Domain Adaptation for Person Re-identification on New Unlabeled Data
Tiago de C. G. Pereira
a
and Teofilo E. de Campos
b
Departamento de Ci
ˆ
encia da Computac¸
˜
ao, Universidade de Bras
´
ılia - UnB, Bras
´
ılia-DF, Brazil
Keywords:
Domain Adaptation, Person Re-identification, Deep Learning.
Abstract:
In the world where big data reigns and there is plenty of hardware prepared to gather a huge amount of non
structured data, data acquisition is no longer a problem. Surveillance cameras are ubiquitous and they capture
huge numbers of people walking across different scenes. However, extracting value from this data is chal-
lenging, specially for tasks that involve human images, such as face recognition and person re-identification.
Annotation of this kind of data is a challenging and expensive task. In this work we propose a domain adapta-
tion workflow to allow CNNs that were trained from one domain to be applied to another domain without the
need for new annotation of the target data. Our results show that domain adaptation techniques really improve
the performance of the CNN when applied in the target domain.
1 INTRODUCTION
The purpose of person re-identification is to match
images of persons in non-overlapping cameras views.
It can be helpful in some important applications as
intelligent video surveillance (Wang, 2013), action
recognition (Wei Niu et al., 2004) and person retrieval
(Sun et al., 2017).
For problems related to identifying people in im-
ages, the first method of choice is usually based on
face recognition. This is because such algorithms
have already matched the human capacity, as we can
see in Taigman et al.’s work (Taigman et al., 2014),
where a 97.35% accuracy was achieved in the LFW
dataset (Huang et al., 2008) while the human accu-
racy on the same data is 97.53%. However, face
recognition algorithms have little value on surveil-
lance images because the subjects are usually far
away from the cameras, so there is not enough resolu-
tion in the area of the face. Furthermore, the surveil-
lance viewpoint is usually such that a high amount of
(self-)occlusion happens, to the point that the faces
are not visible at all. For these reasons, person re-
identification algorithms usually take the whole body
into account. The typical workflow to train a person
re-identification system follows this steps:
1. Use a CCTV system to gather non structured data;
2. Filter this data using a person detector and tracker;
a
https://orcid.org/0000-0002-9200-9795
b
https://orcid.org/0000-0001-6172-0229
3. Annotate person bounding boxes;
4. Train a metric learning CNN in the annotated
data;
5. Deploy the trained CNN to match people that ap-
pear in different cameras.
The biggest problem with this workflow is step 3,
because CNNs need a huge amount of data to be prop-
erly trained and the process of annotating all this data
is very expensive (in terms of time and manpower).
We therefore propose to replace this step by an unsu-
pervised domain adaptation technique. According to
Pan and Yang (Pan and Yang, 2010), domain adapta-
tion is a type of transfer learning where only source
domain data is labeled and both domains have the
same task.
In our technique, we use a public dataset as our
source domain and the non structured data from the
CCTV as our target domain. In our source domain all
the annotation and image filtering have already been
done, then we use unsupervised image-image transla-
tion to create an intermediate dataset. This dataset has
the labels of the source domain, but the appearance of
people is similar to those in the target domain. Next,
we proceed to the metric learning step using that inter-
mediate dataset. As the intermediate dataset is similar
to the target domain, we expect that the CNN trained
in it will perform well in the target domain.
In addition, we use this learned metric to annotate
the target domain using a clustering algorithm. That
way, we have pseudo labels available for the target
Pereira, T. and E. de Campos, T.
Domain Adaptation for Person Re-identification on New Unlabeled Data.
DOI: 10.5220/0008973606950703
In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 4: VISAPP, pages
695-703
ISBN: 978-989-758-402-2; ISSN: 2184-4321
Copyright
c
 2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
695

