• Laser & Optoelectronics Progress
  • Vol. 60, Issue 14, 1410007 (2023)
Dongdong Huo and Haishun Du*
Author Affiliations
  • School of Artificial Intelligence, Henan University, Zhengzhou 450046, Henan, China
  • show less
    DOI: 10.3788/LOP221850 Cite this Article Set citation alerts
    Dongdong Huo, Haishun Du. Cross-Modal Person Re-Identification Based on Channel Reorganization and Attention Mechanism[J]. Laser & Optoelectronics Progress, 2023, 60(14): 1410007 Copy Citation Text show less
    Overall framework of DCA-Net
    Fig. 1. Overall framework of DCA-Net
    Intra-modal feature channel grouping and reorganization module
    Fig. 2. Intra-modal feature channel grouping and reorganization module
    Aggregated feature attention mechanism module
    Fig. 3. Aggregated feature attention mechanism module
    Channel attention module
    Fig. 4. Channel attention module
    Spatial attention module
    Fig. 5. Spatial attention module
    Position attention module
    Fig. 6. Position attention module
    Comparison of visible images and infrared images
    Fig. 7. Comparison of visible images and infrared images
    Settingall-searchindoor-search
    Methodr=1r =10r =20mAP /%r=1r =10r =20mAP /%
    HOG142.7618.3031.904.243.2224.7044.507.25
    BDTR3317.0155.4371.9619.66
    HSME2320.6832.7477.9523.12
    D2RL2228.9070.6082.4029.20
    MAC3433.2679.0490.0936.2236.4362.3671.6337.03
    MSR3537.3583.4093.3438.1139.6489.2997.6650.88
    AlignGAN1142.4085.0093.7040.7045.9087.6094.4054.30
    cmGAN2626.9767.5180.5631.4931.6377.2389.1842.19
    HPILN3641.3684.7894.3142.9545.7791.8298.4656.52
    LZM3745.0089.0695.7745.9449.6692.4797.1559.81
    AGW147.5084.3992.1447.6554.1791.1495.9862.97
    X-modal3849.9289.7995.9650.73
    DDAG1254.7590.3995.8153.0261.0294.0698.4167.98
    Proposed method59.2391.8396.6356.5563.2294.3998.2069.54
    Table 1. Performance comparison of DCA-Net and current state-of-the-art methods on the SYSU-MM01 dataset
    SettingVisible to thermalThermal to visible
    Methodr=1r =10r =20mAP /%r=1r =10r =20mAP /%
    HCML2424.4447.5356.7820.0821.7045.0255.5822.24
    BDTR3333.5658.6167.4332.7632.9258.4668.4331.96
    D2RL2243.4066.1076.3044.10
    HSME2350.8573.3681.6647.0050.1572.4081.0746.16
    MAC3936.4362.3671.6337.0336.2061.6870.9936.63
    MSR3548.4370.3279.9548.67
    EDFL4052.5872.1081.4752.9851.8972.0981.0452.13
    AlignGAN1157.9053.6056.3053.40
    LZM3757.0376.1084.3458.06
    X-modal3862.2183.1391.7260.18
    AGW170.0586.2191.5566.3770.4987.1291.8465.90
    DDAG1269.3486.1991.4963.4668.0685.1590.3161.80
    Proposed method78.1691.7594.6671.1877.6291.6094.4770.56
    Table 2. Performance comparison of DCA-Net and current state-of-the-art methods on RegDB dataset
    BaselineCGSAAFAICGRSYSU-MM01
    Rank-1mAP
    48.1847.64
    50.7549.73
    57.7354.42
    59.2356.55
    Table 3. Experimental study of ablation on SYSU-MM01 dataset
    BaselineConv2Conv3Conv4SYSU-MM01
    Rank-1mAP
    57.7354.42
    58.2754.73
    59.1956.19
    59.2356.55
    Table 4. Experimental results of ICGR inserted different position under SYSU-MM01 dataset
    Loss functionSYSU-MM01RegDB
    Rank-1mAPRank-1mAP
    Lid56.8954.7570.6362.03
    Ltri+Lid57.7354.4272.1866.03
    Le+Ltri+Lid59.2356.5578.1671.18
    Table 5. Effect of different loss functions on model performance
    ModelModel memory /MBTraining time /s
    AGW273234.33
    DDAG362.48299.82
    DCA-Net364237.07
    Table 6. Model complexity analysis
    Dongdong Huo, Haishun Du. Cross-Modal Person Re-Identification Based on Channel Reorganization and Attention Mechanism[J]. Laser & Optoelectronics Progress, 2023, 60(14): 1410007
    Download Citation