Unconstrained Face Verification and Open-World Person
Re-identification via Densely-connected Convolution Neural Network
Donghwuy Ko, Jongmin Yu, Ahmad Muqeem Sheri and Moongu Jeon
Gwangju Institute of Science and Technology 123, Cheomdangwagi-ro, Buk-gu, Gwangju, South Korea
Keywords:
Face Verification, Person Re-Identification, Unconstrained Condition, Metric Learning.
Abstract:
Although various methods based on the hand-crafted features and deep learning methods have been developed
for various applications in the past few years, distinguishing untrained identities in testing phase still remains
a challenging task. To overcome these difficulties, we propose a novel representation learning approach to
unconstrained face verification and open-world person re-identification tasks. Our approach aims to reinforce
the discriminative power of learned features by assigning the weight to each training sample. We demonstrate
the efficiency of the proposed method by testing on datasets which are publicly available. The experimental
results for both face verification and person re-identification tasks show that its performance is comparable to
state-of-the-art methods based on hand-crafted feature and general convolutional neural network.
1 INTRODUCTION
Image recognition has been studied extensively in
computer vision which has resulted in the develop-
ment of various methods based on the hand-crafted fe-
atures as well as deep learning methods. Recently, the
introduction of representation learning methods based
on the deep convolutional neural network (DCNN)
provided automatic feature representation and have
shown outstanding accuracies. Face verification and
person re-identification have also been extensively
studied in computer vision studies. Face verification
is a binary classification task to classify whether two
input images share the same identity, while person re-
identification task is identification over dataset where
images of each identity are in different environment
variables, such as different illumination and angle of
view. To be more specific, person re-identification
is the task of identifying test images from a gallery
where test images and training images are from dif-
ferent cameras. For both tasks, representation lear-
ning methods based on DCNN showed outstanding
accuracies compared to hand-crafted methods. Ho-
wever, in case of person re-identification, it is a still
difficult problem to handle various images distorted
by their extrinsic factors such as illumination and ca-
mera angles. Even though recently proposed methods
based on deep learning can cover these various ima-
ges, these methods only focus on feature which can be
extracted using training samples. In case of face veri-
fication, it is challenging to perform face verification
on unconstrained condition, where identities of test
dataset and training dataset are mutually exclusive.
To overcome these difficulties over person re-
identification and face verification, it is essential
to develop a method that works precisely in both
tasks under unconstrained condition and various en-
vironments. Both face verification and person re-
identification tasks under the unconstrained condition
and various environments are certainly challenging is-
sues. To achieve high accuracy in these conditions, it
is necessary to extract discriminative features. Since
identities in training dataset and test dataset are dif-
ferent, it is obvious that generalized features would
lead to better performance. Especially in person re-
identification, it is important to minimize the effect
of noises caused by its different environmental condi-
tion.
Minimization of noise and classification using ex-
tracted features can be interpreted as reducing intra-
class variance and increasing inter-class variance, re-
spectively. Inter-class variance is variance among
centroid of different identities, whereas intra-class va-
riance is variance among images that share the same
identity. In other words, intra-class variance determi-
nes the size of clusters, and inter-class variance deter-
mines the size of the overlapping area among them.
To make outstanding networks that can extract discri-
minative representation, both inter-class variance and
intra-class variance should be considered.
Ko, D., Yu, J., Sheri, A. and Jeon, M.
Unconstrained Face Verification and Open-World Person Re-identification via Densely-connected Convolution Neural Network.
DOI: 10.5220/0007381104430449
In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 443-449
ISBN: 978-989-758-354-4
Copyright
c
 2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
443

𝜶
l2-normalization
I𝑛𝑝𝑢𝑡 𝑖𝑚𝑎𝑔𝑒 (𝒙)
Multi Convolutional
Layer
Inception
Block
Dense
Block
Densely-connected Convolutional Layer
Softmax
Classifier
N-Classes
Identification
𝑳(𝜶)
128-D hyper-sphere
Threshold
Verification
Figure 1: Overall architecture of proposed network. α represents the output of encoder. In case of verification, latent features
are projected into 128-dimension hypersphere by performing L2-normalization. Other projections which have smaller distance
than given threshold is interpreted as same identity.
