• Laser & Optoelectronics Progress
  • Vol. 59, Issue 12, 1215009 (2022)
Dexiang Zhang1、2、*, Peicheng Yuan1、**, and Jun Wang1、***
Author Affiliations
  • 1School of Electrical Engineering and Automation, Anhui University, Hefei 230601, Anhui , China
  • 2School of Electronic and Electrical Engineering, Anhui Sanlian University, Hefei 230601, Anhui , China
  • show less
    DOI: 10.3788/LOP202259.1215009 Cite this Article Set citation alerts
    Dexiang Zhang, Peicheng Yuan, Jun Wang. Person Reidentification Based on Multiscale Batch Feature-Discarding Network[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215009 Copy Citation Text show less
    Structure of IBN module
    Fig. 1. Structure of IBN module
    Process of batch feature discarding
    Fig. 2. Process of batch feature discarding
    Comparison of different discarding methods with the same batch
    Fig. 3. Comparison of different discarding methods with the same batch
    Structure of person re-identification network based on multi-scale batch feature discarding network
    Fig. 4. Structure of person re-identification network based on multi-scale batch feature discarding network
    Learning process of triplet loss
    Fig. 5. Learning process of triplet loss
    Experimental results under different ε. (a) Experimental results under Market1501 dataset; (b) experimental results under DukeMTMC-reID dataset
    Fig. 6. Experimental results under different ε. (a) Experimental results under Market1501 dataset; (b) experimental results under DukeMTMC-reID dataset
    Visualization results of Market1501
    Fig. 7. Visualization results of Market1501
    Visualization results of DukeMTMC-reID
    Fig. 8. Visualization results of DukeMTMC-reID
    DatasetDetail information
    IDImageCameraLabelYear
    Market15011501326686Hand&Auto2015
    DukeMTMC-reID1404364118Hand&Auto2017
    CUHK031467140972Hand&Auto2014
    Occluded-Duke1221332798Hand&Auto2019
    Table 1. Statistics for datasets
    MethodMarket1501DukeMTMC-reID
    Rank-1mAPRank-1mAP
    ResNet50(without FD and IBN)93.884.086.773.1
    ResNet50(with IBN)95.186.188.175.2
    ResNet50(with FD)94.084.487.174.0
    ResNet50(with FD and IBN)95.386.888.575.9
    Table 2. Experimental results of different ResNet50
    εMarket1501DukeMTMC-reID
    Rank-1mAPRank-1mAP
    0.0595.186.288.274.7
    0.195.386.888.575.9
    0.1594.785.688.475.5
    0.294.684.988.175.2
    0.2594.184.687.874.3
    0.393.884.587.973.9
    Table 3. Experimental results under different ε
    MethodMarket1501DukeMTMC-reID
    Rank-1mAPRank-1mAP
    ResNet50(without FD and IBN)92.882.185.169.8
    ResNet50(without FD and IBN)+RE93.182.685.771.5
    ResNet50(without FD and IBN)+Cut93.884.086.773.1
    ResNet50(with FD and IBN)94.084.487.773.9
    ResNet50(with FD and IBN)+RE94.385.087.975.4
    ResNet50(with FD and IBN)+Cut95.386.888.575.9
    Table 4. Experimental results of Cut and RE
    MethodMarket1501DukeMTMC-reID
    Rank-1mAPRank-1mAP
    Baseline91.276.983.565.1
    +warmup92.782.686.072.3
    +Cutout93.983.986.373.5
    +LS94.585.287.674.8
    +stride=195.386.888.575.9
    Table 5. Experimental results under different training tricks
    MethodMarket1501DukeMTMC-reID
    Rank-1mAPRank-1mAP
    IDE72.546.067.747.1
    SVDNet82.362.176.756.8
    HA-CNN91.275.780.563.8
    SVDNet+Era87.171.379.362.4
    IANet94.483.187.173.4
    PCB92.477.381.965.3
    PCB+RPP93.881.683.369.2
    MGN95.786.988.778.4
    Ours95.386.888.575.9
    Ours+Re-ranking95.493.190.487.5
    Table 6. Comparison of experimental results of different methods on Market1501 and DukeMTMC-reID
    MethodCUHK03-LabeledCUHK03-Detected
    Rank-1mAPRank-1mAP
    IDE22.221.021.319.7
    SVDNet41.537.3
    HA-CNN44.441.041.738.6
    SVDNet+Era49.445.048.737.2
    PCB61.354.2
    PCB+RPP62.856.7
    MGN68.067.466.866.0
    Ours80.977.877.974.9
    Ours+Re-ranking86.588.184.485.9
    Table 7. Comparison of experimental results of different methods on CUHK03
    MethodRank-1mAP
    Part-Aligned28.820.2
    PCB42.633.7
    SFR42.332
    PGFA51.437.3
    HOReID55.143.8
    Ours58.746.4
    Ours+Re-ranking61.761.2
    Table 8. Comparison of experimental results of different methods on Occluded-Duke
    Dexiang Zhang, Peicheng Yuan, Jun Wang. Person Reidentification Based on Multiscale Batch Feature-Discarding Network[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215009
    Download Citation