domain, then we fine tune our CNN in these pseudo
labels and learn specific characteristics from the tar-
get domain. As the training is performed with the ac-
tual target domain images we expect to increase even
more our performance, even though the pseudo label
generate a noisy label space for the target domain.
In our experiments, we evaluated the CNN perfor-
mance in the target domain (direct transfer), we eval-
uated the same CNN trained with the dataset adapted
by a cycleGAN to the target distribution and we also
evaluated the same CNN trained in the target domain
using our pseudo label method. Our method surpasses
the baseline accuracy in all test cases. All of that is
achieved by replacing step 3 by our technique and will
be explained in more details in sections 3 and 4.
In addition, we observed that the highly unbal-
anced nature of the person re-identification problem
means that training batches may be heavily biased to-
wards negative samples. To deal with that, we use a
batch scheduler algorithm that allows to train a CNN
with a triplet loss in cases where the data is noisy.
Next section discusses related work. Section 3
presents our method and Section 4 presents experi-
ments and results. This paper concludes in Section 5.
2 RELATED WORK
The state-of-art on person re-identification follows
a pattern of using either attention-based neural net-
works (Liu et al., 2017), factorization neural networks
(Chang et al., 2018) or body parts detection (Zhao
et al., 2017). The common point in these works is try-
ing to disregard the background information, so they
can give the proper weight for the image areas where
the person is visible. These methods achieve great re-
sults, but have a high complexity, as they are based
on combinations of several elements. However, dif-
ferent datasets have different characteristics and cer-
tain combination of methods may not work across all
datasets. In this paper, our focus is on the exploitation
of domain adaptation for this application. To design
more controlled experiments, we use a relatively sim-
ple end-to-end system based on the ResNet-50 (He
et al., 2016) as a backbone.
Typically, the person re-identification challenge
is approached as a metric learning task (Zhao et al.,
2017; Deng et al., 2018). But it can also be ap-
proached as a classification task where each person
from the dataset is a class (Liu et al., 2017; Chang
et al., 2018). The problem of the classification-based
approach is that the space of labels is fixed and has
a large cardinality. Such methods are rarely applica-
ble in practice, unless the set of identities of people
who transit through a set of environments is always
the same. Our target application is public spaces,
therefore it is not possible to restrict the set of labels.
Therefore we approach this as a metric learning chal-
lenge. Further to being applicable to public spaces,
the task of comparing samples is the same across dif-
ferent domains. This enables the application of un-
supervised domain adaptation methods to adapt the
marginal distribution of the data.
Recently, some works presented domain adapta-
tions techniques for person re-identification. (Zhao
et al., 2017) created a new dataset to evaluate the gen-
eralization capacity of his model. Their CNN was
evaluated in it without further training. (Zhong et al.,
2018) used a cycleGAN to approximate the camera
views in a dataset trying to learn a camera latent space
metric. (Xiao et al., 2016) trained his CNN with a
super dataset created concatenating multiple datasets.
They proposed a domain guided dropout to further
specialize their CNN for each dataset. In this work,
we consider that the target domains have no labeled
data, then we cannot use the approaches of (Zhong
et al., 2018) or (Xiao et al., 2016). The approach of
(Zhao et al., 2017) can be called direct transfer, be-
cause it just evaluates a CNN on a target domain. We
shall demonstrate that our method outperforms direct
transfer.
3 PROPOSED METHOD
Our technique is based on training a CNN to learn a
metric, so we can ensure that distinct domains will
have the same task. Therefore, we train a ResNet-50
(Section 3.1) with the triplet loss (Section 3.2) to learn
the desired metric in an Euclidean vector space. The
core of the domain adaptation method is based in a
cycleGAN that will perform an image-image transla-
tion to approximate source and target domains (Sec-
tion 3.3). Then, we use the CNN trained in the in-
termediate dataset to extract the features of the tar-
get domain images and use a clustering algorithm to
generate pseudo-labels for the target domain (Section
3.4).
3.1 Baseline CNN
As said in Section 2, the state-of-art in person re-
identification use techniques that exploit information
from CNNs at multiple levels, bringing multiple se-
mantic levels to the final features. Those semantic
levels may carry specific person attributes like gen-
der, clothing, textures and clothing, which are impor-
tant for matching people across views.
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
696