The main contribution of this paper is a propo-
sal of a simple and novel way to handle both inter-
class variance and intra-class variance. We applied a
method called loss-weighted gradient where elements
with a higher loss in a mini-batch are weighed more
to overcome difficulties of training hard samples. The
network structure in this paper is inspired by Siamese
Network (Bromley et al., 1994), DenseNet (Huang
et al., 2016), and Inception model (Szegedy et al.,
2015). We employed the densely connected convo-
lutional neural network and inception model. Loss-
weighted gradient is applied to identification loss to
weight the gradient computed from hard samples than
easy samples.
2 PROPOSED APPROACH
The proposed network consists of three parts: fea-
ture extraction, identification, and verification. In fea-
ture extraction, 128 dimension latent features of given
images are extracted. Extracted features are then pro-
cessed by a softmax layer for identification. Latent
features are also projected on 128 dimension hyper-
sphere using L2 normalization and passed into the ve-
rification process. These processes aim to optimize
inter-class variance and intra-class variance.
2.1 Structural Details
The network structure of proposed model is inspired
by DenseNet (Huang et al., 2016) and Inception (Sze-
gedy et al., 2015). Low-level features are extracted
through six conventional convolutional layers and one
max-pooling layer. Eight DenseNet blocks and two
inception blocks are applied subsequently. From pro-
posed network, latent features α are encoded. In the
case of identification, softmax classification is perfor-
med using α as input. In case of verification task, un-
like other methods (Bromley et al., 1994) which uses
latent features itself for verification, we reform latent
features using L2-normalization. L2-normalized fea-
tures described by:
L(α) =k α k
2
=
α
max(k α k, ε)
(1)
where ε is a parameter that prevents the output
value from diverging to ∞. L(α) can be understood
as the projected point of α onto the surface of 128-
dimensional hypersphere. In our method, the distance
between two L2-normalized features is interpreted as
dissimilarity. When the distance between two featu-
res is close, it implies that two image have high simi-
larity. Figure 1 illustrates the overall architecture of
the proposed network.
2.2 Construction of Training Dataset
In many applications based on deep learning (Yin
et al., 2011; Sun et al., 2013; Zhu et al., 2014; Sun
et al., 2014), it is essential to train model with a lot of
training data. In case of Person re-identification, we
merge 4 publicly available datasets: CUHK-02 data-
set (Li and Wang, 2013), CUHK-03 dataset (Li et al.,
2014), CAVIAR4REID dataset (Cheng et al., 2011),
and iLIDs-VID dataset(Wang et al., 2014), and ap-
plied the data augmentation. We applied the simple
linear transformation to artificially enlarge the origi-
nal dataset. Rotation, horizontal flipping, cropping,
and blurring were employed to transform the original
image dataset. Specifically, because the CUHK-02
dataset contains the images of the CUHK-01 dataset
used in evaluation, we manually remove the duplicate
images. Consequently, the training dataset consists of
217,942 images with 1,592 identities. For face verifi-
cation, our model was trained with CASIA-Webface
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
444

dataset (Yi et al., 2014), which contains 10,575 diffe-
rent identities with 494,414 facial images. Data aug-
mentation with both horizontal flip and random crop-
ping is applied.
2.3 Loss Function
2.3.1 Overall Loss Function
To make outstanding networks that can extract discri-
minative representation, both inter-class variance and
intra-class variance should be considered. Network
with large inter-class variance focuses on distinguis-
hable features that are required to verify whether in-
put images are same or not. On the other hand, small
intra-class variance implies that latent features of gi-
ven images are stable regardless of its transformation
such as rotation, blurring, etc. In other words, it gua-
rantees the robustness of extracted features.
