Laser & Optoelectronics Progress, Volume. 61, Issue 4, 0415004(2024)

3D Reconstruction of Neural Radiation Field Based on Improved Multiple Layer Perceptron

Yaofei Hou1, Haisong Huang1,2、*, Qingsong Fan1, Jing Xiao1, and Zhenggong Han1
Author Affiliations
  • 1Key Laboratory of Advanced Manufacturing Technology of the Ministry of Education, Guizhou University, Guiyang 550025, Guizhou, China
  • 2Information Engineering Institute, Chongqing Vocational and Technical University of Mechatronics, Chongqing 402760, China
  • show less
    Figures & Tables(14)
    Structure of NeRF network
    Structure of IP-NeRF network
    Multi-feature joint learning module
    Gated channel tranformation MLP module
    Visualized new view reconstruction results under the modules in Trex and Lego scenes
    Visualized new view reconstruction results of different methods in the selected scenes of the three datasets
    • Table 1. Ablation experiment of the new view reconstruction in Trex scene

      View table

      Table 1. Ablation experiment of the new view reconstruction in Trex scene

      Ablation experimentPSNR/dBSSIMLPIPS
      A26.90(+0.37%)0.881(+0.11%)0.057
      B28.23(+5.33%)0.925(+5.11%)0.051(-10.53%)
      C28.56(+6.56%)0.930(+5.68%)0.046(-19.30%)
      D28.74(+7.24%)0.934(+6.13%)0.042(-26.31%)
    • Table 2. Ablation experiment of the new view reconstruction in Lego scene

      View table

      Table 2. Ablation experiment of the new view reconstruction in Lego scene

      Ablation experimentPSNR/dBSSIMLPIPS
      A32.62(+0.24%)0.9610.024
      B33.94(+4.33%)0.968(+0.73%)0.017(-29.17%)
      C34.37(+5.62%)0.971(+1.04%)0.016(-33.33%)
      D34.77(+6.85%)0.974(+1.35%)0.015(-37.50%)
    • Table 3. Parameter comparison of different methods on Realistic Synthetic 360° dataset

      View table

      Table 3. Parameter comparison of different methods on Realistic Synthetic 360° dataset

      SceneNeRFNeRF-IDIP-NeRF
      PSNR/dBSSIMLPIPSPSNR/dBSSIMLPIPSPSNR/dBSSIMLPIPS
      Mean31.010.9470.04132.340.9570.02932.750.9600.026
      Chair33.000.9670.01934.540.9780.01435.170.9830.010
      Drums25.010.9250.05825.150.9260.05725.800.9310.051
      Ficus30.130.9640.02232.240.9760.01531.860.9730.016
      Hotdog36.180.9740.01637.260.9810.01338.480.9860.010
      Lego32.540.9610.02434.730.9740.01534.770.9740.015
      Materials29.620.9490.02930.370.9560.02431.900.9770.011
      Mic32.910.9800.02334.710.9880.00934.210.9820.018
      Ship28.650.8560.11929.750.8760.08129.790.8760.081
    • Table 4. Parameter comparison of different methods on Real Forward-Facing dataset

      View table

      Table 4. Parameter comparison of different methods on Real Forward-Facing dataset

      SceneNeRFNeRF-IDIP-NeRF
      PSNR /dBSSIMLPIPSPSNR /dBSSIMLPIPSPSNR /dBSSIMLPIPS
      Mean26.500.8110.07326.760.8220.07028.080.8870.061
      Fern25.200.7920.09225.010.8000.08927.080.8680.079
      Flower27.400.8270.06127.850.8420.05828.820.9010.053
      Fortress31.160.8810.03031.510.8880.02832.940.9330.024
      Horns27.450.8280.06827.880.8430.06529.300.9110.057
      Leaves20.920.6900.11121.090.7080.10822.530.8250.100
      Orchids20.360.6410.12120.380.6430.12021.440.7640.100
      Room32.700.9480.04132.930.9540.03933.860.9610.035
      Trex26.800.8800.05727.450.8970.05128.740.9340.042
    • Table 5. Parameter comparison of different methods on DTU dataset

      View table

      Table 5. Parameter comparison of different methods on DTU dataset

      SceneNeRFNeRF-IDIP-NeRF
      PSNR /dBSSIMLPIPSPSNR /dBSSIMLPIPSPSNR /dBSSIMLPIPS

      Scan1

      Scan22

      Scan55

      Scan109

      23.490.7540.28223.800.7650.26624.470.7780.248
      21.550.7080.23821.980.7150.22622.680.7580.196
      26.540.7940.22926.760.8000.21927.230.8120.206
      28.330.8600.23628.630.8700.22629.460.8810.185
      Mean24.980.7790.24625.290.7870.23425.960.8070.208
    • Table 6. Calculation cost comparison of different methods

      View table

      Table 6. Calculation cost comparison of different methods

      DatasetNeRFNeRF-IDIP-NeRF
      PSNR /dBTrain-time /hRender-time /(s/it)PSNR /dBTrain-time /hRender-time /(s/it)PSNR /dBTrain-time /hRender-time/(s/it)
      Realistic Synthetic 360° Real Forward-Facing DTU31.0118.421.1832.3414.917.1032.7519.522.24
      26.5016.520.1026.7613.316.1728.0817.521.20
      24.9819.436.6025.2915.630.4825.9620.538.63
    • Table 7. Calculation cost comparison of simplified network

      View table

      Table 7. Calculation cost comparison of simplified network

      DatasetNeRF-IDSIP-NeRF
      PSNR /dBTrain-time /hRender-time /(s/it)PSNR /dBTrain-time /hRender-time /(s/it)
      Realistic Synthetic 360° Real Forward-Facing DTU32.3414.917.1032.4015.017.08
      26.7613.316.1727.6813.316.10
      25.2915.630.4825.6415.630.39
    • Table 8. Comprehensive performance analysis of different methods

      View table

      Table 8. Comprehensive performance analysis of different methods

      DatasetParameterNeRF13NSVF15GRF17NeuSample21NeXT20IP-NeRF
      Realistic Synthetic 360°PSNR /dB31.0131.7532.0631.1534.4032.75
      Train-time/h18.41.523.014.052.719.5
      Real Forward-FacingPSNR /dB26.50/26.6426.83/28.08
      Train-time /h16.5/20.612.5/17.5
    Tools

    Get Citation

    Copy Citation Text

    Yaofei Hou, Haisong Huang, Qingsong Fan, Jing Xiao, Zhenggong Han. 3D Reconstruction of Neural Radiation Field Based on Improved Multiple Layer Perceptron[J]. Laser & Optoelectronics Progress, 2024, 61(4): 0415004

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Machine Vision

    Received: Dec. 13, 2022

    Accepted: Mar. 1, 2023

    Published Online: Feb. 26, 2024

    The Author Email: Haisong Huang (hshuang@gzu.edu.cn)

    DOI:10.3788/LOP223312

    CSTR:32186.14.LOP223312

    Topics