Novelty Detection for Person Re-identiﬁcation in an Open World

George Galanakis

1,2

, Xenophon Zabulis

and Antonis A. Argyros

1,2

Computer Science Department, University of Crete, Greece

Institute of Computer Science, FORTH, Greece

Keywords:

Person Re-identiﬁcation, Open World, Novelty Detection.

Abstract:

A fundamental assumption in most contemporary person re-identiﬁcation research, is that all query persons

that need to be re-identiﬁed belong to a closed gallery of known persons, i.e., they have been observed and a

representation of their appearance is available. For several real-world applications, this closed-world assump-

tion does not hold, as image queries may contain people that the re-identiﬁcation system has never observed

before. In this work, we remove this constraining assumption. To do so, we introduce a novelty detection

mechanism that decides whether a person in a query image exists in the gallery. The re-identiﬁcation of per-

sons existing in the gallery is easily achieved based on the persons representation employed by the novelty

detection mechanism. The proposed method operates on a hybrid person descriptor that consists of both super-

vised (learnt) and unsupervised (hand-crafted) components. A series of experiments on public, state of the art

datasets and in comparison with state of the art methods shows that the proposed approach is very accurate in

identifying persons that have not been observed before and that this has a positive impact on re-identiﬁcation

accuracy.

1 INTRODUCTION

A key problem in vision-based person tracking is en-

countered when a person exits and then re-enters the

ﬁeld(s) of view of the cameras observing it. On that

occasion, we wish to re-identify this person and as-

sociate it with its previous detection(s). Such a re-

identiﬁcation can support person tracking in large en-

vironments that are covered by multiple cameras, or

improve single-view tracking against occlusions and

viewing limitations.

Typically, re-identiﬁcation methods assume that a

gallery set contains person representations associated

with person IDs. These representations may comprise

of images as well as of global or local feature descrip-

tors. A query or probe image of a person is compared

against the gallery set, seeking a match with any of the

persons therein. In most cases, the gallery is assumed

to be a closed-set, i.e., it contains a representation for

every person whose identity is going to be queried. In

contrast, in open-world person re-identiﬁcation, the

gallery is assumed to be an open-set. As it is possi-

ble that this person has never been observed before,

the query image may not match any of the persons in

the gallery. In this context, prior to re-identiﬁcation,

it needs to be decided whether a person has been ob-

served before or not. From a technical point of view,

the open-world version of the problem is more chal-

lenging. At the same time, its solution can support a

much more diverse set of application domains.

In this work, we treat the problem of open-world

person re-identiﬁcation. We do so by casting the

problem of deciding whether a person belongs or

not to the gallery as a novelty detection problem.

A recent review on person re-identiﬁcation (Zheng

et al., 2016a) indicates novelty detection as an open

issue. Novelty detection and the subsequent person

re-identiﬁcation operate on a newly proposed person

descriptor that consists of a supervised (learnt) and an

unsupervised (hand crafted) component.

Two major advantages of the proposed approach

over existing ones are the following. First, it can be

applied to a new setting (e.g., environment, set of ca-

meras, gallery of known persons) without any additio-

nal training, as required by other learning approaches

(Zheng et al., 2016b; Zhu et al., 2017). Second, it can

be naturally adapted to operate in an online fashion.

That is in contrast to methods as in (Wang et al.,

2016) which require that the entire probe set is avai-

lable, rather than gradually introduced. Therefore, we

regard that our method constitutes a very attractive

candidate for solving the person re-identiﬁcation pro-

Galanakis, G., Zabulis, X. and Argyros, A.

Novelty Detection for Person Re-identiﬁcation in an Open World.

DOI: 10.5220/0007368304010411

In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 401-411

ISBN: 978-989-758-354-4

401

blem in open, real-world setups and scenarios. Such

a scenario is considered in the CONNEXIONs Hori-

zon 2020 project, which is funded by the European

Commission, and will develop and demonstrate next-

generation detection, prediction, prevention, and in-

vestigation services. In this context, the tracking and

re-identiﬁcation of persons in multiple heterogeneous

cameras is of paramount importance.

2 RELATED WORK

Research topics that are relevant to person re-

identiﬁcation methods include the investigation of

descriptors that can be used to represent persons

as well as methods for comparing them. Re-

identiﬁcation methods themselves are categorized ba-

sed on whether they operate under the assumption of

an open or closed world. We focus in the second ca-

tegory, as this is the one treated in this paper by incor-

porating novelty detection mechanisms.

Features and Descriptors for Person Representa-

tion. Data-driven features have recently gained at-

tention due to the proliferation of Convolutional Neu-

ral Networks (CNNs) and the availability of large-

scale datasets for training. These factors gave rise

to learned features, which are robust to illumina-

tion, scale and pose variations (Hermans et al., 2017;

Chen et al., 2017; Zhou et al., 2017; Su et al., 2016;

Wang et al., 2017; Qian et al., 2017; Li et al., 2018a;

Song et al., 2018; Xu et al., 2018; Sarfraz et al.,

2018), but also speciﬁc to the training data. Hand-

crafted representations have also witnessed advan-

ces. Departing from conventional features such as

color histograms, SIFT and HoG features, more so-

phisticated descriptors have been proposed (Faren-

zena et al., 2010; Liao et al., 2015; Matsukawa et al.,

2016a; Gou et al., 2017). Context-speciﬁc know-

ledge has been also utilized, i.e., body-part segmen-

tation (Su et al., 2017; Qian et al., 2017; Zheng et al.,

2017a; Zhao et al., 2017) and attribute recognition

(Shi et al., 2015; Su et al., 2016; Matsukawa and Su-

zuki, 2016; Qian et al., 2017; Chang et al., 2018). In

greater relation to this work, learnt and hand-crafted

features have been combined to support person re-

identiﬁcation (Wu et al., 2016).