Our approach aims to minimize intra-class vari-
ance and maximize inter-class variance. We imple-
mented both identification and verification to modu-
late inter-class variance and intra-class variance. The
total loss function for our network its loss is formula-
ted as follows:
L = L
h
id
+ λL
ve
(2)
where L
h
id
, L
ve
, and λ are identification loss
with loss-weighted gradient, verification loss, and the
hyper-parameter that modulates ratio between iden-
tification loss and verification loss. Identification is
performed with softmax classification. Using extrac-
ted latent features α over softmax classifier, it is for-
mulated as:
L
id
= −
N
∑
i=1
[ ˆo
i
logo
i
+ (1 − ˆo
i
)log(1 − o
i
)]
(3)
where o
i
is the output of softmax classifier, and ˆo
i
is the desired output value. For identification loss of
all training data:
L
h
id
=
1
M
M
∑
j=1
L
h
id, j
(4)
where M is the number of training examples, and
h is hyper-parameter that is used for loss-weighted
gradient. This identification task takes a role of ma-
king the margins among groups of feature that have
the same identity. This role could be interpreted as
optimization of inter-class variance.
To achieve smaller intra-class variance, we em-
ployed the verification task using L2 normalized la-
tent features. In the verification task, we employ the
triplet loss as follows:
L
ve
=
M
∑
i=1
[d(L(a
i
),L(p
i
)) − d(L(a
i
),L(n
i
)) + α]
(5)
where a
i
, p
i
and n
i
are i-th latent feature of anchor,
positive, and negative sample respectively. Verifica-
tion task considers not only samples of target identity
but also that of different identities.
2.3.2 Details of Loss-weighted Gradient
Calculating the gradient of a single batch from equa-
tion (6), it is formulated as:
g(L
id, j
) =
∂L
id, j
∂w
(6)
where j implies index of given batch. For weight
gradient of the element in single batch with respect to
its loss by L
id, j
raised to the power of h :
g((L
id, j
)
h
) =
∂(L
id, j
)
h
∂w
=
∂(L
id, j
)
h
∂L
id, j
∂L
id, j
∂w
= h((L
id, j
)
h−1
)g(L
id, j
)
(7)
Meaning that weight update of individual loss is
multiplied by its loss. A batch with a larger loss will
be weighted more than one with a smaller loss. To
prevent the gradient from getting too small, or having
too large value, we implemented L
h
id, j
that has the fol-
lowing gradient:
g(L
h
id, j
) =
Mh · L
h−1
id, j
∑
M
k=1
L
h−1
id,k
g(L
id, j
)
(8)
Where M is the size of mini batch. Average of
(L
id, j
)
h−1
is divided after calculation of g(L
id, j
). Ap-
plying loss weighted gradient on identification, hard-
sample for each identity will have majority over upda-
ting weights. To show its effect over the network, we
conducted additional experiment on MNIST Dataset.
3 EXPERIMENTAL RESULTS
3.1 Loss-weight Gradient on MNIST
Dataset
We conducted an experiment with classification net-
work with loss-weight gradient using MNIST data-
set (LeCun et al., 1998) to verify the efficiency of the
loss-weighted gradient. In this experiment, we visua-
lize activation results of last hidden layer. The dimen-
sionality of last hidden layer is equal to 2, therefore
value of the output can be interpreted as coordinates
Unconstrained Face Verification and Open-World Person Re-identification via Densely-connected Convolution Neural Network
445

Figure 2: Visualization results on MNIST dataset with re-
spect to h.
in the x-y plane. 10-way softmax classification is per-
formed with cross-entropy loss.
Figure 2 illustrates the visualization result of 2-D
latent features on the last hidden layer. By analyzing
visualization results of softmax with respect to hyper
parameters h, we show that different value of hyper-
parameter h result in the different distribution of sam-
ples’ embedding. The case with h = 1, which is ge-
neral softmax classifier, shows well-spreading shape
of embeddings. However, spreads of individual class
shrinks when value of h becomes large. Especially
in case of h = 5, the inter-class distance becomes
too small that it is barely possible to perform clas-
sification among them. Since unconstrained condi-
tions does not ensure the non-existence of unknown
class between known ones, larger h will result in lo-
wer accuracy. In our case, we conduct experiments to
find appropriate h value to get the best results.