We choose ResNet-50 architecture in our work be-
cause we believe that residuals blocks help to propa-
gate information from multiple semantic levels when
they are relevant for the output. Although the residual
blocks may not perform as well as a specific architec-
ture, the main goal of our work is to propose a domain
adaptation workflow.
To have an initial boost (Donahue et al., 2014),
we start with a ResNet-50 CNN pre-trained on Ima-
geNet (Deng et al., 2009). We then transfer learn it to
the problem of person re-identification using a public
dataset. This is done by replacing the last fully con-
nect layer by a new fully connected layer with 128
features which are used as an embedding for metric
learning. We use Adam optimizer and the triplet loss.
3.2 Triplet Loss and Batching Strategies
A siamese-like loss is ideal when trying to learn a
metric because it allows one to perform an end-to-end
learning from a dataset to an embedding space. The
siamese loss receives as input a pair of feature vec-
tors and tries to approximate them if they are from the
same person or set them apart if they are from differ-
ent people. This generates an embedding space where
feature vectors from the same person tend to lie near
each other.
The triplet loss is an upgrade from the siamese
loss which instead of using a pair of samples as in-
put, it uses an anchor, a positive sample and a nega-
tive sample. Therefore, the triplet loss approximates
feature vectors from the same person while it also sep-
arates features of different people, according to equa-
tion 1 (defined for each anchor sample x
a
). This way,
one can expect better samples separation in the em-
bedding space.
L(x
a
) = max

0 , m + D

f
a
, f
p

− D

f
a
, f
n


, (1)
where m is a margin so the loss does not go to zero, f
is the CNN output, i.e., a lower dimensional embed-
ding of image x; (sub indexes a, p and n mean anchor,
positive and negative, respectively) and D(·) can be
any distance measurement algorithm, in our case is
the Euclidean distance defined by
D(u, v) =
v
u
u
t
d
∑
i=1
(u
i
− v
i
)
2
. (2)
A question that arises from the triplet loss use
is “how to choose the positive/negative examples?”
(Hermans et al., 2017) investigated this problem and
came to a conclusion that the best learning is achieved
when using the hardest positive/negative samples dur-
ing training. This approach was coined batch hard
and it works as follows: for each anchor sample x
a
from the batch, the choice of positive sample x
p
is
chosen as the one that maximizes D(f
a
, f
p
) and the
negative sample x
n
is chosen as the one that mini-
mizes D(f
a
, f
n
). Using this strategy, equation 1 can
be rewritten as
L
BH
(x
a
) = max

0 , m + max
p
D

f
a
, f
p

(3)
− min
n
D

f
a
, f
n


,
where positive and negative samples are chosen
within each batch and the losses across all anchors
in a batch are averaged out.
Figure 1 illustrates how samples are chosen for a
batch. All the rectangles at the top represent sam-
ples from a person and the rectangles at the bottom
represent sample of another person. The triplet will
choose each rectangle as anchor at a time, calculate
the loss for it and in the final sum all the losses. From
the green rectangle as an anchor, the numbered ar-
rows indicate the distance D(·) from it to the samples,
where Pos i, i = 1, 2,3, are possible positive samples
and Neg j, j = 1, 2,3, 4, are the possible negative sam-
ples. In a batch hard approach, Pos 2 is selected
as positive sample, Neg 3 as negative sample and
L
BH
= m + 0.361 − 0.490.
Figure 1: Example of a batch hard triplet selection.
(Hermans et al., 2017) proved the batch hard ef-
fectiveness, but choosing the hardest samples at each
batch increases the training complexity. Furthermore,
we work with an intermediate dataset that can be
noisy, meaning that the separation between positive
and negative samples may be less trivial, which in-
creases the training cost even more. The consequence
is that the training process may never converge with
this strategy. When using the triplet loss, a non
Domain Adaptation for Person Re-identification on New Unlabeled Data
697