Comparing Person Descriptors. Conventional mea-

sures such as Euclidean distance and cosine similarity

have been adopted for this purpose (Farenzena et al.,

2010; Wang et al., 2017). Learnt measures have been

also used with better results (Zheng et al., 2016a). In

such methods, discrepancies due to viewpoint and il-

lumination variation between views are learned for

speciﬁc view combinations (Yang et al., 2014; Liao

et al., 2015; Martinel et al., 2015; Chen et al., 2015;

Jose and Fleuret, 2016; Yu et al., 2017). CNNs have

been also employed in this task (Hermans et al., 2017;

Chen et al., 2017; Zhou et al., 2017; Wang et al.,

2018; Xu et al., 2018). Pertinent methods are dataset-

dependent, as they learn the change of person appea-

rance across speciﬁc views.

Closed-world Person Re-identiﬁcation. Closed-set

re-identiﬁcation methods mainly focus on a proper

deﬁnition of the above components, i.e., discrimina-

tive features and/or distance metrics. Subsequently,

most works perform pairwise comparison of query

images against gallery ones. Pairwise comparisons

result to a ranked list, on top of which lays the most

probable gallery person (nearest neighbor). To im-

prove results, some works exploit information about

top-ranked persons to perform re-ranking (Zheng

et al., 2015b; Lisanti et al., 2015; Zhong et al., 2017;

Sarfraz et al., 2018). Other works, especially those

that operate on videos, maintain images of a person

in a set. Thus, comparison among query and gallery

persons is formulated as set comparison (Wang et al.,

2014) or graph matching (Ye et al., 2017).

During the evaluation of learning-based re-

identiﬁcation methods, datasets are split into training

and testing subsets to avoid learning bias. Persons in

these subsets are referred as non-target and target, re-

spectively. Learning based on non-target persons is

not practical in real-world applications, because it re-

quires manual annotation of a large number of ima-

ges.

Open-world Person Re-identiﬁcation. Only a few

works deal with open-world person re-identiﬁcation.

The decision on whether a person belongs to the gal-

lery set has been approached as a novelty detection

problem. Novelty detection was initially applied to

object classiﬁcation. In (Bodesheim et al., 2013; Liu

et al., 2017) a discriminative null space is recommen-

ded, where images from the same, known, class are

mapped to a single point in the null space. The most

recent method (Liu et al., 2017) addresses the incre-

mental case, where the gallery set is progressively

expanded with new object representations. The no-

velty detection methods in (Bendale and Boult, 2016;

unther et al., 2017) employ normalization in class

inclusion scores by applying Extreme Value Theory

(EVT). In (Kliger and Fleishman, 2018), novel ob-

jects are detected using the Generative Adversarial

Network (GAN) framework. Deep neural networks

have been utilized by three recent works (Ruff et al.,

2018; Masana et al., 2018; Perera and Patel, 2018)

in an effort towards end-to-end feature extraction and

novelty detection. The works in (Ruff et al., 2018;

Perera and Patel, 2018) target one-class classiﬁcation,

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

402

thus, in contrast to (Bodesheim et al., 2013), they do

not expand naturally to multi-class novelty detection,

but depend on some aggregation similar to One-Class

SVMs (see Sec. 4). We also note that (Kliger and

Fleishman, 2018; Ruff et al., 2018; Masana et al.,

2018) are evaluated on datasets which are simpler

and/or contain fewer classes (MNIST, CIFAR-10 etc).

Moreover, if only few per class samples are available

for training, application of DNNs is limited. Thereaf-

ter, we consider previous works as options for multi-

class novelty detection.

A ﬁrst example on how novelty detection is ap-

plied to open-world person re-identiﬁcation appears

in (Brun et al., 2011), where the kernel PCA algo-

rithm is applied to graph representations of persons.

An input representation is considered as novel if its

squared distance from the ﬁrst few of the principal

components is above a certain threshold. This met-

hod is subject to careful parameter selection (number

of principal components, kernel size), which is avoi-

ded in null space methods (Bodesheim et al., 2013).

Three recent works (Wang et al., 2016; Zheng et al.,

2016b; Zhu et al., 2017) deal with the open-world pro-

blem in very challenging settings. (Wang et al., 2016;

Zheng et al., 2016b) use only one training image per

person, while (Zhu et al., 2017) achieves computa-

tional efﬁciency in addition to accuracy. Although

(Wang et al., 2016) depends only on target persons

(unsupervised setting), its objective function requires

the entire query set as input and can be applied only

to a pair of views. This is particularly constraining

as, in real-world settings, the entire query set is not

available at once and more than two views are em-

ployed to achieve full coverage of the environment.