3.2 Face Verification on LFW and YTF
Datasets
For model evaluation, Labeled Faces in the Wild
(LFW) dataset (Huang et al., 2007) and Youtube Fa-
ces (YTF) (Wolf et al., 2011) dataset, known to be
exclusive to CASIA-Webface, are used. We follo-
wed unrestricted with labeled outside data protocol
for evaluation. The face verification performance of
the proposed model is evaluated on 6,000 of face pairs
from LFW dataset, and 5,000 of video pairs from YTF
dataset.
We evaluated the proposed model on LFW and
YTF dataset in unconstrained protocol. Table 1
shows proposed method in comparison with others
for LFW and YTF dataset. Each input facial image
is resized into 160x160 resolution. Value of hyper-
parameter h is assigned to 1.1 which showed the best
performance. Although proposed method did not
achieve the highest accuracy, it showed comparable
performance to models trained with the public data-
Table 1: Accuracy(%) on LFW and YTF dataset. Private
dataset type means that used dataset for training is not pu-
blicly accessible.
Method Dataset Dataset type LFW YTF
DeepFace (Taigman et al., 2014) 4M Private 97.35 91.4
GaussianFace(Lu and Tang, 2015) 850K Private 98.52 N/A
Facenet (Schroff et al., 2015) 200M Private 99.65 95.1
DeepID2+ (Sun et al., 2015) 200K Private 99.47 93.2
Baidu (Liu et al., 2015) 1.3M Private 99.13 N/A
Associate-Predict(Yin et al., 2011) Multi-PIE Public 90.57 N/A
DDML(combined)(Hu et al., 2014) LFW-train Public 90.68 82.34
ConvNet-RBM(Sun et al., 2013) CelebFaces Public 92.52 N/A
high-dim LBP(Chen et al., 2013) WDRef Public 95.17 N/A
TL Joint Bayesian(Cao et al., 2013) WDRef Public 96.33 N/A
FR+FCN(Zhu et al., 2014) CelebFaces Public 96.45 N/A
DeepID(Sun et al., 2014) CelebFaces Public 97.45 N/A
Liu etal (Liu et al., 2016) CASIA Public 99.10 94.0
Ours(h=1.1) CASIA Public 98.83 92.40
set. It showed 98.83% of accuracy over LFW dataset,
and 92.40% of accuracy over YTF dataset.
3.3 Person Re-identification on ViPeR
Dataset and CUHK-01 Dataset
To demonstrate an efficiency of the proposed lear-
ning approach in person re-identification problem, we
initially conduct the experiment using ViPeR dataset
(Gray et al., 2007). The VIPeR dataset contains 632
pedestrian paired images taken from various view-
points with diverse illumination conditions. The data-
set is collected in an academic setting over the course
of several months. We also carry out the experiment
using CUHK-01 dataset (Li et al., 2012). The CUHK-
01 dataset is composed of 971 identities with 4 images
for each. Each image is resized to 128x48 resolution.
In addition, we conducted additional experiments to
get appropriate value of h for loss-weighted gradient.
Table 3 shows the accuracies of the proposed model
with respect to values of h on CUHK-01 dataset.
We use the one-trial Cumulative Matching Cha-
racteristic (CMC) result to compare the proposed met-
hod and others. Table 2 shows the comparison re-
sults using the ViPeR dataset and CUHK-01 dataset
respectively. Due to the lack of experiments about the
unconstrained person re-identification, we have com-
pared the proposed method with approaches which
focus in constrained condition. Despite the disad-
vantage of the unconstrained condition, the proposed
method has achieved the state-of-the-art performance.
The proposed method shows 59.40% and 73.17% ma-
tching rates in 1-rank and 5-rank respectively in Vi-
PeR dataset. In CUHK-01 dataset, proposed met-
hod achieved 61.65% of 1-ranked matching rate, and
the work presented the state-of-the-art performance
among the 1-ranked matching rate across the person
re-identification methods.
