Journal of Applied Optics, Volume. 44, Issue 5, 1010(2023)

Multi-scale oriented object detection based on improved RoI Transformer in remote sensing images

Minhao LIU1...2, Kun WANG1,2, Ruijiao JIN1,2, Tian LU2, and Zhang LI12,* |Show fewer author(s)
Author Affiliations
  • 1College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410000, China
  • 2Hunan Province Key Laboratory of Image Measurement and Vision Navigation, National University of Defense Technology, Changsha 410000, China
  • show less
    Figures & Tables(17)
    Comparison between remote sensing images (the first row) and natural images (the second row)
    Structure diagram of HRD-ROI Transformer
    Structure diagram of HRNet[18]
    Schematic diagram of angle boundary discontinuity
    Schematic diagram of square-like problem
    Comparison of detection results (false detection)
    Comparison of detection results (missed detection)
    Comparison of detection results (objects of large aspect ratios)
    Effectiveness of KLD on DIOR-R dataset
    Detection results of airport
    Detection results of golf course
    • Table 1. Performance comparison of different methods on DOTAv1.0 dataset

      View table
      View in Article

      Table 1. Performance comparison of different methods on DOTAv1.0 dataset

      MethodBackboneLossAP/%mAP/%
      PLBDBRGTFSVLVSHTCBCSTSBFRAHASPHC
      One-stage
      Rotated RetinaNetResNet50Smooth L189.775.040.864.166.567.785.890.762.665.754.462.062.652.254.566.3
      R3DetResNet50Smooth L189.573.244.465.366.977.287.290.857.966.251.363.272.153.054.667.5
      S2ANetResNet50Smooth L189.073.843.667.164.974.279.190.562.766.356.864.861.254.242.066.0
      SASM reppointsResNet50GIoU89.576.045.370.759.974.678.090.364.167.346.267.170.356.344.366.7
      Oriented reppointsResNet50GIoU89.775.749.870.774.180.588.490.565.168.647.164.670.457.854.669.8
      Two-stage
      Rotated Faster RCNNResNet50Smooth L188.574.744.170.063.771.479.490.558.762.054.764.563.258.250.166.3
      Oriented RCNNResNet50Smooth L189.175.850.068.362.384.088.890.668.762.357.063.666.457.339.168.2
      RoI TransformerResNet50Smooth L189.477.746.871.968.477.980.090.771.362.559.163.667.360.245.468.8
      ReDetReResNet50Smooth L189.678.047.468.865.882.487.490.667.569.763.465.967.353.048.769.7
      OursHRNetKLD89.875.454.778.968.878.689.390.775.762.867.067.275.360.752.172.5
    • Table 2. Performance comparison of different methods on DIOR-R dataset

      View table
      View in Article

      Table 2. Performance comparison of different methods on DIOR-R dataset

      MethodBackboneLossAP/%mAP/%
      APLAPOBFBCBRCHESAETSDAMGFGTFHAOPSHSTASTOTCTSVEWM
      One-stage
      Rotated RetinanetResNet50Smooth L159.115.070.481.114.572.664.946.614.670.974.724.830.267.069.150.181.241.632.561.952.1
      Rotated Retinanet-GResNet50GWD64.621.172.981.113.172.768.545.814.770.175.127.230.668.966.157.981.247.434.861.553.8
      R3DetResNet50Smooth L153.327.968.981.022.972.666.449.619.268.476.022.141.568.357.955.481.145.535.754.053.4
      R3Det-KResNet50KLD57.834.969.481.228.572.771.853.216.171.877.136.447.674.562.560.881.350.039.856.257.2
      S2ANetResNet50KFIoU67.228.076.080.827.372.661.260.317.968.678.226.244.677.765.867.481.348.942.263.157.8
      SASM reppointsResNet50GIoU61.252.174.582.732.472.576.058.134.971.377.138.651.579.164.866.380.760.541.764.262.0
      Oriented reppointsResNet50GIoU68.741.975.184.035.475.479.565.832.175.078.643.451.880.366.566.485.454.046.265.063.5
      Two-stage
      Rotated Faster RCNNResNet50Smooth L162.018.171.381.022.972.561.058.510.067.678.834.338.980.458.862.481.344.741.364.355.5
      Oriented RCNNResNet50Smooth L161.826.771.681.333.872.674.058.423.766.880.029.952.081.062.562.481.450.642.365.058.9
      RoI TransformerResNet50Smooth L163.130.771.881.533.972.775.864.624.367.482.535.751.181.270.570.881.544.443.466.060.7
      ReDetReResNet50Smooth L171.028.371.588.731.372.771.661.120.861.881.936.748.881.163.162.581.649.242.864.659.6
      OursHRNetKLD63.141.679.088.042.172.676.665.828.271.082.942.257.181.372.570.489.753.349.166.364.7
    • Table 3. Detection effects of small object on DOTAv1.0 and DIOR-R datasets

      View table
      View in Article

      Table 3. Detection effects of small object on DOTAv1.0 and DIOR-R datasets

      MethodBackboneLossDIOR-R/%DOTAv1.0/%
      SHVEWMSVSH
      One-stage
      Rotated RetinaNetResNet50Smooth L167.032.561.966.585.8
      R3DetResNet50Smooth L168.335.754.066.987.2
      S2ANetResNet50Smooth L177.742.263.164.979.1
      SASM reppointsResNet50GIoU79.141.764.259.978.0
      Oriented reppointsResNet50GIoU80.346.265.074.188.4
      Two-stage
      Rotated Faster RCNNResNet50Smooth L180.441.364.363.779.4
      Oriented RCNNResNet50Smooth L181.042.365.062.388.8
      RoI TransformerResNet50Smooth L181.243.466.068.480.0
      ReDetReResNet50Smooth L181.142.864.665.887.4
      OursHRNetKLD81.349.166.368.889.3
    • Table 4. Comparison of effectiveness of KLD and HRNet on DOTAv1.0 dataset

      View table
      View in Article

      Table 4. Comparison of effectiveness of KLD and HRNet on DOTAv1.0 dataset

      MethodKLDHRNetmAP/%
      Rotated Faster RCNN66.3
      RoI Transformer68.8
      Ours(a)70.3
      Ours(b)71.7
      Ours(c)72.5
    • Table 5. Comparison of effectiveness of KLD and HRNet on DIOR-R dataset

      View table
      View in Article

      Table 5. Comparison of effectiveness of KLD and HRNet on DIOR-R dataset

      MethodKLDHRNetmAP/%
      Rotated Faster RCNN55.5
      RoI Transformer60.7
      Ours(a)61.5
      Ours(b)63.9
      Ours(c)64.7
    • Table 6. Comparison of mAP for three loss function models

      View table
      View in Article

      Table 6. Comparison of mAP for three loss function models

      Loss FunctionDOTAv1.0/%DIOR-R/%
      GWD69.261.4
      KFIOU68.960.3
      KLD70.361.5
    Tools

    Get Citation

    Copy Citation Text

    Minhao LIU, Kun WANG, Ruijiao JIN, Tian LU, Zhang LI. Multi-scale oriented object detection based on improved RoI Transformer in remote sensing images[J]. Journal of Applied Optics, 2023, 44(5): 1010

    Download Citation

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

    Category: Research Articles

    Received: Jul. 7, 2023

    Accepted: --

    Published Online: Mar. 12, 2024

    The Author Email: LI Zhang (李璋)

    DOI:10.5768/JAO202344.0502001

    Topics