converging training process can be identified if the
loss is stuck at the margin (m), because that means
D(f
a
, f
p
) = D(f
a
, f
n
), meaning that all the features are
converging to vectors of 0s.
While training with the triplet loss, the goal is to
make D(f
a
, f
p
) < D(f
a
, f
n
). However, if the batch is
big, the number of negative examples is way bigger
than the number of positive examples, particularly in
the case of person re-identification. It is therefore
possible to have a negative sample that is nearer to
the anchor than the hardest positive sample. This
way the loss will always be greater than the margin
(L
BH
> m), then the optimizer learns that outputting
vectors of 0s will reduce the loss to the margin, i.e.,
(L
BH
= m).
Our solution was to use a batch scheduler algo-
rithm to decrease the number of negative samples and
lower the training complexity. This way we ease the
training convergence, and once the training is con-
verging we slowly increase the batch size (and there-
fore its complexity, having an impact in the loss). This
enables us to learn step by step and converge the train-
ing even with a noisy dataset. Our batch scheduler
algorithm is shown in Algorithm 1.
Algorithm 1: Batch Scheduler.
batch size = 8
m = 0.5 // m is the loss margin of Eq. 1
for i = 0 to num epochs do
loss = train(i, batch size)
if loss < 0.8 × m then
batch size = batch size +8
end if
end for
3.3 Image-image Translation for
Domain Adaptation
To give some background, the definitions and nota-
tions used in this paper are based on (Csurka, 2017)
and (Pan and Yang, 2010). A domain D is com-
posed of a d dimensional feature space X ⊂ IR
d
with
a marginal probability distribution P(X) and a task
T defined by a label space Y and the conditional
probability distribution P(Y|X), where X and Y are
sets of random variables (which usually are multivari-
ate). Given a particular sample set X = {x
1
, ··· , x
n
} ∈
X , with corresponding labels Y = {y
1
, ··· , y
n
} ∈ Y ,
P(Y|X) in general can be learned in a supervised
manner from these feature-label pairs {x
i
, y
i
}.
For simplicity, let us assume that there are two
domains: a source domain D
s
= {X
s
, P(X
s
)} with
T
s
= {Y
s
, P(Y
s
|X
s
)} and a target domain D
t
=
{X
t
, P(X
t
)} with T
t
= {Y
t
, P(Y
t
|X
t
)}. Those do-
mains are different D
s
6= D
t
, because P(X
s
) 6= P(X
t
)
due to domain shift. Also, we do not have the target
domain labels Y
t
, so we do not have the feature-label
pairs {x
i
, y
i
} to learn P(Y|X
t
) in a supervised manner.
The person re-identification task T consists in
learning a projection from x ∈ X to a feature f in
a Euclidean space where f is closer to other vectors
if they originated from the same person, more dis-
tant to vectors from other people. The set of labels
can be thought of as the space of all possible person
identities in the world, which impractical. Alterna-
tively, the person re-ID problem can be seen as a bi-
nary problem that takes two samples as input, indi-
cating whether or not they come from the same per-
son. Therefore, each person re-ID dataset (or indeed
each camera surveillance environment) can be seen
as a different domain, however the task is always the
same, i.e., telling if two images contain the same per-
son or not. Domain adaptation are transductive trans-
fer learning methods where it is assumed T
s
= T
t
,
according to Csurka (Csurka, 2017). Therefore, we
can use domain adaptation to exploit the related in-
formation from {D
s
, T
s
} to learn P(Y
t
|X
t
).
In our method, we have images from source do-
main X
s
and target domain X
t
, but we do not have
the labels from target domain Y
t
. So, we approxi-
mate data from images of a known source domain to
images of a target domain generating an intermediate
dataset.
We use, as source domain, a public dataset which
has ground truth annotation of positive/negative ex-
amples for each anchor. An unsupervised domain
adaptation method can be used to generate an inter-
mediate dataset D
i
that leverages the source domain
annotation Y
s
and is similar to the target domain. For
that, we follow an approach based on Generative Ad-
versarial Networks – GANs (Goodfellow et al., 2014).
More specifically, we use the cycleGAN method pro-
posed by (Zhu et al., 2017) and applied to person re-
identification by (Deng et al., 2018).
The idea is to use images from the source domain
(X
s
) as input and train a GAN to generate outputs
which are similar to the images from the target do-
main (X
t
). However, once we have no paired images
between domains the problem has a high complexity.
Zhu et al. proposed to train two generators G and F
where G : X
s
→ X
t
is a mapping from the source do-
main to the target and F : X
t
→ X
s
is a mapping from
the target domain to the source. Also, a cyclic com-
ponent is added to the loss:
L(G, F, D
X
s
, D
X
t
) = L
GAN
(G, D
X
t
, X
s
, X
t
)+
L
GAN
(F, D
X
s
, X
t
, X
s
)+ (4)
λL
cyc
(G, F),
where both L
GAN
components are the basic GAN loss
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
698

proposed by Goodfellow et al. and the L
cyc
is the
cyclic component added by Zhu et al., wich is given
by:
L
cyc
(G, F) = E
X
s
∼p
data
(X
s
)

k
F(G(X
s
)) − X
s
k
1

+
E
X
t
∼p
data
(X
t
)