The experimental results show that proposed met-
hod outperforms the existing state-of-the-art approa-
ches under the disadvantage of the unconstrained con-
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
446

Table 2: Top-ranked matching rates (%) on ViPeR and CUHK-01 dataset. p-scale denotes the scale of the set of probe
images. The bolded figures represent the best values across their rank. Protocol denotes the environmental condition of the
experiment.
ViPeR Dataset CUHK-01 Dataset
Method Protocol p-scale r = 1 r = 5 r = 10 r = 20 Protocol p-scale r = 1 r = 5 r = 10 r = 20
`
1
-norm (Zhao et al., 2013a) - - - - - - Closed-set 486 10.33 20.64 26.34 33.52
`
2
-norm (Zhao et al., 2013a) - - - - - - Closed-set 486 9.84 19.84 26.42 33.13
SDALF (Farenzena et al., 2010) Closed-set 316 19.87 38.89 49.37 65.73 Closed-set 486 9.90 22.57 30.33 41.03
CPS (Cheng et al., 2011) Closed-set 316 21.84 44.00 57.21 71.00 - - - - - -
eSDC (Zhao et al., 2013b) Closed-set 316 26.31 46.61 58.86 72.77 Closed-set 486 19.67 32.72 40.29 50.58
SalMatch (Zhao et al., 2013a) Closed-set 316 30.16 52.31 65.54 79.15 Closed-set 486 28.45 45.85 55.67 67.95
MLF (Zhao et al., 2014) Closed-set 316 29.11 52.34 65.95 79.87 Closed-set 486 34.30 55.06 64.96 74.94
LADF (Li et al., 2013) Closed-set 316 29.34 61.04 75.98 88.10 - - - - - -
MFA (Xiong et al., 2014) Closed-set 316 32.24 65.99 79.66 90.64 - - - - - -
kLFDA (Xiong et al., 2014) Closed-set 316 32.33 65.78 79.72 90.95 Closed-set 486 32.76 59.01 69.63 79.18
DeepRank (Chen et al., 2016) Closed-set 316 38.37 69.22 81.33 90.43 Closed-set 486 50.41 75.93 84.07 91.32
Ours(h=1.1) Open-set 632 59.40 73.17 78.90 85.55 Open-set 971 61.65 74.95 80.62 85.57
Table 3: Accuracies of the proposed model depending on h
over CUHK-01 dataset.
h value r = 1 r = 5 r = 10 r = 20
1.0 60.06 72.89 78.66 84.02
1.1 61.65 74.95 80.62 85.57
1.2 61.03 74.97 80.00 84.95
1.3 61.03 75.05 80.21 84.95
1.4 61.03 74.02 80.21 84.95
1.5 60.72 74.43 79.48 84.23
dition. The interesting point is that other methods
tend to have dependency over the dataset that its accu-
racy changes when test dataset changes. Our method
seems to be independent of its test dataset. It implies
that our network is well trained to extract more robust
and general features that can be found in any different
dataset.
4 CONCLUSION
This paper proposes a densely-connected convolution
network with loss-weighted gradient. It has shown
great performance over both face verification and per-
son re-identification tasks. The presented approach
was evaluated by comparing face verification and per-
son re-identification task with other methods. Alt-
hough it did not show the best performance in face
verification among models trained with the public da-
taset, it showed comparatively good performance with
98.83% of accuracy over LFW dataset. In the case
of person re-identification, it outperformed previous
state-of-the-art approaches in both CUHK-01 dataset
and ViPeR dataset. Moreover, the presented approach
showed its superior stability over dataset compared to
other methods that the accuracy of other approaches
tends to change over datasets.
ACKNOWLEDGEMENTS
This work was supported by the ICT R&D program
of MSIP/IITP. [2014-0-00077, Development of global
multi-target tracking and event prediction techniques
based on real-time large-scale video analysis]
REFERENCES
Bromley, J., Guyon, I., LeCun, Y., S
¨
ackinger, E., and Shah,
R. (1994). Signature verification using a” siamese”
time delay neural network. In Advances in Neural In-
formation Processing Systems, pages 737–744.
