Fig 1.
Self-attention mechanism.
Fig 2.
Multi-head self-attention mechanism.
Fig 3.
Overall model architecture.
Fig 4.
The process of regular window partition and reverse.
Fig 5.
The process of shifting window partition and reverse.
Table 1.
Performance comparison of our method with baselines on the Market1501, DukeMTMC-reID and MSMT17 dataset.
Fig 6.
Loss and top1 error curve with Market1501 dataset.
Fig 7.
Loss and top1 error curve with DukeMTMC-reID dataset.
Fig 8.
Loss and top1 error curve with MSMT17 dataset.
Fig 9.
ROC curve with Market1501 dataset.
Fig 10.
ROC curve with DukeMTMC-reID dataset.
Fig 11.
ROC curve with MSMT17 dataset.
Fig 12.
Example 1 of ranking results.
Fig 13.
Example 2 of ranking results.
Fig 14.
Example 3 of ranking results.
Fig 15.
The examples of the feature visualization for different methods.
Table 2.
Ablation experiments of our method on the Market1501, DukeMTMC-reID and MSMT17 datasets.
Table 3.
Comparison of computation efficiency among different methods.