G(F(X
t
)) − X
t


1

(5)
the cyclic component is there to do an identity match
between source domain images X
s
and their dou-
ble transformed pairing images F(G(X
s
)), and vice-
versa. By minimizing this cyclic loss we expect to
have transformations that can map both domains.
Therefore, we use the generator G : X
s
→ X
t
in all
images of our source domain to generate an interme-
diate dataset. That is, we create a dataset that lever-
ages from the labeled data of the source domain and
have similar characteristics to the target domain. This
way we can expect that a training in this intermediate
dataset will perform well in the target domain.
3.4 Pseudo Labels for Re-identification
In Section 3.2, we used the triplet loss to learn a dis-
tance metric in an Euclidean vector space. In Section
3.3, we showed that both source and target domains
have the same label space Y . We also presented a
method to train our CNN in an intermediate dataset
that leverages from the labeled data of the source do-
main and have similar characteristics to the target do-
main. The CNN therefore should already present a
reasonable performance in target domain.
We use the CNN to extract all features f
t
i
from tar-
get domain images X
t
and these features belong to an
Euclidean vector space. Then, we used a clustering
algorithm to group these features, using the obtained
group identifications as target domain with pseudo-
labels Y
t
. In addition, we fine tune the CNN using the
feature-label pairs {x
i
, y
i
} with the real images from
target domain and the pseudo-labels generated by the
clustering algorithm.
Even though the pseudo labels generated may
contain a lot of errors, this next training step uses the
real images from target domain X
t
. Therefore, the
CNN is be able to learn more robust features for the
target domain, because it learns the exact characteris-
tics of the target domain.
We choose the k-means (Hartigan and Wong,
1979) clustering algorithm to group the features in the
Euclidean vector space. The value of k was chosen as
a proportion of the size of each target dataset. Table
1 indicates the values used in this paper (the datasets
are discussed later). However, the naive assignment
of samples to clusters is a flawed strategy to anno-
tate the data, because a simple look at the data may
cluster viewpoints rather than people. In other words,
features from different people taken from the same
camera view are often more similar to each other than
features from the same person from different camera
views.
Table 1: The chosen k for each dataset when using k-means
algorithm.
Dataset k
CUHK03 2000
Market1501 1600
Viper 632
Our solution is to use k-means algorithm to gener-
ate k clusters for each camera view, then use a nearest
neighbor algorithm to group these clusters across the
camera views. This way, we guarantee that every per-
son from our pseudo-labels space have images from
each camera. That results in a noisy annotation, be-
cause that assumption is not a true in the real label
space of the dataset. However, using this approach
we ease the CNN task of learning features robust for
multiple cameras views and achieve better results in
validation.
4 EXPERIMENTAL RESULTS
In our work, we produced results using three well
known person re-identification datasets, they are the
CUHK03 (Li et al., 2014), the Market1501 (Zheng
et al., 2015) and the Viper (Gray et al., 2007). For all
the experiments, we did not use any label information
in the target domain, except to evaluate the results.
Our work produces two kind of results that must
be analyzed to understand the method effectiveness.
These results are the generation of a intermediate
dataset (discussed in section 4.1) and the CNN eval-
uation in the target domain after the complete work-
flow was done using pseudo-labels (discussed in sec-
tion 4.2).
4.1 Intermediate Dataset
As said in section 3.3 our method tries to approximate
the source domain to the target domain. This is done
training a cycleGAN between both domains and us-
ing the generator to create an intermediate dataset that
shifts the source domain samples so that they become
more similar to the target domain data. The idea is to
generate images that preserve the person morphology,
but are visually adapted to the target domain. While
there is no guarantee that a GAN preserves person
morphology, the cyclic loss contributes towards this
goal, as it has an identity match component.
Domain Adaptation for Person Re-identification on New Unlabeled Data
699