Cao, X., Wipf, D., Wen, F., Duan, G., and Sun, J. (2013). A
practical transfer learning algorithm for face verifica-
tion. In Computer Vision (ICCV), 2013 IEEE Interna-
tional Conference on, pages 3208–3215. IEEE.
Chen, D., Cao, X., Wen, F., and Sun, J. (2013). Blessing of
dimensionality: High-dimensional feature and its effi-
cient compression for face verification. In Computer
Vision and Pattern Recognition (CVPR), 2013 IEEE
Conference on, pages 3025–3032. IEEE.
Chen, S.-Z., Guo, C.-C., and Lai, J.-H. (2016). Deep ran-
king for person re-identification via joint representa-
tion learning. IEEE Transactions on Image Proces-
sing, 25(5):2353–2367.
Cheng, D. S., Cristani, M., Stoppa, M., Bazzani, L., and
Murino, V. (2011). Custom pictorial structures for re-
identification. In Bmvc, volume 2, page 6.
Farenzena, M., Bazzani, L., Perina, A., Murino, V., and
Cristani, M. (2010). Person re-identification by
symmetry-driven accumulation of local features. In
Computer Vision and Pattern Recognition (CVPR),
2010 IEEE Conference on, pages 2360–2367. IEEE.
Gray, D., Brennan, S., and Tao, H. (2007). Evaluating ap-
pearance models for recognition, reacquisition, and
tracking. In Proc. IEEE International Workshop on
Performance Evaluation for Tracking and Surveil-
lance (PETS), volume 3, pages 1–7. Citeseer.
Hu, J., Lu, J., and Tan, Y.-P. (2014). Discriminative deep
metric learning for face verification in the wild. In
Unconstrained Face Verification and Open-World Person Re-identification via Densely-connected Convolution Neural Network
447

Proceedings of the IEEE Conference on Computer Vi-
sion and Pattern Recognition, pages 1875–1882.
Huang, G., Liu, Z., Weinberger, K. Q., and van der Maaten,
L. (2016). Densely connected convolutional networks.
arXiv preprint arXiv:1608.06993.
Huang, G. B., Ramesh, M., Berg, T., and Learned-Miller, E.
(2007). Labeled faces in the wild: A database for stu-
dying face recognition in unconstrained environments.
Technical report, Technical Report 07-49, University
of Massachusetts, Amherst.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998).
Gradient-based learning applied to document recogni-
tion. Proceedings of the IEEE, 86(11):2278–2324.
Li, W. and Wang, X. (2013). Locally aligned feature trans-
forms across views. In Proceedings of the IEEE Con-
ference on Computer Vision and Pattern Recognition,
pages 3594–3601.
Li, W., Zhao, R., and Wang, X. (2012). Human reidentifi-
cation with transferred metric learning. In ACCV.
Li, W., Zhao, R., Xiao, T., and Wang, X. (2014). Deep-
reid: Deep filter pairing neural network for person re-
identification. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pages
152–159.
Li, Z., Chang, S., Liang, F., Huang, T. S., Cao, L., and
Smith, J. R. (2013). Learning locally-adaptive deci-
sion functions for person verification. In Computer
Vision and Pattern Recognition (CVPR), 2013 IEEE
Conference on, pages 3610–3617. IEEE.
Liu, J., Deng, Y., Bai, T., Wei, Z., and Huang, C. (2015).
Targeting ultimate accuracy: Face recognition via
deep embedding. arXiv preprint arXiv:1506.07310.
Liu, W., Wen, Y., Yu, Z., and Yang, M. (2016). Large-
margin softmax loss for convolutional neural net-
works. In ICML, pages 507–516.
Lu, C. and Tang, X. (2015). Surpassing human-level face
verification performance on lfw with gaussianface. In
AAAI, pages 3811–3819.
Schroff, F., Kalenichenko, D., and Philbin, J. (2015). Fa-
cenet: A unified embedding for face recognition and
clustering. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pages 815–
823.