The work in (Zheng et al., 2016b) capitalizes on a

non-target dataset from all views, in order to project

person representations to a new feature space. Intuiti-

vely, in this space, intra-identity representations are

similar and inter-identity representations dissimilar.

The work in (Zhu et al., 2017) proposes to learn two

hashing functions, one for gallery and another one for

probe images (obtained from disjoint views). Then,

it compares hash codes using the Hamming distance.

Both methods (Zheng et al., 2016b; Zhu et al., 2017)

depend on a single dataset, containing images from

speciﬁc viewpoints, captured at arbitrary illumination

conditions and do not generalize to new conditions.

In the most recent work, (Li et al., 2018b), the

GAN framework is utilized for learning how to discri-

minate between known and unknown persons. Their

network architecture comprises of two types of dis-

criminators; one between persons and other visual

content, and the other between known and unknown

persons. They consider a subset of the ids (persons)

as known, while the rest of the ids are divided into

training and testing. We argue that this method, alt-

hough promising, is impractical in real-world scena-

rios. The reason is the following. Learning to discri-

minate between known versus unknown persons con-

sists of utilization of both (a) known ids and (b) a set

of few hundred unknown ids, during training. The la-

ter set essentially represents the negative samples, i.e.

unknown persons. The method in (Li et al., 2018b)

is demonstrated on training and testing datasets from

same period of time and camera network ((Li et al.,

2018b), Sec. 4.2). Therefore it depends on time-

demanding acquisition and manual labeling of person

images, apart from those of known ids. This assump-

tion is fairly limiting. Moreover, the occurrence of an

unknown id requires overly time-consuming training

of the neural network that prohibits the real-time ope-

ration of the system. This time-consuming training is

due to the incremental augmentation of the gallery of

known persons.

Our Approach. In this work, we propose OW-

REID, a novel method for open-world person re-

identiﬁcation. OW-REID operates on a person des-

criptor that consists of both hand-crafted and learnt

features. It capitalizes on the Kernel Null Folley-

Summon Transform (KNFST) (Bodesheim et al.,

2013), a parameter-free novelty detection technique,

to decide whether a person belongs to the gallery set

or not. Novelty detection based on KNFST is pos-

sible without requiring an annotated dataset of non-

target persons. This increases the exploitability of

the method in real world conditions as it reduces dra-

matically its setup time and costs. The KNFST al-

gorithm has been applied to re-identiﬁcation (Zhang

et al., 2016) but for addressing the closed-world ver-

sion of the problem. In this context it is used as a

feature mapping/reprojection technique and not as a

novelty detection mechanism. More speciﬁcally, it is

used to learn the transfer function normalizing person

appearance across a number of views so that known

people can be re-identiﬁed if observed by these ca-

meras. In contrast, OW-REID learns a person-speciﬁc

representation from a number of cameras, so that (a)

these persons can be re-identiﬁed by these or other ca-

meras and (b) unknown persons are identiﬁed as such.

Our Contribution. In summary, the contributions

of this work are the following: (1) We propose OW-

REID, a novel method for open-world person re-

identiﬁcation by treating it as a novelty detection pro-

blem; (2) We propose a hybrid person descriptor that

consists of supervised (hand-crafted) and unsupervi-

sed (learnt) features. This contributes to better no-

velty detection accuracy and independence from the

training data; (3) We provide an extensive, compa-

Novelty Detection for Person Re-identiﬁcation in an Open World

403

rative study and evaluation of the proposed approach

against baseline methods and person descriptors on

standard datasets.

3 OPEN-WORLD

RE-IDENTIFICATION

(OW-REID)

The proposed approach for novelty detection and

open-world person re-identiﬁcation (OW-REID) capi-

talizes on KNFST (Bodesheim et al., 2013), a met-

hod for detecting novel entities among general known

object categories. In our framework, each category

corresponds to a single person and contains different

views of that person. KNFST is applied to person des-

criptors that represent a view of each person as a mul-

tidimensional feature vector.

3.1 Person Description

As person descriptors, we have considered both learnt

and hand-crafted features that have been designed to

optimize the accuracy of novelty detection and person

re-identiﬁcation.

As a ﬁrst candidate, we considered the features

proposed in (Hermans et al., 2017) where each image

of a person is represented as an 128-D feature vector,

called TriNet. TriNets are learned through training on

two recent large-scale datasets, Market1501 (Zheng

et al., 2015a) and MARS (zhe, 2016), achieving state-

of-the-art re-identiﬁcation accuracy in those particu-

lar datasets.

As a second candidate, we considered the recently

proposed, hand-crafted GOG features (Matsukawa

et al., 2016b). In (Matsukawa et al., 2016b), two

variants for GOG are presented. The ﬁrst, GOG

rgb

is obtained by applying the GOG descriptor on an

RGB color representation of an image. In the second,

GOG

fusion

, the GOG descriptor is applied indepen-

dently to the RGB, Lab, HSV and nRnG color en-

codings and the results are concatenated. In prelimi-

nary experiments, the GOG

rgb

descriptor yielded bet-

ter novelty detection accuracy and was adopted there-

after. In line with (Matsukawa et al., 2016b), we ap-

ply mean removal and L2 normalization. Also, using

default settings, the GOG

rgb

descriptor has 7,567

dimensions. To retain only expressive components

and improve accuracy, we retained 1000 components

through PCA dimensionality reduction. Henceforth,

we refer to the resultant descriptor simply as GOG.

