Author Affiliations
Jiangsu Provincial Engineering Laboratory of Pattern Recognition and Computational Intelligence, School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, Chinashow less
Fig. 1. Multi-branch person re-identification framework based on multi-scale attention mechanism
Fig. 2. Structure of multi-scale attention module
Fig. 3. Structure of multi-scale attention-aware feature DropBlock module
Fig. 4. Comparison of two partition strategies
| Method | Market-1501 | DukeMTMC-ReID |
|---|
| Rank-1 | mAP | Rank-1 | mAP |
|---|
| Baseline | 88.63 | 71.55 | 82.27 | 64.57 | | Baseline(with MSA) | 90.15 | 73.03 | 84.47 | 67.02 | | Ours(without MSA) | 94.42 | 86.65 | 89.74 | 79.18 | | Ours | 95.37 | 88.02 | 90.57 | 80.92 |
|
Table 1. Effects of multi-scale attention module unit: %
| Method | Market-1501 | DukeMTMC-ReID |
|---|
| Rank-1 | mAP | Rank-1 | mAP |
|---|
| Baseline | 88.63 | 71.55 | 82.27 | 64.57 | | Baseline(with BDB) | 92.68 | 84.53 | 86.94 | 75.33 | | Baseline(with MSA-FD) | 93.76 | 85.42 | 88.60 | 76.74 | | Ours(without MSA-FD) | 94.94 | 86.72 | 89.59 | 79.61 | | Ours | 95.37 | 88.02 | 90.57 | 80.92 |
|
Table 2. Effects of multi-scale attention-aware feature DropBlock module unit: %
| Partition strategy | Market-1501 |
|---|
| Rank-1 | mAP |
|---|
| UP branch (H=8) | 94.17 | 86.07 | | FP branch (H=9) | 94.43 | 86.29 | | FP branch (H=10) | 94.83 | 86.58 | | FP branch (H=11) | 94.71 | 86.60 | | FP branch (H=12) | 94.50 | 86.42 |
|
Table 3. Effects of different partition strategies unit: %
| Method | Market-1501 | DukeMTMC-ReID |
|---|
| Rank-1 | mAP | Rank-1 | mAP |
|---|
| Baseline | 88.63 | 71.55 | 82.27 | 64.57 | | Baseline+MSA-FD branch | 93.76 | 85.42 | 88.60 | 76.74 | | Baseline+FP branch | 94.83 | 86.58 | 89.50 | 79.47 | | All | 95.37 | 88.02 | 90.57 | 80.92 |
|
Table 4. Effect of joint multiple local feature strategy unit: %
| Partition strategy | Market-1501 |
|---|
| Rank-1 | mAP |
|---|
| MSCAN[23] | 80.31 | 57.53 | | MGCAM[22] | 83.55 | 74.25 | | HA-CNN[11] | 91.20 | 75.70 | | AACN[9] | 85.90 | 66.87 | | SPReID[25] | 90.80 | 76.56 | | MLFN[24] | 90.00 | 74.30 | | PCB[5] | 93.80 | 81.60 | | MGN[6] | 95.70 | 86.90 | | Ours | 95.37 | 88.02 | | Ours+Re-ranking | 96.35 | 94.50 |
|
Table 5. Comparison of results on Market-1501 unit: %
| Method | DukeMTMC-ReID |
|---|
| Rank-1 | mAP |
|---|
| JLML[27] | 73.30 | 56.40 | | SVDNet-ResNet50[28] | 76.70 | 56.80 | | AACN[9] | 76.84 | 59.25 | | SPReID[25] | 80.48 | 63.27 | | HACNN[11] | 80.50 | 63.80 | | MLFN[24] | 81.00 | 62.80 | | PCB[5] | 83.30 | 69.20 | | MGN[6] | 88.70 | 78.40 | | Ours | 90.57 | 80.92 | | Ours+Re-ranking | 93.00 | 90.74 |
|
Table 6. Comparison of results on DukeMTMC-ReID unit: %
| Method | CUHK03(Labeled) | CUHK03(Detected) |
|---|
| Rank-1 | mAP | Rank-1 | mAP |
|---|
| DPFL[29] | 43.00 | 40.50 | 40.70 | 37.00 | | MGCAM[22] | 49.29 | 49.89 | 46.29 | 46.74 | | HA-CNN[11] | 44.40 | 41.00 | 41.70 | 38.60 | | MLFN[24] | 54.10 | 49.20 | 52.80 | 47.80 | | SVDNet-ResNet50[28] | 40.90 | 37.80 | 41.50 | 37.30 | | PCB[5] | - | - | 63.70 | 57.50 | | MGN[6] | 68.00 | 67.40 | 66.80 | 66.00 | | Ours | 77.43 | 75.84 | 75.57 | 73.28 | | Ours+Re-ranking | 85.30 | 87.15 | 83.64 | 85.17 |
|
Table 7. Comparison of results on CUHK03 unit: %
| Method | Computationtime perbatch /s | Rank-1 /% | mAP /% |
|---|
| Mobilenet_v2[30] | 0.138 | 87.0 | 68.5 | | HA-CNN[11] | 0.237 | 91.2 | 75.7 | | MLFN[24] | 0.585 | 90.0 | 74.3 | | PCB[5] | 0.331 | 93.8 | 81.6 | | MGN[6] | 0.561 | 95.7 | 86.9 | | Ours | 0.544 | 95.37 | 88.02 |
|
Table 8. Comparison of computation speed on Market-1501
