In Defense of the Triplet Loss for Person Re-Identification

22 Mar 2017  ·  Alexander Hermans, Lucas Beyer, Bastian Leibe ·

In the past few years, the field of computer vision has gone through a revolution fueled mainly by the advent of large datasets and the adoption of deep convolutional neural networks for end-to-end learning. The person re-identification subfield is no exception to this. Unfortunately, a prevailing belief in the community seems to be that the triplet loss is inferior to using surrogate losses (classification, verification) followed by a separate metric learning step. We show that, for models trained from scratch as well as pretrained ones, using a variant of the triplet loss to perform end-to-end deep metric learning outperforms most other published methods by a large margin.

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Benchmark
Person Re-Identification CUHK03 TriNet Rank-1 89.63 # 5
Rank-5 99.01 # 3
Person Re-Identification DukeMTMC-reID TriNet Rank-1 72.44 # 66
mAP 53.50 # 68
Person Re-Identification Market-1501 LuNet Rank-1 81.38 # 86
Rank-5 92.34 # 13
mAP 60.71 # 92
Person Re-Identification Market-1501 TriNet Rank-1 84.92 # 80
Rank-5 94.21 # 11
mAP 69.14 # 82
Person Re-Identification Market-1501 LuNet (RK) Rank-1 84.59 # 81
Rank-5 91.89 # 14
mAP 75.62 # 75
Person Re-Identification Market-1501 TriNet (RK) Rank-1 86.67 # 76
Rank-5 93.38 # 12
mAP 81.07 # 70
Person Re-Identification MARS TriNet (RK) mAP 77.43 # 14
Rank-1 81.21 # 11
Rank-5 90.76 # 6
Person Re-Identification MARS LuNet (RK) mAP 73.68 # 15
Rank-1 78.48 # 13
Rank-5 88.74 # 8
Person Re-Identification MARS LuNet mAP 60.48 # 19
Rank-1 75.56 # 15
Rank-5 89.70 # 7
Person Re-Identification MARS TriNet mAP 67.70 # 18
Rank-1 79.80 # 12
Rank-5 91.36 # 5

Methods


No methods listed for this paper. Add relevant methods here