Fig. 7. Results on Sceneflow, from the top to bottom: the left input, actual disparity map, predicted disparity map
    Comparison with other algorithms, from the top to the bottom: the PSM-Net results, the GWC-Net results, our results, our improvement results for the parts framed
    Fig. 8. Comparison with other algorithms, from the top to the bottom: the PSM-Net results, the GWC-Net results, our results, our improvement results for the parts framed
    MethodSceneflowKITTI2012KITTI2015
    EPE2 pixel3 pixel4 pixelALL DOC
    Out-NocOut-AllOut-NocOut-AllOut-NocOut-AllD1-allD1-all
    MC-CNN (Žbontar et al., 2016) [2]3.793.905.452.433.631.902.853.883.33
    GC-Net (Cao et al., 2019)[3]2.512.713.461.772.301.361.772.672.45
    iResNet-i2 (Liang et al., 2018)[20]1.402.693.341.712.161.301.632.442.19
    PSM-Net (Chang et al., 2018) [5]1.092.443.011.491.891.121.422.322.14
    SegStereo (Yang et al., 2018)[14]1.452.663.191.682.031.251.522.252.08
    GA-Net (Zhang et al., 2019)[7]0.842.182.791.361.801.031.371.931.73
    Ours0.952.332.981.421.760.921.212.222.07
    Table 1. Comparison with other algorithms
    Mingyang Cheng, Shaoyan Gai, Feipeng Da. A Stereo-Matching Neural Network Based on Attention Mechanism[J]. Acta Optica Sinica, 2020, 40(14): 1415001
    Download Citation