In addition, we introduce GOG+TriNet, a new

descriptor formed by the concatenation of the GOG

and TriNet descriptors, in an effort to retain the ad-

vantages of both worlds.

3.2 Novelty Detection

Given a gallery set G and a query image q, we wish to

deduce whether q depicts a person of known identity

(i.e., whether q ∈ G). Let c be the number of known

persons. Query and gallery images are assumed to en-

velope persons, showing as less background as pos-

sible. Below we describe how KNFST (Bodesheim

et al., 2013) is applied to solve this novelty detection

problem.

Each image x (either query or gallery ones) are

mapped to a descriptor/feature space F as discussed

in Section 3.1. After this mapping has been perfor-

med, the original images are not utilized anymore. Let

i ∈ [1, c] enumerate persons in G. Let also v

be the

number of images of person i, and j ∈ [1, v

] enume-

rate the images for person i. Then, x

i j

denotes the

person descriptor of the j

image of the i

person.

Let X ∈ R

n×d

be a matrix containing a descriptor

i j

in each of its rows (n =

∑

i=1

is the number of

gallery images and d the dimensionality of x

i j

). The

kernel K ∈ R

n×n

is constructed to contain pairwise si-

milarities for these n vectors. As in (Liu et al., 2017),

pairwise similarities are measured utilizing the Radial

Basis Function (RBF) kernel.

Using KNFST, a projection P(x) : F → N is lear-

ned, where N denotes a c −1 null space. Each person

i is represented by a point t

∈ N , i ∈ [1,c], deﬁned as

= mean

i j

). The input parameters of KNFST are

K and corresponding person labels. Details on how

the projection is computed can be found in (Bodes-

heim et al., 2013). To provide some intuition, Fig. 1

provides a 2D visualization of how KNFST operates.

Figure 1 left, shows 100 images of the RAiD data-

set (10 images for each of 10 randomly chosen per-

sons). Originally, these are 1128-D GOG+TriNet vec-

tors in F that are projected on the 2D plane using t-

SNE (Maaten and Hinton, 2008) to aid visualization.

Images of the same person correspond to the same

marker and color. We can see that images of the same

person project to different points on the plane. Actu-

ally, for each person there are two clusters of points,

each corresponding to images acquired by a different

camera. Figure 1 right, shows a 2D projection of the

representation of the same images in the N space. It

can be veriﬁed that in this space, images of the same

person are much more tightly clustered.

Novelty detection is then performed as follows.

The feature vector z in F corresponding to a query

image q is projected to point t in N . A novelty score

is computed, as the distance of t to its nearest neig-

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

404

Figure 1: Visualization of feature vectors for 10 images per person and for 10 persons in the original feature space F (left)

and in the KNFST’s projection space N (right). Each person identity is represented by a unique marker and color. Both F

and N spaces are projected to the 2D plane for visualization purposes (see text for details).

hbor t

in N (ideally, the different persons in the gal-

lery G). If this distance is larger than a threshold θ,

the query image is considered to be novel, i.e., depict

a person that is not a member of the gallery. At the

same time, if this distance is lower than θ, the iden-

tity of the nearest neighbor t

of t in N solves the

re-identiﬁcation problem.

As soon as a novel person is detected, its views

can be used to augment the gallery G and re-learn

the KNFST-based mapping from F to N . Thus,

the proposed method can be turned into an on-

line/incremental approach that extends an original

gallery G of known persons, as views of previously

unseen persons are encountered..

4 EXPERIMENTAL EVALUATION

To evaluate novelty detection, we adopt the recently

proposed cross-view scheme (Zhu et al., 2017). In

this scheme, G contains images from a subset of

views (cameras). The query set Q contains images

of both known and unknown persons from the remai-

ning views. Considering the remaining views is par-

ticularly important because in a real-world setting, a

person is typically ﬁrst observed in a view and needs

to be re-identiﬁed in another view, which usually is

subject to different illumination conditions.

Evaluation datasets comprise of multiple images

per person, per view. Persons in G and Q are imaged

in a wide range of viewpoints and lighting conditi-

ons. Let L

and L

denote the disjoint sets of views

in G and Q , respectively; L = L

∪ L

denotes the

set of all views. Let Q

and Q

denote the subsets

of images of known and unknown persons, respecti-

vely; thus, Q = Q

∪Q

. In the experiments, we rand-

omly choose the views in L

, to avoid bias to speciﬁc

views. Similarly, to avoid bias over speciﬁc persons,

we randomly chose the identities in Q

and Q

. Fi-

nally, to avoid bias over a speciﬁc random selection of

views or persons, we perform 300 random trials and

report average accuracy, in each case. In all perfor-

med experiments, G is not incrementally updated, as

it would be in a real-world system. This is to measure

the accuracy of novelty detection over multiple trials

and under the same conditions, regarding varying gal-

lery sizes c and numbers of gallery images per person,

. Also, in line with (Bodesheim et al., 2013), relati-

vely small values of c are considered (< 50).

