Opto-Electronic Engineering, Volume. 51, Issue 1, 230304-1(2024)

Design of Swin Transformer for semantic segmentation of road scenes

Hao Hang... Yingping Huang*, Xurui Zhang and Xin Luo |Show fewer author(s)
Author Affiliations
  • School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
  • show less
    Figures & Tables(13)
    Network architecture
    Swin Transformer architecture
    Swin Transformer block
    Patch Merging module
    FCM module
    AFM module
    Comparison of segmentation effects of multiple methods in Cityscapes scenes
    Comparison of ablation experiment effects
    • Table 1. Experimental environment

      View table
      View in Article

      Table 1. Experimental environment

      实验环境配置实验环境配置
      CPUAMD5600XdCPU核心数6
      GPUNVIDIA RTX3070主频3.7 GHz
      内存32 G显存11 G
      操作系统Ubuntu18.04编程语言Python 3.7
      深度学习框架Pytorch 1.10.0CUDA10.2
    • Table 2. IoU and MIoU of various models on the Cityscapes dataset

      View table
      View in Article

      Table 2. IoU and MIoU of various models on the Cityscapes dataset

      ClassesFCNPSPNetUNetDeepLabv3SwinTOurs
      Road97.198.098.098.198.098.1
      Sidewalk79.981.884.284.584.786.2
      Building89.391.191.191.791.491.6
      Wall44.248.248.751.254.455.5
      Fence48.350.351.553.657.359.9
      Pole30.645.748.250.355.557.2
      Traffic Light44.750.051.753.761.963.2
      Traffic Sign56.862.365.868.273.574.4
      Vegetation87.189.290.190.190.292.4
      Terrain60.462.865.364.261.363.2
      Sky90.894.293.895.394.295.1
      Person64.171.272.674.575.576.9
      Rider38.245.646.149.555.755.9
      Car90.492.092.292.693.893.5
      Truck51.368.563.474.473.672.5
      Bus72.080.377.683.279.479.9
      Train74.477.478.581.577.778.1
      Motocycle52.550.155.553.556.559.2
      Bicycle59.160.163.464.271.273.2
      MIoU/%64.9269.2870.4573.7173.1775.18
    • Table 3. PA and MPA of various models on the Cityscapes dataset

      View table
      View in Article

      Table 3. PA and MPA of various models on the Cityscapes dataset

      ClassesFCNPSPNetUNetDeepLabv3SwinTOurs
      Road98.198.598.899.199.199.1
      Sidewalk89.989.390.292.091.292.7
      Building96.394.796.196.296.596.8
      Wall52.272.160.773.171.472.3
      Fence60.369.368.572.571.474.6
      Pole36.674.759.274.374.177.7
      Traffic Light56.772.062.769.270.472.1
      Traffic Sign68.879.375.876.576.779.3
      Vegetation94.193.295.193.695.397.7
      Terrain74.479.878.378.179.280.3
      Sky95.897.297.897.597.597.9
      Person77.182.284.684.286.387.9
      Rider58.268.655.171.272.473.7
      Car96.496.096.296.397.697.6
      Truck62.379.576.475.573.576.2
      Bus85.087.389.691.785.687.7
      Train78.483.492.588.479.382.9
      Motocycle66.573.567.577.577.379.2
      Bicycle77.173.180.476.280.384.2
      MPA/%74.6479.9780.0682.3181.5984.83
    • Table 4. Performance comparison of various semantic segmentation algorithms

      View table
      View in Article

      Table 4. Performance comparison of various semantic segmentation algorithms

      方法MIoU/%MPA/%Param/MFLOPs/GFPS
      FCN64.9274.6434.9066.3858.61
      PSPNet69.2879.9751.86152.9781.25
      UNet70.4580.0649.10166.9254.52
      DeepLabv373.7182.3168.37235.3736.59
      SwinT73.1781.59121.25297.5712.22
      Ours75.1884.83123.77305.4614.83
    • Table 5. Ablation experiment

      View table
      View in Article

      Table 5. Ablation experiment

      实验序号AFMFCMASPPMIoU/%MPA/%
      注:“√”表示网络中包含该结构,“×”表示在网络中去掉该结构。
      ×××73.181.6
      ××73.880.5
      ×74.983.3
      75.284.8
    Tools

    Get Citation

    Copy Citation Text

    Hao Hang, Yingping Huang, Xurui Zhang, Xin Luo. Design of Swin Transformer for semantic segmentation of road scenes[J]. Opto-Electronic Engineering, 2024, 51(1): 230304-1

    Download Citation

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

    Category: Article

    Received: Dec. 14, 2023

    Accepted: Jan. 24, 2024

    Published Online: Apr. 19, 2024

    The Author Email: Huang Yingping (黄影平)

    DOI:10.12086/oee.2024.230304

    Topics