• Acta Optica Sinica
  • Vol. 42, Issue 19, 1915001 (2022)
Kaiyi Zhang1、2, Ru Hong1、2, Shaoyan Gai1、2, and Feipeng Da1、2、3、*
Author Affiliations
  • 1School of Automation, Southeast University, Nanjing 210096, Jiangsu , China
  • 2Key Laboratory of Measurement and Control of Complex Engineering Systems, Ministry of Education, Southeast University, Nanjing 210096, Jiangsu , China
  • 3Shenzhen Research Institute, Southeast University, Shenzhen 518036, Guangdong , China
  • show less
    DOI: 10.3788/AOS202242.1915001 Cite this Article Set citation alerts
    Kaiyi Zhang, Ru Hong, Shaoyan Gai, Feipeng Da. Three-Dimensional Human Hand Pose Estimation Based on Finger-Point Reinforcement and Multi-Level Feature Fusion[J]. Acta Optica Sinica, 2022, 42(19): 1915001 Copy Citation Text show less
    Overview of overall algorithm flow
    Fig. 1. Overview of overall algorithm flow
    Comparison of three groups of gestures (black dots on top are parts with higher confidence parameters, and original point clouds are shown on bottom)
    Fig. 2. Comparison of three groups of gestures (black dots on top are parts with higher confidence parameters, and original point clouds are shown on bottom)
    SE module and MFSE module. (a) SE module; (b) MFSE module
    Fig. 3. SE module and MFSE module. (a) SE module; (b) MFSE module
    Error of each finger joint point by different methods
    Fig. 4. Error of each finger joint point by different methods
    Proportion of test frames with errors within different thresholds
    Fig. 5. Proportion of test frames with errors within different thresholds
    Comparison of experimental results under ICVL dataset
    Fig. 6. Comparison of experimental results under ICVL dataset
    Comparison of experimental results under MSRA dataset
    Fig. 7. Comparison of experimental results under MSRA dataset
    JointBaselineBaseline+FPR
    Mean error8.55308.1930

    Palm

    Index_R

    Index_T

    Mid_R

    Mid_T

    Ring_R

    Ring_T

    8.6665

    6.8979

    11.1541

    5.4392

    10.9231

    5.7619

    9.7697

    8.2581

    6.4721

    10.5203

    5.1762

    10.3742

    5.5396

    9.4127

    Pinky_R7.44917.1871

    Pinky_T

    Thumb_R

    Thumb_T

    10.2487

    7.5354

    13.9441

    9.6389

    7.2211

    13.7389

    Table 1. Error distance of each joint of FPR strategy on MSRA dataset
    JointBaselineBaseline+SEBaseline+MFSE
    Mean error8.53208.32808.2050
    Palm8.66658.49588.3172
    Index_R6.89796.58296.5038
    Index_T11.154110.696010.4891
    Mid_R5.43925.21405.1937
    Mid_T10.923110.283510.1443
    Ring_R5.76195.61995.5399
    Ring_T9.76979.41829.2983
    Pinky_R7.44917.33867.1515
    Pinky_T10.248710.16669.7651
    Thumb_R7.53547.41017.2387
    Thumb_T13.944113.907213.7871
    Table 2. Error distance of each joint on MSRA dataset by different methods
    MethodMean error in MSRAMean error in ICVL
    Hand PointNet128.5056.935
    3D DenseNet217.986.77
    SHPR-NET227.967.2
    CNN Model238.37.1
    Bayesian DeepPrior2510.1
    Pose REN248.66.79
    SO-HandNet137.7
    PCHPS267.1178.893
    PointNet+MFSE8.2036.854
    PointNet+FPR8.1926.728
    PointNet+MFSE+FPR7.9426.673
    Table 3. Average error distance of each method on MSRA and ICVL datasets
    Kaiyi Zhang, Ru Hong, Shaoyan Gai, Feipeng Da. Three-Dimensional Human Hand Pose Estimation Based on Finger-Point Reinforcement and Multi-Level Feature Fusion[J]. Acta Optica Sinica, 2022, 42(19): 1915001
    Download Citation