Datasets. To evaluate OW-REID, three recent data-

sets have been employed, Market1501 (Zheng et al.,

2015a), DukeMTMC-reID (Zheng et al., 2017b) and

RAiD (Das et al., 2014). All datasets contain multi-

ple persons and multiple views per person. For the

ﬁrst two datasets, images are split into training and

testing sets. For these datasets, only the testing set

is utilized in the evaluation, as the training set is re-

served for learning methods (see Sec. 2). For RAiD,

no such splitting is provided and the entire dataset is

utilized.

The testing subsets of Market-1501 and

DukeMTMC-reID datasets contain 750 and 1110

persons, observed from up to 6 and 8 cameras, re-

spectively. For DukeMTMC-reID, we are interested

in persons that are imaged in at least two views (702

in total), as required by the cross-view protocol.

RAiD contains images for 43 walking persons from

4 views (2 indoor and 2 outdoor), with an exception

of 2 persons that are imaged by only 3 views.

Novelty Detection Evaluation Metrics. Novelty de-

tection accuracy is measured in (Bodesheim et al.,

2013) using the area under the ROC curve, hence-

forth auROC. The ROC curve plots the true positive

rate (TPR) against the false positive rate (FPR), for

different values of θ. In our case, TPR denotes the ra-

tio of persons correctly classiﬁed as unknown, while

FPR denotes the ratio of known persons misclassi-

ﬁed as unknown. The works in (G

unther et al., 2017;

Zheng et al., 2016b; Zhu et al., 2017) use the DIR

Novelty Detection for Person Re-identiﬁcation in an Open World

405

curve (Jain and Li, 2011), which correspondingly to

the ROC curve, plots the Detection Identiﬁcation Rate

(DIR) over the False Alarm Rate (FAR), as a function

of threshold θ. This curve has been also referred to

as “True Target Rate (TTR) over False Target Rate

(FTR)” in (Zheng et al., 2016b; Zhu et al., 2017).

This measure is similar to ROC but, besides novelty

detection, assesses identiﬁcation accuracy as well. In

this work, we borrow the idea from (Bodesheim et al.,

2013) and use the area under the DIR curve, hence-

forth auDIR, which is determined based on FAR and

DIR, where FAR is the ratio of unknown persons mis-

classiﬁed as known and DIR is the ratio of persons

correctly classiﬁed as known and, also, correctly re-

identiﬁed.

Experiment Types. We investigated two parameters

that affect the performance of novelty detection: (a)

the number of images v

for each known person i, and

(b) the number of known persons, c. A large value of

, implies that persons are imaged in a wide range of

viewpoints and illumination conditions. In this case,

novelty detection is provided with diverse gallery data

for each person, which we anticipate as beneﬁcial for

novelty detection accuracy. In contrast, a large value

of c implies greater ambiguity in novelty detection, as

there is a greater chance for accidental similarity of

an unknown person to a person in G.

Given the above, novelty detection accuracy was

assessed in a wide variety of settings, in two experi-

ments, the VPV (varying person views) and the VGS

(varying gallery size) experiments. More speciﬁcally,

in the VPV experiment, the number of known persons

was constant (c = 20) and the number of available

images per person, v

, was modulated as follows. We

required v

to be in range [v

] and systematically

varied the limits of this range. The speciﬁc value of v

depends on how many images are available for person

i in the particular trial. If there are not enough gallery

images (v

< v

), person i is considered as unknown.

If more than enough images are available (v

> v

we randomly choose a subset of v

images from the

available ones. As mentioned earlier, the datasets do

not have the same number of images per person. For

this reason, we evaluate these datasets in suitable ran-

ges (see Section 4). In the VGS experiment, values

= 20, v

= 50 were kept constant and c varied in

the range {1,5,10,. . .,50}, for all datasets.

Given the dataset sizes, in all experiments, we

chose |L

| = 5 for the Market1501 and DukeMTMC-

reID and |L

| = 2 for RAiD. In all experiments, the

TriNet features used were as originally trained on the

Market1501 dataset.

Comparison with Baselines. We compare the per-

formance of OW-REID against two baseline met-

hods: (a) distance thresholding in the original feature

space and (b) Multiple One-Class SVMs. In (a), we

compute pairwise distances of a query z ∈ F against

all x

i j

, as measured by the Euclidean (L2) distance.

Query z is considered to belong to an unknown per-

son, if the maximum of pairwise distances is larger

than a threshold θ. For (b), we used the νSVM vari-

ant (Sch

olkopf et al., 2000), with ν = 0.5 and γ = 1/d,

where d is the dimensionality of F . We utilized c

One-Class SVM models, one for each person. Each

model m

is learned by utilizing all the v

feature vec-

tors x

i j

for person i. Then, we compute the pairwise

scores of z against every m

, select the best score and

compare it to threshold θ, as above. We note that we

use the inverse distance as novelty score.

Tables 1 and 2 summarize mean auROC, of the

compared methods for the VPV experiment (varia-

ble v

) and VGS experiment (variable c), respectively.

