Codes for Deep Attention Aware Feature Learning for Person Re-Identification
We have tested the performance of our method on two typical networks: TriNet and Bag of Tricks. The results are as follows (mAP (rank-1)).
Method | Market-1501 | CUHK-03(labeled) | CUHK-03(Detected) | Duke-MTMC |
---|---|---|---|---|
TriNet | 65.48 ( 82.51 ) | 46.39 ( 51.07 ) | 46.74 ( 51.93 ) | 53.50 ( 72.44 ) |
DAAF-TriNet | 70.40 ( 85.87 ) | 53.21 ( 58.23 ) | 51.98 ( 57.03 ) | 58.91 ( 76.63 ) |
TriNet* | 69.14 ( 84.92 ) | 54.45 ( 55.86 ) | 51.98 ( 52.64 ) | 58.18 ( 75.36 ) |
DAAF-TriNet* | 72.63 ( 87.17 ) | 55.01 ( 60.34 ) | 54.48 ( 58.71 ) | 60.12 ( 77.29 ) |
Bag of Tricks | 85.90 ( 94.50 ) | 60.90 ( 63.30 ) | 58.00 ( 59.10 ) | 76.40 ( 86.40 ) |
DAAF-BoT | 87.90 ( 95.10 ) | 67.60 ( 69.00 ) | 63.10 ( 64.90 ) | 77.90 ( 87.90 ) |
Bag of Tricks** | 94.20 ( 95.40 ) | 77.10 ( 74.40 ) | 73.40 ( 70.40 ) | 89.10 ( 90.30 ) |
DAAF-BoT** | 95.00 ( 96.40 ) | 82.00 ( 78.70 ) | 77.30 ( 73.90 ) | 89.60 ( 91.70 ) |
* represents five crops and flip
** represents re-ranking