Figure 2 presents examples of transformation re-
sults between all domains. It is interesting to note that
the person morphology have been well preserved and
the changes have been more in the colors, texture and
background. That means we could produce a great
approximation of how a person would look like in the
view of another dataset.
The CUHK03 dataset was created using surveil-
lance cameras from a university in Hong Kong with
an elevated viewpoint, so normally the background of
their images consists in a granular floor. While the
Market1501 dataset was created with cameras in a
park, so the images usually have grass in the back-
ground of their views. Viper is the oldest dataset used
in this work, it was published in 2007 and is com-
posed of low resolution outdoor images.
These characteristics of the datasets make it easy
to understand the effects seen in Figure 2. When us-
ing CUHK03 as the target domain, the transformed
images tend to have a granular background to approx-
imate the floor texture in CUHK03 images. When
using Market1501 as target domain, images from
CUHK03 had a background transformation from the
granular floor to grass, and images from Viper had
just a color transformation, because both datasets are
from outdoor images. When using Viper as target do-
main, images from Market1501 had a color transfor-
mation and images from CUHK03 had a texture back-
ground transformation and a brightness enhancement.
4.2 Domain Adaptation Results
4.2.1 Image-image Translation Method
After successfully generating an intermediate dataset
that approximates both domains we used that inter-
mediate dataset to fine-tune the CNN trained in the
source domain. We evaluated all the results in the tar-
get domain using the CMC score with rank-1, rank-5
and rank-10.
The cycleGAN method was compared with the di-
rect transfer method, where the direct transfer method
consists in evaluating in the target domain a CNN
trained in the source domain without further training.
The direct transfer method therefore shows how dif-
ferent are both domains and is used as a baseline.
As one can see in Table 2 the cycleGAN method
presents huge rank-1 improvements when using
CUHK03 as target domain (26% improvement for
Viper as source domain and 14.9% improvement
for Market1501 as source domain). This happens
because the CUHK03 images have granular back-
ground texture as a strong characteristic that was eas-
ily learned by our cycleGAN.
A great rank-1 improvement was also obtained for
Market1501 as target domain and CUHK03 as source
domain, where the cycleGAN method achieved a
9% improvement compared with the baseline. Fur-
thermore, for Market1501 as target and Viper as
source domain our method achieved 1% improve-
ment, meaning that the color transformation helped to
approximate these domains, but this was not as signif-
icant as texture changes that occurred when working
with CUHK03 images.
For Viper as a target domain the cycleGAN
method achieved 1.5% rank-1 improvement using
CUHK03 as source domain and 1.9% rank-5 im-
provement for Market1501 as source domain. Again,
this means that texture transformations are more sig-
nificant than color transformations. Although those
are not our best results, they are very significant be-
cause as Viper is an old dataset it has a lot less images
than the others (only 1264 images), so learning to cre-
ate the intermediate dataset in a unsupervised manner
without much data is extremely hard.
4.2.2 Pseudo-labels Method
Section 4.2.1 proved the effectiveness of domain
adaptation and that the cycleGAN successfully
shifted images to the target domain appearance, car-
rying their source label with them. Also, it was clear
that texture transformations are more significant than
color transformations.
Although the cycleGAN did a great job shifting
images between domains, when using the pseudo-
labels method we achieved even better results. This is
because the training is now performed with the actual
target domain images and estimated pseudo-labels.
So, there is no longer the problem of images in which
the person morphology was not preserved. The target
dataset characteristics are better represented. Figure 3
illustrates the dataset created using pseudo-labels – as
one can see the estimated labels are not perfect, but
the grouped images show a strong color similarity.
As one can see in Table 2, our method showed
great improvements in all test cases. Even when us-
ing the Viper dataset as target domain our method
could improve the cycleGAN results in 2% or more.
For the Market1501 dataset the rank-1 improvement
was around 2% also and for the CUHK03 our method
achieved improvements of 4% in rank-1 accuracy.
It is important to notice that the pseudo-labels
have a stronger positive impact on smaller target
datasets. This is because small datasets require fewer
clusters to annotate the data. This was very significant
for the great results presented for Viper dataset.
In summary our method is significantly better than
direct transfer without adaptation. It is important to
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
700