Tables 3 and 4 summarize auDIR quantities, for the

same experiments. TriNet features dominate perfor-

mance in the Market1501 dataset, independently of

the novelty detection method. That is attributed to the

fact that TriNet features were deﬁned through training

on exactly this dataset. For the other two datasets,

we observe signiﬁcant performance drop when Tri-

Net features are utilized, independently of the novelty

detection method. We conclude that learning featu-

res in a certain dataset and context, proves inadequate

for supporting accurate novelty detection in other da-

tasets, obtained under different conditions. We also

stress that re-training features in a new setting is a

cumbersome task. First, because it requires labori-

ous manual annotation for the gallery images and, se-

cond, because in an open-world setting, examples of

unseen persons are not available. In contrast, the pro-

posed combination of TriNet with hand-crafted featu-

res yields the best performance, when an appropriate

novelty detection method is utilized.

Indeed, the selection of the novelty detection met-

hod plays a signiﬁcant role in the obtained perfor-

mance. When hand-crafted (GOG) features are uti-

lized, better results are obtained using OW-REID.

Even more importantly, OW-REID combined with

GOG+TriNet achieves a better overall performance

against the other compared methods. OW-REID per-

forms worse only for very small values of c, v

. In

these cases, the limited number of samples make it

difﬁcult for KNFST to learn an appropriate projection

space N . Nonetheless, OW-REID performs well

when v

≥ 10, an assumption that can easily be met

in practice. OW-REID is computationally more ef-

ﬁcient than the L2 baseline, because it compares a

query with a single point in N , instead of comparing

it to all images in the gallery.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

406

Table 1: Mean auROC for the VPV experiment (varying person views, v

Market1501

method features 1 5 10 15 20 25 30 35 40 45 50

OW-REID

TriNet 0.98 0.98 0.98 0.98 0.98 0.98 0.98 0.98 0.98 - -

GOG 0.81 0.82 0.83 0.85 0.85 0.86 0.86 0.87 0.88 - -

GOG+TriNet 0.95 0.96 0.96 0.97 0.97 0.97 0.96 0.97 0.97 - -

OSVM

TriNet 0.97 0.97 0.98 0.98 0.98 0.98 0.98 0.98 0.98 - -

GOG 0.82 0.82 0.82 0.84 0.83 0.83 0.81 0.83 0.86 - -

GOG+TriNet 0.96 0.96 0.96 0.97 0.96 0.96 0.95 0.96 0.97 - -

TriNet 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 - -

GOG 0.80 0.79 0.79 0.80 0.79 0.80 0.79 0.80 0.84 - -

GOG+TriNet 0.94 0.94 0.94 0.94 0.93 0.93 0.93 0.93 0.95 - -

DukeMTMC-reID

method features 1 5 10 15 20 25 30 35 40 45 50

OW-REID

TriNet 0.71 0.71 0.72 0.72 0.74 0.74 0.76 0.79 0.79 0.85 -

GOG 0.70 0.71 0.76 0.79 0.82 0.79 0.80 0.80 0.80 0.84 -

GOG+TriNet 0.74 0.75 0.79 0.82 0.84 0.82 0.83 0.83 0.83 0.86 -

OSVM

TriNet 0.70 0.70 0.72 0.71 0.71 0.73 0.75 0.78 0.77 0.79 -

GOG 0.68 0.69 0.71 0.74 0.74 0.70 0.71 0.73 0.73 0.77 -

GOG+TriNet 0.74 0.74 0.76 0.77 0.77 0.73 0.74 0.77 0.78 0.80 -

TriNet 0.76 0.76 0.76 0.75 0.76 0.77 0.78 0.80 0.79 0.82 -

GOG 0.69 0.69 0.71 0.72 0.73 0.71 0.70 0.71 0.70 0.74 -

GOG+TriNet 0.74 0.75 0.75 0.76 0.77 0.73 0.72 0.73 0.73 0.76 -

RAiD

method features 2 5 10 15 20 30 40 50 60 70 80

OW-REID

TriNet 0.65 0.69 0.69 0.69 0.69 0.69 0.69 0.69 0.69 0.70 0.71

GOG 0.63 0.66 0.69 0.71 0.72 0.72 0.73 0.75 0.74 0.75 0.76

GOG+TriNet 0.67 0.73 0.76 0.78 0.79 0.78 0.79 0.80 0.76 0.78 0.79

OSVM

TriNet 0.64 0.67 0.68 0.69 0.70 0.69 0.69 0.70 0.69 0.69 0.70

GOG 0.62 0.67 0.69 0.71 0.71 0.70 0.71 0.71 0.71 0.71 0.72

GOG+TriNet 0.69 0.73 0.74 0.76 0.76 0.75 0.76 0.76 0.75 0.75 0.76

TriNet 0.69 0.72 0.72 0.74 0.74 0.73 0.73 0.74 0.74 0.74 0.75

GOG 0.61 0.64 0.66 0.67 0.67 0.66 0.68 0.68 0.68 0.68 0.69

GOG+TriNet 0.68 0.72 0.73 0.74 0.74 0.72 0.73 0.73 0.73 0.73 0.74

Table 2: Mean auROC for the VGS experiment (varying gallery size, c).