Sun, Y., Wang, X., and Tang, X. (2013). Hybrid deep lear-
ning for face verification. In Computer Vision (ICCV),
2013 IEEE International Conference on, pages 1489–
1496. IEEE.
Sun, Y., Wang, X., and Tang, X. (2014). Deep learning face
representation from predicting 10,000 classes. In Pro-
ceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 1891–1898.
Sun, Y., Wang, X., and Tang, X. (2015). Deeply learned
face representations are sparse, selective, and robust.
In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pages 2892–2900.
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., An-
guelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.,
et al. (2015). Going deeper with convolutions. Cvpr.
Taigman, Y., Yang, M., Ranzato, M., and Wolf, L. (2014).
Deepface: Closing the gap to human-level perfor-
mance in face verification. In Proceedings of the IEEE
conference on computer vision and pattern recogni-
tion, pages 1701–1708.
Wang, T., Gong, S., Zhu, X., and Wang, S. (2014). Person
re-identification by video ranking. In European Con-
ference on Computer Vision, pages 688–703. Springer.
Wolf, L., Hassner, T., and Maoz, I. (2011). Face recognition
in unconstrained videos with matched background si-
milarity. In Computer Vision and Pattern Recognition
(CVPR), 2011 IEEE Conference on, pages 529–534.
IEEE.
Xiong, F., Gou, M., Camps, O., and Sznaier, M. (2014).
Person re-identification using kernel-based metric le-
arning methods. In European conference on computer
vision, pages 1–16. Springer.
Yi, D., Lei, Z., Liao, S., and Li, S. Z. (2014). Lear-
ning face representation from scratch. arXiv preprint
arXiv:1411.7923.
Yin, Q., Tang, X., and Sun, J. (2011). An associate-predict
model for face recognition. In Computer Vision and
Pattern Recognition (CVPR), 2011 IEEE Conference
on, pages 497–504. IEEE.
Zhao, R., Ouyang, W., and Wang, X. (2013a). Person re-
identification by salience matching. In Proceedings of
the IEEE International Conference on Computer Vi-
sion, pages 2528–2535.
Zhao, R., Ouyang, W., and Wang, X. (2013b). Unsupervi-
sed salience learning for person re-identification. In
Proceedings of the IEEE Conference on Computer Vi-
sion and Pattern Recognition, pages 3586–3593.
Zhao, R., Ouyang, W., and Wang, X. (2014). Learning
mid-level filters for person re-identification. In Pro-
ceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 144–151.
Zhu, Z., Luo, P., Wang, X., and Tang, X. (2014). Reco-
ver canonical-view faces in the wild with deep neural
networks. arXiv preprint arXiv:1404.3543.
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
448

APPENDIX
Convolutional Layer DenseBlock Inception
Layers Input Kernel Stride Layer Growth L1 L2 L3 L4 num(Parameters)
Conv1 160x160x3 3x3 2 0.9K
Conv2 79x79x32 3x3 1 9.2K
Conv3 77x77x32 3x3 1 18.4K
MaxPool 77x77x64 3x3 1 0
Conv4 75x75x64 1x1 1 5.1K
Conv5 75x75x80 3x3 1 138.2K
Conv6 73x73x192 3x3 2 442.4K
DenseBlock 36x36x256 6 12 148.4k
Inception a 36x36x164 192 256 256 384 1.4M
DenseBlock 17x17x804 8 24 1.3M
Inception b 17x17x498 256 384 256 256 3.0M
DenseBlock 8x8x1394 6 12 2.6M
DenseBlock 8x8x1466 6 12 2.8M
AvgPool 0
FC 1x1x1538 196.9K
Concat 0
Softmax 128 1.4M
Total 13.5M
Parameter details of proposed network. Size of the input image is 160x160x3. Value of Layer and Growth
in DenseBlock implies the number of sequential convolution layer and depth of output layer respectively. Value
of parameters in Inception a and Inception b implies depth of filters in the first layer, and the second layer in
inception block respectively.
Unconstrained Face Verification and Open-World Person Re-identification via Densely-connected Convolution Neural Network
449