Figure 2: Examples of the cycleGAN transformations between domains.
Figure 3: Images from a final cluster when using the pseudo-labels method. The cluster were achieved using Viper as source
dataset and Market1501 as target dataset.
emphasize that our method does not make use of any
label from the target domain, completely removing
the burden of annotating new data when the applica-
tion domain changes.
5 CONCLUSIONS
In person re-identification, each type of environment
(e.g. airport, shopping center, university campus, etc.)
has its own typical appearance, so a system that is
trained in one environment does not perform very
well in another environment. This observation was
confirmed by our cross-dataset (direct transfer) exper-
iments, indicating that each dataset can be treated as
a domain. Therefore, we showed that a domain adap-
tation method based on cycleGAN can be applied to
transform the marginal distribution of samples from
a source dataset to a target dataset. This enables us
to retrain a triplet CNN on adapted samples so that
Domain Adaptation for Person Re-identification on New Unlabeled Data
701

Table 2: Results with all domains combinations.
Accuracy
Target Domain Source Domain Method Rank-1 Rank-5 Rank-10
Market1501
Viper
Direct Transfer 5.7% 15.5% 22.2%
CycleGAN 6.7% 17.0% 23.7%
Ours 8.6% 20.5% 28.4%
CUHK03
Direct Transfer 26.8% 45.9% 55.1%
CycleGAN 35.8% 56.5% 65.7%
Ours 37.3% 60.4% 70.4%
CUHK03
Viper
Direct Transfer 5.9% 18.1% 29.0%
CycleGAN 31.9% 64.4% 77.5%
Ours 36.1% 69.2% 81.3%
Market1501
Direct Transfer 19.9% 49.4% 63.2%
CycleGAN 34.8% 66.7% 79.1%
Ours 38.2% 69.7% 81.6%
Viper
CUHK03
Direct Transfer 10.1% 22.5% 29.0%
CycleGAN 11.6% 25.5% 34.7%
Ours 13.6% 33.9% 46.0%
Market1501
Direct Transfer 12.5% 25.0% 33.1%
CycleGAN 9.8% 26.9% 36.4%
Ours 13.9% 29.0% 40.7%
their performance is improved on the target dataset
without using a single labeled sample from the target
set. Furthermore, we showed that using this CNN and
a clustering algorithm to generate pseudo-labels and
retrain the triplet CNN leads to a significant boost in
the performance on target dataset. This opens doors
for the deployment of person re-ID software to real
applications, as it completely removes the burden of
annotating new data.
Further to proposing a domain adaptation tech-
nique for this problem, we also presented the use of
a batch scheduler which increases the batch size as
training starts to converge.
For future works, we believe it would be inter-
esting to try our technique with other datasets, using
more robust CNN architectures as backbone and with
different clustering algorithms. But it is proved that
this technique brings great contribution to the field of
person re-identification.
ACKNOWLEDGEMENTS
The authors would like to thank FAPDF (http://www.
fap.df.gov.br) and CNPq grant PQ 314154/2018-3
(url http://cnpq.br/).
REFERENCES
Chang, X., Hospedales, T. M., and Xiang, T. (2018). Multi-
level factorisation net for person re-identification. In
Proc of the IEEE Conf on Computer Vision and Pat-
tern Recognition (CVPR).
Csurka, G. (2017). A comprehensive survey on domain
adaptation for visual applications. In Csurka, G., ed-
itor, Domain Adaptation in Computer Vision Applica-
tions, pages 1–35. Springer International Publishing,
Cham.
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-
Fei, L. (2009). ImageNet: A Large-Scale Hierarchical
Image Database. In CVPR09.
Deng, W., Zheng, L., Ye, Q., Kang, G., Yang, Y., and
Jiao, J. (2018). Image-image domain adaptation with
preserved self-similarity and domain-dissimilarity for
person re-identification. In The IEEE Conference on
Computer Vision and Pattern Recognition (CVPR).
Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N.,
Tzeng, E., and Darrell, T. (2014). Decaf: A deep con-
volutional activation feature for generic visual recog-
nition. In Xing, E. P. and Jebara, T., editors, Proceed-
ings of the 31st International Conference on Machine
Learning, Bejing, China. PMLR.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,
Warde-Farley, D., Ozair, S., Courville, A., and Ben-
gio, Y. (2014). Generative adversarial nets. In Ghahra-
mani, Z., Welling, M., Cortes, C., Lawrence, N. D.,
and Weinberger, K. Q., editors, Advances in Neu-
ral Information Processing Systems (NIPS) 27, pages
2672–2680. Curran Associates, Inc.
Gray, D., Brennan, S., and Tao, H. (2007). Evaluating ap-
pearance models for recognition, reacquisition, and
tracking. In In IEEE International Workshop on Per-
formance Evaluation for Tracking and Surveillance,
Rio de Janeiro.
Hartigan, J. A. and Wong, M. A. (1979). Algorithm AS
136: A K-means clustering algorithm. Journal of the
Royal Statistical Society. Series C (Applied Statistics).
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
702