Market1501

method features 1 5 10 15 20 25 30 35 40 45 50

OW-REID

TriNet 0.72 0.99 0.99 0.98 0.98 0.98 0.98 0.97 0.97 0.97 0.97

GOG 0.34 0.52 0.75 0.83 0.86 0.87 0.87 0.87 0.83 0.86 0.86

GOG+TriNet 0.99 0.71 0.91 0.96 0.97 0.97 0.97 0.97 0.96 0.96 0.96

OSVM

TriNet 0.55 0.99 0.99 0.98 0.98 0.98 0.97 0.97 0.97 0.97 0.97

GOG 0.67 0.87 0.87 0.86 0.84 0.83 0.81 0.81 0.81 0.80 0.79

GOG+TriNet 0.58 0.98 0.98 0.97 0.96 0.96 0.95 0.95 0.95 0.94 0.93

TriNet 1.00 1.00 0.99 0.99 0.99 0.99 0.99 0.98 0.98 0.98 0.98

GOG 0.65 0.84 0.84 0.83 0.81 0.79 0.78 0.77 0.77 0.75 0.75

GOG+TriNet 0.96 0.97 0.97 0.95 0.94 0.93 0.92 0.91 0.91 0.90 0.89

DukeMTMC-reID

method features 1 5 10 15 20 25 30 35 40 45 50

OW-REID

TriNet 0.79 0.77 0.76 0.74 0.74 0.73 0.72 0.72 0.70 0.69 0.69

GOG 0.65 0.64 0.78 0.81 0.82 0.82 0.81 0.81 0.77 0.77 0.78

GOG+TriNet 0.74 0.72 0.82 0.84 0.84 0.84 0.83 0.83 0.78 0.77 0.78

OSVM

TriNet 0.78 0.77 0.76 0.72 0.70 0.71 0.69 0.68 0.66 0.65 0.65

GOG 0.79 0.77 0.78 0.75 0.74 0.74 0.71 0.71 0.68 0.68 0.67

GOG+TriNet 0.84 0.82 0.82 0.78 0.77 0.76 0.74 0.74 0.71 0.71 0.70

TriNet 0.85 0.85 0.81 0.77 0.76 0.75 0.73 0.72 0.71 0.70 0.70

GOG 0.76 0.73 0.77 0.73 0.73 0.73 0.71 0.72 0.70 0.69 0.69

GOG+TriNet 0.83 0.81 0.81 0.77 0.77 0.76 0.73 0.74 0.72 0.71 0.71

RAiD

method features 1 5 10 15 20 25 30 35 40 45 50

OW-REID

TriNet 0.62 0.76 0.73 0.71 0.69 0.68 0.67 0.67 0.64 - -

GOG 0.42 0.55 0.68 0.73 0.74 0.74 0.75 0.76 0.75 - -

GOG+TriNet 0.82 0.70 0.77 0.79 0.79 0.76 0.76 0.77 0.75 - -

OSVM

TriNet 0.63 0.78 0.73 0.71 0.69 0.68 0.67 0.67 0.65 - -

GOG 0.64 0.75 0.73 0.72 0.71 0.71 0.69 0.69 0.68 - -

GOG+TriNet 0.62 0.83 0.79 0.77 0.75 0.74 0.73 0.73 0.71 - -

TriNet 0.92 0.83 0.79 0.75 0.73 0.72 0.71 0.70 0.68 - -

GOG 0.60 0.70 0.70 0.69 0.67 0.68 0.67 0.66 0.65 - -

GOG+TriNet 0.80 0.81 0.77 0.75 0.73 0.72 0.71 0.70 0.68 - -

Comparison with the State of the Art. The met-

hod in (Wang et al., 2016) is not compatible to our

evaluation protocol, because it only applies to a sin-

gle view pair and requires all query data as a batch,

rather than as a stream, as typically happens in real-

world applications and in OW-REID. Also, none of

the methods presented in (Zheng et al., 2016b; Zhu

et al., 2017) is compatible to our experimental proto-

col, as their operation requires the availability of an

annotated, non-target dataset. Aiming at a ﬂexible

and easy-to-setup method, OW-REID does not have

this requirement. Additionally, there is no available

implementation for (Zheng et al., 2016b; Zhu et al.,

2017), so they cannot be tested according to our ex-

perimental protocol. However, in order to provide a

basis of comparison of OW-REID to state of the art

Novelty Detection for Person Re-identiﬁcation in an Open World

407

Table 3: Mean auDIR for the VPV experiment (varying person views, v

Market1501

method features 1 5 10 15 20 25 30 35 40 45 50

KNFST

TriNet 0.97 0.98 0.98 0.98 0.97 0.97 0.96 0.96 0.96 - -

