Laser & Optoelectronics Progress, Volume. 58, Issue 20, 2015004(2021)

Human Pose Estimation Model Based on Improved Hourglass Network

Hong Liu, Jie Ma*, and Yujing Chai
Author Affiliations
  • School of Electronics and Information Engineering, Hebei University of Technology, Tianjin 300401, China
  • show less
    Figures & Tables(12)
    Structure of the HN model
    Structure of the IHN module
    Structure of the prediction model
    Principle of expanding resolution. (a) Unpooling; (b) upsampling; (c) convolution; (d) deconvolution
    Experimental results of different methods on the COCO data set. (a) DeepPose[11]; (b) CPM[12]; (c) 8-SHN[13]; (d) Baseline[19]; (e) IHN-OHKM
    AP curves for different methods in COCO data set
    Detection results of our method on the COCO data set
    • Table 1. Test results of different methods on the COCO data set unit: %

      View table

      Table 1. Test results of different methods on the COCO data set unit: %

      MethodAPAP@50AP@75AP@mAP@lAR
      DeepPose[11]66.580.673.663.472.771.9
      CPM[12]66.982.376.365.375.974.4
      8-SHN[13]69.485.379.867.478.676.8
      Baseline[19]72.988.580.269.079.378.2
      IHN73.688.780.569.179.079.1
      IHN-OHKM74.588.880.269.478.579.6
    • Table 2. Calculation efficiency of different methods

      View table

      Table 2. Calculation efficiency of different methods

      MethodAverage processing time /msGFLOPs /(109 times)Number of parameters
      8-SHN[13]6620.34.6×107
      IHN206.41.3×107
      IHN-OHKM236.91.5×107
    • Table 3. Test results of different methods on MPII data set unit: %

      View table

      Table 3. Test results of different methods on MPII data set unit: %

      MethodHeadSho.Elb.Wri.Hip.KneeAnk.Mean
      DeepPose[11]95.494.391.784.089.787.081.389.1
      CPM[12]96.295.092.084.990.587.782.089.8
      8-SHN[13]97.695.792.385.391.388.682.590.5
      Baseline[19]97.096.593.488.592.090.783.291.8
      IHN97.397.894.289.092.690.585.992.5
      IHN-OHKM97.098.995.388.694.090.286.492.9
    • Table 4. Ablation experiments on the COCO validation data set

      View table

      Table 4. Ablation experiments on the COCO validation data set

      Residual module layerDeconvolution layerAP /%Average processing time /ms
      3271.216.6
      4374.618.0
      5475.021.4
    • Table 5. Impact of different M on AP

      View table

      Table 5. Impact of different M on AP

      M6810121417
      AP /%70.475.574.173.072.772.5
    Tools

    Get Citation

    Copy Citation Text

    Hong Liu, Jie Ma, Yujing Chai. Human Pose Estimation Model Based on Improved Hourglass Network[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2015004

    Download Citation

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

    Category: Machine Vision

    Received: Dec. 2, 2020

    Accepted: Jan. 13, 2021

    Published Online: Oct. 14, 2021

    The Author Email: Ma Jie (jma@hebut.edu.cn)

    DOI:10.3788/LOP202158.2015004

    Topics