• Opto-Electronic Engineering
  • Vol. 52, Issue 1, 240238 (2025)
Zhongmin Liu1,*, Fujun Yang1, and Wenjin Hu2
Author Affiliations
  • 1Department of Electrical Engineering and Information Engineering, Lanzhou University of Technology, Lanzhou, Gansu 730050, China
  • 2College of Mathematics and Computer Science, Northwest Minzu University, Lanzhou, Gansu 730030, China
  • show less
    DOI: 10.12086/oee.2025.240238 Cite this Article
    Zhongmin Liu, Fujun Yang, Wenjin Hu. Multi-scale feature interaction pseudo-label unsupervised domain adaptation for person re-identification[J]. Opto-Electronic Engineering, 2025, 52(1): 240238 Copy Citation Text show less
    Overall framework of the MSFINet model
    Fig. 1. Overall framework of the MSFINet model
    FSA mechanism
    Fig. 2. FSA mechanism
    Framework of the ReFIM
    Fig. 3. Framework of the ReFIM
    Structure of PCBN module
    Fig. 4. Structure of PCBN module
    Partial convolution ablation experiment results of the PCBN module
    Fig. 5. Partial convolution ablation experiment results of the PCBN module
    Performance of different train epochs on mainstream datasets
    Fig. 6. Performance of different train epochs on mainstream datasets
    Ranking list of retrieval results with different models
    Fig. 7. Ranking list of retrieval results with different models
    MethodDuke-to-MarketMarket-to-DukeMarket-to-MSMT
    mAPRank-1Rank-5mAPRank-1Rank-5mAPRank-1Rank-5
    “—”代表结果未显示
    SECRET[5]79.892.367.180.324.349.9
    FSA81.192.896.568.383.589.424.551.863.8
    PCBN80.792.597.467.180.688.724.350.963.5
    ReFIM80.092.997.267.280.489.224.651.763.9
    FSA+PCBN81.292.497.166.080.489.224.751.964.1
    FSA+ReFIM81.392.897.464.179.088.225.352.364.7
    PCBN+ReFIM82.793.098.067.680.788.224.552.564.3
    FSA +PCBN+ReFIM82.993.797.468.782.789.826.654.767.5
    Table 1. Ablation experimental results of FSA,PCBN, and ReFIM on mainstream datasets (Unit: %)
    MethodDuke-to-MarketMarket-to-DukeMarket-to-MSMT
    mAPRank-1Rank-5mAPRank-1Rank-5mAPRank-1Rank-5
    “—”代表结果未显示
    SECRET[5]79.892.367.180.324.349.9
    CDC79.892.596.965.479.788.824.450.964.1
    FIM79.891.897.166.180.189.524.351.463.7
    ReFIM80.092.997.267.280.489.224.651.763.9
    Table 2. Experiments results of the CDC and FIM ablation in the ReFIM module (Unit: %)
    MethodEvaluation metrics
    Params/MFLOPs/GLatency/ms
    Bottleneck (SECRET)[5]0.52.223
    Conv1_spling0.52.020
    Conv2_spling0.41.717
    Conv3_spling0.31.512
    Table 3. Partial convolution efficiency ablation experiments of the PCBN module
    MethodDuke-to-MarketMarket-to-DukeMarket-to-MSMT
    mAPRank-1Rank-5mAPRank-1Rank-5mAPRank-1Rank-5
    “—”代表结果未显示
    SECRET[5]79.892.367.180.324.349.9
    SE80.092.396.362.481.887.023.348.362.5
    CBAM79.892.696.465.382.288.124.451.263.3
    CA67.787.095.363.280.586.422.550.662.9
    FSA81.192.896.568.383.589.424.551.863.8
    Table 4. Efficiency ablation experiment results of the FSA (Unit: %)
    MethodEvaluation metrics
    Horizontal squeezeVertical squeezeDetail enhancementParams/MFLOPs/GLatency/ms
    1.31.643
    1.21.544
    1.41.847
    1.11.339
    Table 5. Experimental results of different parts in the FSA module
    ModelsDuke-to-MarketMarket-to-DukeMarket-to-MSMT
    mAPRank-1Rank-5mAPRank-1Rank-5mAPRank-1Rank-5
    “—”代表结果未显示
    HHL[30]31.462.278.827.246.961.0
    ECN[31]43.075.187.640.463.375.810.215.240.4
    UDL[32]44.776.888.941.364.977.0
    PDA-Net[32-33]47.675.286.345.163.277.0
    PCB-PAST[34]54.678.454.372.4
    SSG[3]58.380.090.053.473.080.613.231.6
    MMCL[35]60.484.492.851.472.482.915.140.8
    SNR[36]61.782.858.176.3
    ECN++[37]63.884.192.854.474.083.715.240.4
    AD-Cluster[38]68.386.794.454.172.682.5
    MMT[4]71.287.794.965.178.088.822.949.263.1
    MEB-NET[39]76.089.996.066.179.688.3
    GCL+ (JVTC+)[40]76.591.696.368.382.689.427.153.866.9
    UNRN[41]78.191.996.169.182.090.725.352.464.7
    SECRET[5] (baseline)79.892.367.180.324.349.9
    HQP[42]80.392.396.968.082.690.223.649.862.4
    DHCL[43]81.592.897.267.381.189.3
    MSFINet (ours)82.993.797.468.782.789.826.654.767.5
    Table 6. Performance comparison among different models on mainstream datasets (Unit: %)
    Zhongmin Liu, Fujun Yang, Wenjin Hu. Multi-scale feature interaction pseudo-label unsupervised domain adaptation for person re-identification[J]. Opto-Electronic Engineering, 2025, 52(1): 240238
    Download Citation