ual learning for image recognition. In The IEEE Con-
ference on Computer Vision and Pattern Recognition
(CVPR).
Hermans, A., Beyer, L., and Leibe, B. (2017). In defense of
the triplet loss for person re-identification. Technical
Report arXiv:1703.07737, Cornell University Library.
http://arxiv.org/abs/1703.07737.
Huang, G. B., Mattar, M., Berg, T., and Learned-Miller,
E. (2008). Labeled Faces in the Wild: A Database
for Studying Face Recognition in Unconstrained En-
vironments. In Workshop on Faces in ‘Real-Life’ Im-
ages: Detection, Alignment, and Recognition, Mar-
seille, France. Erik Learned-Miller and Andras Fer-
encz and Fr
´
ed
´
eric Jurie.
Li, W., Zhao, R., Xiao, T., and Wang, X. (2014). Deep-
reid: Deep filter pairing neural network for person re-
identification. In Proc of the IEEE Conf on Computer
Vision and Pattern Recognition (CVPR), pages 152–
159.
Liu, X., Zhao, H., Tian, M., Sheng, L., Shao, J., Yi, S., Yan,
J., and Wang, X. (2017). Hydraplus-net: Attentive
deep features for pedestrian analysis. In The IEEE
International Conference on Computer Vision (ICCV).
Pan, S. J. and Yang, Q. (2010). A survey on transfer learn-
ing. IEEE Transactions on Knowledge and Data En-
gineering, 22(10):1345–1359.
Sun, Y., Zheng, L., Deng, W., and Wang, S. (2017). Svdnet
for pedestrian retrieval. In IEEE International Confer-
ence on Computer Vision (ICCV), pages 3820–3828.
Taigman, Y., Yang, M., Ranzato, M., and Wolf, L. (2014).
Deepface: Closing the gap to human-level perfor-
mance in face verification. In The IEEE Conference
on Computer Vision and Pattern Recognition (CVPR).
Wang, X. (2013). Intelligent multi-camera video surveil-
lance: A review. Pattern Recognition Letters, 34(1):3
– 19. Extracting Semantics from Multi-Spectrum
Video.
Wei Niu, Jiao Long, Dan Han, and Yuan-Fang Wang
(2004). Human activity detection and recognition
for video surveillance. In IEEE International Con-
ference on Multimedia and Expo (ICME) (IEEE Cat.
No.04TH8763), volume 1, pages 719–722 Vol.1.
Xiao, T., Li, H., Ouyang, W., and Wang, X. (2016). Learn-
ing deep feature representations with domain guided
dropout for person re-identification. In The IEEE Con-
ference on Computer Vision and Pattern Recognition
(CVPR).
Zhao, H., Tian, M., Sun, S., Shao, J., Yan, J., Yi, S., Wang,
X., and Tang, X. (2017). Spindle net: Person re-
identification with human body region guided feature
decomposition and fusion. In Proc 30th IEEE Conf
on Computer Vision and Pattern Recognition (CVPR),
Honolulu, Hawaii, July 21-26.
Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., and Tian,
Q. (2015). Scalable person re-identification: A bench-
mark. In IEEE International Conference on Computer
Vision.
Zhong, Z., Zheng, L., Zheng, Z., Li, S., and Yang,
Y. (2018). Camera style adaptation for person re-
identification. In Proc of the IEEE Conf on Computer
Vision and Pattern Recognition (CVPR).
Zhu, J.-Y., Park, T., Isola, P., and Efros, A. A. (2017).
Unpaired image-to-image translation using cycle-
consistent adversarial networks. In The IEEE Inter-
national Conference on Computer Vision (ICCV).
Domain Adaptation for Person Re-identification on New Unlabeled Data
703