• Laser & Optoelectronics Progress
  • Vol. 59, Issue 10, 1010009 (2022)
Sen Xiang1、2、*, Nanting Huang1、2, Huiping Deng1、2, and Jin Wu1、2
Author Affiliations
  • 1School of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan 430081, Hubei , China
  • 2Engineering Research Center for Metallurgical Automation and Measurement Technology, Ministry of Education, Wuhan University of Science and Technology, Wuhan 430081, Hubei , China
  • show less
    DOI: 10.3788/LOP202259.1010009 Cite this Article Set citation alerts
    Sen Xiang, Nanting Huang, Huiping Deng, Jin Wu. Estimation of Light Field Depth Based on Multi-Level Network Optimization[J]. Laser & Optoelectronics Progress, 2022, 59(10): 1010009 Copy Citation Text show less
    Proposed network architecture
    Fig. 1. Proposed network architecture
    Feature extraction module
    Fig. 2. Feature extraction module
    Padding diagram
    Fig. 3. Padding diagram
    MSE of different levels of training model
    Fig. 4. MSE of different levels of training model
    Comparison of depth map results of validation datasets. (a) Comparison of qualitative results; (b) comparison of details
    Fig. 5. Comparison of depth map results of validation datasets. (a) Comparison of qualitative results; (b) comparison of details
    Comparison of depth map results of test datasets. (a) Experimental results; (b) detail drawing
    Fig. 6. Comparison of depth map results of test datasets. (a) Experimental results; (b) detail drawing
    MSE under different spatial resolution of input light field
    Fig. 7. MSE under different spatial resolution of input light field
    Comparison of experimental results of real light field
    Fig. 8. Comparison of experimental results of real light field
    SceneBP1 /%BP3 /%BP7 /%MSE×100
    WSProposedWSProposedWSProposedWSProposed
    Average32.0425.8414.1011.846.016.002.402.24
    Boxes37.5232.7823.6119.4712.8013.877.436.91
    Cotton44.1930.0613.168.022.790.990.470.54
    Dino25.4619.399.447.913.562.600.450.30
    Sideboard20.9821.1210.2011.964.906.531.261.22
    Table 1. Performance comparison between the weight-sharing (WS) strategy and proposed strategy
    MethodBP7 /%MSE×100
    BxCoDnSbMnBdAveBxCoDnSbMnBdAve
    LF23.07.819.022.010.710.315.517.439.171.165.070.691.15.77
    LF_PAC22.67.49.110.19.42.910.29.927.050.921.100.780.603.40
    SPO14.33.22.57.36.51.85.910.884.140.390.990.860.432.95
    CAE17.93.45.09.84.71.57.18.421.500.380.880.410.462.01
    Shi22.13.34.710.16.42.38.29.160.940.501.370.410.432.14
    EPINET15.10.91.96.67.81.45.65.200.250.190.800.600.261.22

    Proposed

    method

    13.91.02.66.54.92.25.26.910.540.301.220.410.451.64
    MethodBP1 /%BP3 /%
    BxCoDnSbMnBdAveBxCoDnSbMnBdAve
    LF65.554.273.859.840.452.757.738.421.445.135.420.823.530.8
    LF_PAC79.262.173.072.946.350.163.946.018.531.330.119.111.826.1
    SPO69.558.365.870.045.647.559.526.711.415.025.314.08.016.7
    CAE72.759.261.156.942.248.056.740.415.521.326.811.38.220.6
    Shi68.438.245.663.639.635.748.538.510.614.526.314.38.018.7
    EPINET62.451.741.058.642.539.949.227.44.96.621.014.55.213.3

    Proposed

    method

    32.830.119.421.140.941.330.919.58.07.912.011.47.711.1
    Table 2. Objective quality comparison of depth maps with different methods
    MethodBoxCottonDinoSideboard
    EPINET237.33125.2644.5894.40
    Proposed method43.202.824.8626.16
    Table 3. Variance of MSE of scenes in Fig. 7
    MethodBoxesCottonDinoSideboardBicycleBedroomOrigamiHerbs
    EPINET1.159111.163941.155581.142251.077071.084611.079971.08541
    Proposed method0.420780.416220.420400.421680.405370.403040.403880.39573
    Table 4. Comparison of test time
    Sen Xiang, Nanting Huang, Huiping Deng, Jin Wu. Estimation of Light Field Depth Based on Multi-Level Network Optimization[J]. Laser & Optoelectronics Progress, 2022, 59(10): 1010009
    Download Citation