In Defense of the Triplet Loss for Person Re-Identification
Alexander Hermans
∗
, Lucas Beyer
∗
and Bastian Leibe
Visual Computing Institute
RWTH Aachen University
Abstract
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, thanks to the
notable publication of the Market-1501 and MARS datasets
and several strong deep learning approaches. Unfortu-
nately, a prevailing belief in the community seems to be that
the triplet loss is inferior to using surrogate losses (classi-
fication, verification) followed by a separate metric learn-
ing 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 any
other published method by a large margin.
1. Introduction
In recent years, person re-identification (ReID) has at-
tracted significant attention in the computer vision commu-
nity. Especially with the rise of deep learning, many new
approaches have been proposed to achieve this task [18, 1,
32, 6, 34, 24, 31, 8, 41, 3, 37, 42]
1
. In many aspects per-
son ReID is similar to image retrieval, where significant
progress has been made and where deep learning has re-
cently introduced a lot of changes. One prominent example
in the recent literature is FaceNet [23], a convolutional neu-
ral network (CNN) used to learn an embedding for faces.
The key component of FaceNet is to use the triplet loss, as
introduced by Weinberger and Saul [33], for training the
CNN as an embedding function. The triplet loss optimizes
the embedding space such that data points with the same
identity are closer to each other than those with different
identities. A visualization of such an embedding is shown
in Fig. 1.
Several approaches to person ReID have already used
some variant of the triplet loss to train their models [15,
7, 22, 6, 32, 24, 26, 20, 5]. However, the recently most suc-
∗
Equal contribution. Ordering determined by a last minute coin flip.
1
A nice overview of the field is given by a recent survey paper [40].
Figure 1: A small crop of the Barnes-Hut t-SNE [30] of
our learned embeddings for the Market-1501 test-set. The
triplet loss learns semantically meaningful features.
cessful person ReID approaches argue that a classification
loss, possibly combined with a verification loss, is superior
for the task [5, 40, 8, 41]. Typically, these approaches train
a deep CNN using one or multiple of these surrogate losses
and subsequently use a part of the network as a feature ex-
tractor, combining it with a metric learning approach to gen-
erate final embeddings. Both of these losses have their prob-
lems, though. The classification loss necessitates a growing
number of learnable parameters as the number of identities
increases, most of which will be discarded after training. On
the other hand, many of the networks trained with a verifi-
cation loss have to be used in a cross-image representation
mode, only answering the question “How similar are these
two images?”. This makes using them for any other task,
1
arXiv:1703.07737v2 [cs.CV] 27 Mar 2017
