Optical Instruments, Volume. 44, Issue 4, 16(2022)

A dual-branch network for action recognition

Xiaofei QIN1... Rui CAI1, Meng CHEN2, Wenqi ZHANG2, Changxiang HE1 and Xuedian ZHANG1 |Show fewer author(s)
Author Affiliations
  • 1School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
  • 2Institute of Aerospace System Engineering Shanghai, Shanghai 201109, China
  • show less
    Figures & Tables(12)
    Model network architecture
    Representations of joints (left) and bones (right)
    Self-attention block
    A variant of spatial self-attention block
    • Table 1. Model architecture of slow-fast network

      View table
      View in Article

      Table 1. Model architecture of slow-fast network

      StageSlow pathwayFast pathwayOutput sizes T × S2
      Input data64×2242
      Data layerStride 16, 12Stride 2, 12Slow : 4×2242Fast: 32×2242
      Conv11×72, 64 Stride 1, 225×72, 8 Stride 1, 22Slow: 4×1122Fast: 32×1122
      Pool11×32 max Stride 1, 221×32 max Stride 1, 22Slow: 4×562Fast: 32×562
      Res1 x 31×12, 64 1×32, 64 1×12, 256 3×12, 8 1×32, 8 1×12, 32 Slow :4×562 Fast :32×562
      Res2 x 41×12, 128 1×32, 128 1×12, 512 3×12, 16 1×32, 16 1×12, 64 Slow: 4×282Fast: 32×282
      Res3 x 63×12, 256 1×32, 256 1×12, 1024 3×12, 32 1×32, 32 1×12, 128 Slow: 4×142Fast: 32×142
      Res4 x 33×12, 512 1×32, 512 1×12, 2048 3×12, 64 1×32, 64 1×12, 256 Slow: 4×72Fast: 32×72
      GAPGAP1×2304
    • Table 2. Comparison of parameters and accuracies between different representations

      View table
      View in Article

      Table 2. Comparison of parameters and accuracies between different representations

      InputParameter amount/106Accuracy/%
      CSCV
      J0.8989.194.9
      B0.8987.494.9
      2-stream1.7890.596.0
      J + B (proposed)0.8990.696.0
    • Table 3. Accuracy comparison between different number of stacked GCN blocks

      View table
      View in Article

      Table 3. Accuracy comparison between different number of stacked GCN blocks

      aAccuracy/%
      CSCV
      273.080.0
      386.590.3
      490.696.0
      590.695.9
    • Table 4. Accuracy change caused by the value of a

      View table
      View in Article

      Table 4. Accuracy change caused by the value of a

      aAccuracy/%
      CSCV
      194.998.1
      1/295.498.6
      1/395.699.0
      1/495.399.0
      1/595.398.7
    • Table 5. Accuracy comparison of dual branch network

      View table
      View in Article

      Table 5. Accuracy comparison of dual branch network

      BranchAccuracy/%
      CSCV
      Skeleton90.294.7
      RGB82.390.1
      Skeleton+RGB95.699.0
    • Table 6. Comparison of the accuracy and parameters on NTU-RGBD60 dataset

      View table
      View in Article

      Table 6. Comparison of the accuracy and parameters on NTU-RGBD60 dataset

      AlgorithmParameter amount/106Accuracy/%
      CSCV
      AS-GCN[12]7.4086.894.2
      GR-GCN[11]84.892.4
      2s-AGCN[8]6.9288.595.1
      AGC-LSTM[14]22.8189.295.0
      2s-SDGCN[18]89.695.7
      SGN[15]0.6989.094.5
      DGNN[24]8.1689.996.1
      Shift-GCN(2s)[10]1.4889.796.0
      Shift-GCN(4s)[10]2.9490.796.5
      MS-G3D(Joint)[6]3.2089.495.0
      MS-G3D(2s)[6]6.4091.596.2
      MST-GCN[1]2.0391.795.7
      Hierarchical Action46.895.398.3
      Skeleton2.8590.696.0
      Skeleton + RGB33.795.699.0
    • Table 7. Comparison of the accuracy on NTU-RGBD120 dataset

      View table
      View in Article

      Table 7. Comparison of the accuracy on NTU-RGBD120 dataset

      AlgorithmAccuracy/%
      c-subc-set
      ST-LSTM[25]55.757.9
      GCA-LSTM[26]58.359.2
      Pose Evolution Map[27]64.666.9
      2s-AGCN[8]82.584.9
      Shift-GCN[10]85.987.6
      MS-G3D[6]86.988.4
      SGN[15]79.281.5
      MST-GCN[1]88.587.8
      Hierarchical Action93.794.5
      Skeleton84.585.6
      Skeleton + RGB94.795.2
    • Table 8. Comparison of the accuracy on Kinetics 400 skeleton dataset

      View table
      View in Article

      Table 8. Comparison of the accuracy on Kinetics 400 skeleton dataset

      AlgorithmAccuracy/%
      Top 1Top 5
      ST-GCN[13]30.752.8
      AS-GCN[12]34.856.5
      2s-AGCN[8]36.158.7
      DGNN[25]36.959.6
      MS-AAGCN[7]37.861.0
      MS-G3D[6]38.060.9
      Slow-fast[5]75.692.1
      Skeleton37.660.1
      Skeleton+RGB78.193.3
    Tools

    Get Citation

    Copy Citation Text

    Xiaofei QIN, Rui CAI, Meng CHEN, Wenqi ZHANG, Changxiang HE, Xuedian ZHANG. A dual-branch network for action recognition[J]. Optical Instruments, 2022, 44(4): 16

    Download Citation

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

    Category: APPLICATION TECHNOLOGY

    Received: Dec. 21, 2021

    Accepted: --

    Published Online: Oct. 19, 2022

    The Author Email:

    DOI:10.3969/j.issn.1005-5630.2022.004.003

    Topics