Laser & Optoelectronics Progress, Volume. 61, Issue 10, 1015001(2024)

Cross-Modal Person Re-Identification Based on Mask Reconstruction with Dynamic Attention

Kuo Zhang*, Xinyue Fan, Jiahui Li, and Gan Zhang
Author Affiliations
  • School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
  • show less
    Figures & Tables(14)
    MRDA network structure
    Structure of the mask reconstruction module
    Schematic diagram of regional center loss
    Dynamic attention module structure
    Schematic diagram of attention coefficient acquisition
    Experimental results of constraint learning at different stages
    Feature distribution visualization. (a) Baseline feature dimensionality reduction results; (b) MRDA network feature dimensionality reduction results; (c) baseline feature distance distribution; (d) MRDA network feature distance distribution
    Grad-CAM visualization results. (a) Input images; (b) MRDA network visualization results
    Top-10 retrieval results obtained on the SYSU-MM01 dataset. (a) Baseline retrieval results; (b) MRDA network retrieval results
    • Table 1. Comparison of DAM and other attention modules

      View table

      Table 1. Comparison of DAM and other attention modules

      ModuleRank-1 /%mAP /%
      DAM70.5563.89
      BAM66.6362.17
      CBAM66.7460.95
      IWPAM67.7961.29
    • Table 2. Experimental results in all-search single-shot mode of the SYSU-MM01 dataset

      View table

      Table 2. Experimental results in all-search single-shot mode of the SYSU-MM01 dataset

      BaselineMRMDAMRank-1 /%mAP /%
      63.2159.87
      68.3961.82
      68.4763.09
      70.5563.89
    • Table 3. Comparison results with other networks in the all-search mode of the SYSU-MM01 dataset

      View table

      Table 3. Comparison results with other networks in the all-search mode of the SYSU-MM01 dataset

      NetworkRank-1 /%Rank-10 /%Rank-20 /%mAP /%
      Zero-Pad214.8054.1271.3315.95
      cmGAN526.9767.5180.5627.80
      D2RL2628.9070.6082.4029.20
      Hi-CMD2734.9477.5835.94
      DDAG1854.7590.3995.8153.02
      AGW147.5084.3992.1447.65
      MCLNet765.4093.3397.1461.98
      DML2858.4091.2095.8056.10
      MAUMG2961.5959.96
      MSFF3062.9393.6897.6760.62
      MRDA70.5594.9098.5363.89
    • Table 4. Comparison results with other networks in the indoor-search mode of the SYSU-MM01 dataset

      View table

      Table 4. Comparison results with other networks in the indoor-search mode of the SYSU-MM01 dataset

      NetworkRank-1 /%Rank-10 /%Rank-20 /%mAP /%
      Zero-Pad220.5868.3885.7926.92
      cmGAN531.6377.2389.1842.19
      DDAG1861.0294.0698.4167.98
      AGW154.1791.1495.9862.97
      MCLNet772.5696.8899.2076.58
      DML2862.4095.2098.7069.50
      MAUMG2967.0773.58
      MSFF3068.0995.7198.2254.51
      MRDA72.6997.1598.7377.14
    • Table 5. Comparison results with other networks in the visible-to-infrared mode of the RegDB dataset

      View table

      Table 5. Comparison results with other networks in the visible-to-infrared mode of the RegDB dataset

      NetworkRank-1 /%Rank-10 /%Rank-20 /%mAP /%
      Zero-Pad217.7534.2144.3518.90
      cmGAN524.4447.5356.7820.08
      D2RL2643.4066.1076.3044.10
      Hi-CMD2734.9477.5835.94
      DDAG1872.3769.09
      AGW170.0586.2191.5566.37
      MCLNet780.3192.7096.0373.07
      DML2877.6084.30
      MAUMG2983.3978.75
      MSFF3078.0691.3696.1272.43
      MRDA91.8097.4698.6782.08
    Tools

    Get Citation

    Copy Citation Text

    Kuo Zhang, Xinyue Fan, Jiahui Li, Gan Zhang. Cross-Modal Person Re-Identification Based on Mask Reconstruction with Dynamic Attention[J]. Laser & Optoelectronics Progress, 2024, 61(10): 1015001

    Download Citation

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

    Category: Machine Vision

    Received: Jul. 17, 2023

    Accepted: Oct. 9, 2023

    Published Online: Apr. 2, 2024

    The Author Email: Zhang Kuo (s210101189@stu.cqupt.edu.cn)

    DOI:10.3788/LOP231742

    Topics