Author Affiliations
1School of Electrical Engineering and Automation, Anhui University, Hefei 230601, Anhui , China2School of Electronic and Electrical Engineering, Anhui Sanlian University, Hefei 230601, Anhui , Chinashow less
Fig. 1. Structure of IBN module
Fig. 2. Process of batch feature discarding
Fig. 3. Comparison of different discarding methods with the same batch
Fig. 4. Structure of person re-identification network based on multi-scale batch feature discarding network
Fig. 5. Learning process of triplet loss
Fig. 6. Experimental results under different ε. (a) Experimental results under Market1501 dataset; (b) experimental results under DukeMTMC-reID dataset
Fig. 7. Visualization results of Market1501
Fig. 8. Visualization results of DukeMTMC-reID
Dataset | Detail information |
---|
ID | Image | Camera | Label | Year |
---|
Market1501 | 1501 | 32668 | 6 | Hand&Auto | 2015 | DukeMTMC-reID | 1404 | 36411 | 8 | Hand&Auto | 2017 | CUHK03 | 1467 | 14097 | 2 | Hand&Auto | 2014 | Occluded-Duke | 1221 | 33279 | 8 | Hand&Auto | 2019 |
|
Table 1. Statistics for datasets
Method | Market1501 | | DukeMTMC-reID |
---|
Rank-1 | mAP | | Rank-1 | mAP |
---|
ResNet50(without FD and IBN) | 93.8 | 84.0 | | 86.7 | 73.1 | ResNet50(with IBN) | 95.1 | 86.1 | | 88.1 | 75.2 | ResNet50(with FD) | 94.0 | 84.4 | | 87.1 | 74.0 | ResNet50(with FD and IBN) | 95.3 | 86.8 | | 88.5 | 75.9 |
|
Table 2. Experimental results of different ResNet50
ε | Market1501 | | DukeMTMC-reID |
---|
Rank-1 | mAP | | Rank-1 | mAP |
---|
0.05 | 95.1 | 86.2 | | 88.2 | 74.7 | 0.1 | 95.3 | 86.8 | | 88.5 | 75.9 | 0.15 | 94.7 | 85.6 | | 88.4 | 75.5 | 0.2 | 94.6 | 84.9 | | 88.1 | 75.2 | 0.25 | 94.1 | 84.6 | | 87.8 | 74.3 | 0.3 | 93.8 | 84.5 | | 87.9 | 73.9 |
|
Table 3. Experimental results under different ε
Method | Market1501 | | DukeMTMC-reID |
---|
Rank-1 | mAP | | Rank-1 | mAP |
---|
ResNet50(without FD and IBN) | 92.8 | 82.1 | | 85.1 | 69.8 | ResNet50(without FD and IBN)+RE | 93.1 | 82.6 | | 85.7 | 71.5 | ResNet50(without FD and IBN)+Cut | 93.8 | 84.0 | | 86.7 | 73.1 | ResNet50(with FD and IBN) | 94.0 | 84.4 | | 87.7 | 73.9 | ResNet50(with FD and IBN)+RE | 94.3 | 85.0 | | 87.9 | 75.4 | ResNet50(with FD and IBN)+Cut | 95.3 | 86.8 | | 88.5 | 75.9 |
|
Table 4. Experimental results of Cut and RE
Method | Market1501 | | DukeMTMC-reID |
---|
Rank-1 | mAP | | Rank-1 | mAP |
---|
Baseline | 91.2 | 76.9 | | 83.5 | 65.1 | +warmup | 92.7 | 82.6 | | 86.0 | 72.3 | +Cutout | 93.9 | 83.9 | | 86.3 | 73.5 | +LS | 94.5 | 85.2 | | 87.6 | 74.8 | +stride=1 | 95.3 | 86.8 | | 88.5 | 75.9 |
|
Table 5. Experimental results under different training tricks
Method | Market1501 | | DukeMTMC-reID |
---|
Rank-1 | mAP | | Rank-1 | mAP |
---|
IDE | 72.5 | 46.0 | | 67.7 | 47.1 | SVDNet | 82.3 | 62.1 | | 76.7 | 56.8 | HA-CNN | 91.2 | 75.7 | | 80.5 | 63.8 | SVDNet+Era | 87.1 | 71.3 | | 79.3 | 62.4 | IANet | 94.4 | 83.1 | | 87.1 | 73.4 | PCB | 92.4 | 77.3 | | 81.9 | 65.3 | PCB+RPP | 93.8 | 81.6 | | 83.3 | 69.2 | MGN | 95.7 | 86.9 | | 88.7 | 78.4 | Ours | 95.3 | 86.8 | | 88.5 | 75.9 | Ours+Re-ranking | 95.4 | 93.1 | | 90.4 | 87.5 |
|
Table 6. Comparison of experimental results of different methods on Market1501 and DukeMTMC-reID
Method | CUHK03-Labeled | | CUHK03-Detected |
---|
Rank-1 | mAP | | Rank-1 | mAP |
---|
IDE | 22.2 | 21.0 | | 21.3 | 19.7 | SVDNet | | | | 41.5 | 37.3 | HA-CNN | 44.4 | 41.0 | | 41.7 | 38.6 | SVDNet+Era | 49.4 | 45.0 | | 48.7 | 37.2 | PCB | | | | 61.3 | 54.2 | PCB+RPP | | | | 62.8 | 56.7 | MGN | 68.0 | 67.4 | | 66.8 | 66.0 | Ours | 80.9 | 77.8 | | 77.9 | 74.9 | Ours+Re-ranking | 86.5 | 88.1 | | 84.4 | 85.9 |
|
Table 7. Comparison of experimental results of different methods on CUHK03
Method | Rank-1 | mAP |
---|
Part-Aligned | 28.8 | 20.2 | PCB | 42.6 | 33.7 | SFR | 42.3 | 32 | PGFA | 51.4 | 37.3 | HOReID | 55.1 | 43.8 | Ours | 58.7 | 46.4 | Ours+Re-ranking | 61.7 | 61.2 |
|
Table 8. Comparison of experimental results of different methods on Occluded-Duke