GOG 0.70 0.71 0.72 0.75 0.74 0.76 0.74 0.74 0.74 - -

GOG+TriNet 0.93 0.94 0.94 0.95 0.94 0.95 0.94 0.93 0.93 - -

OSVM

TriNet 0.96 0.96 0.97 0.98 0.97 0.97 0.96 0.97 0.96 - -

GOG 0.71 0.71 0.71 0.74 0.73 0.74 0.71 0.72 0.72 - -

GOG+TriNet 0.93 0.94 0.94 0.95 0.94 0.94 0.93 0.93 0.93 - -

TriNet 0.98 0.98 0.98 0.98 0.98 0.98 0.97 0.97 0.97 - -

GOG 0.69 0.69 0.69 0.71 0.70 0.71 0.68 0.68 0.70 - -

GOG+TriNet 0.92 0.92 0.92 0.93 0.91 0.91 0.90 0.90 0.91 - -

DukeMTMC-reID

method features 1 5 10 15 20 25 30 35 40 45 50

KNFST

TriNet 0.54 0.55 0.56 0.56 0.58 0.54 0.53 0.52 0.54 0.63 -

GOG 0.47 0.49 0.54 0.60 0.64 0.59 0.59 0.56 0.57 0.65 -

GOG+TriNet 0.56 0.58 0.62 0.67 0.69 0.64 0.63 0.60 0.62 0.70 -

OSVM

TriNet 0.51 0.53 0.54 0.54 0.55 0.51 0.51 0.49 0.51 0.57 -

GOG 0.46 0.48 0.51 0.56 0.59 0.52 0.52 0.52 0.53 0.60 -

GOG+TriNet 0.56 0.57 0.59 0.63 0.64 0.58 0.58 0.56 0.59 0.65 -

TriNet 0.55 0.56 0.57 0.57 0.58 0.54 0.53 0.52 0.53 0.59 -

GOG 0.46 0.48 0.51 0.55 0.59 0.54 0.53 0.51 0.52 0.59 -

GOG+TriNet 0.56 0.58 0.59 0.62 0.65 0.59 0.57 0.55 0.57 0.64 -

RAiD

method features 2 5 10 15 20 30 40 50 60 70 80

KNFST

TriNet 0.37 0.46 0.48 0.49 0.49 0.48 0.48 0.48 0.49 0.50 0.51

GOG 0.36 0.44 0.49 0.52 0.54 0.53 0.55 0.56 0.56 0.57 0.59

GOG+TriNet 0.46 0.56 0.60 0.63 0.64 0.63 0.64 0.65 0.63 0.64 0.66

OSVM

TriNet 0.37 0.43 0.46 0.48 0.48 0.47 0.47 0.48 0.48 0.49 0.50

GOG 0.37 0.46 0.50 0.53 0.54 0.52 0.54 0.54 0.54 0.54 0.56

GOG+TriNet 0.45 0.55 0.58 0.61 0.62 0.60 0.61 0.61 0.62 0.62 0.63

TriNet 0.40 0.47 0.49 0.52 0.53 0.51 0.51 0.51 0.52 0.53 0.54

GOG 0.35 0.43 0.47 0.50 0.51 0.49 0.51 0.51 0.52 0.52 0.53

GOG+TriNet 0.44 0.54 0.57 0.60 0.60 0.58 0.60 0.60 0.60 0.60 0.61

Table 4: Mean auDIR for the VGS experiment (varying gallery size, c).

Market1501

method features 1 5 10 15 20 25 30 35 40 45 50

KNFST

TriNet 0.72 0.98 0.98 0.98 0.97 0.97 0.97 0.96 0.96 0.96 0.95

GOG 0.34 0.48 0.68 0.75 0.76 0.75 0.74 0.74 0.69 0.70 0.69

GOG+TriNet 0.99 0.71 0.90 0.94 0.95 0.95 0.94 0.94 0.92 0.92 0.92

OSVM

TriNet 0.55 0.98 0.98 0.98 0.97 0.97 0.96 0.96 0.96 0.96 0.95

GOG 0.67 0.83 0.81 0.78 0.75 0.72 0.70 0.69 0.68 0.67 0.64

GOG+TriNet 0.58 0.97 0.97 0.95 0.94 0.94 0.93 0.92 0.91 0.91 0.90

TriNet 1.00 0.99 0.99 0.98 0.98 0.98 0.97 0.97 0.97 0.97 0.96

GOG 0.65 0.80 0.78 0.76 0.72 0.69 0.67 0.66 0.65 0.63 0.61

GOG+TriNet 0.96 0.97 0.95 0.94 0.92 0.91 0.89 0.89 0.88 0.87 0.86

DUKEMTMC-reID

method features 1 5 10 15 20 25 30 35 40 45 50

KNFST

TriNet 0.72 0.70 0.66 0.61 0.58 0.56 0.54 0.53 0.51 0.50 0.49

GOG 0.57 0.55 0.66 0.65 0.65 0.63 0.60 0.60 0.56 0.56 0.55

GOG+TriNet 0.67 0.66 0.73 0.72 0.70 0.69 0.66 0.65 0.59 0.59 0.59

OSVM

TriNet 0.71 0.69 0.65 0.58 0.54 0.53 0.50 0.48 0.45 0.44 0.43

GOG 0.71 0.68 0.67 0.61 0.59 0.57 0.53 0.53 0.50 0.49 0.48

GOG+TriNet 0.77 0.75 0.73 0.67 0.65 0.63 0.59 0.59 0.55 0.55 0.53

TriNet 0.77 0.76 0.69 0.61 0.58 0.56 0.52 0.51 0.49 0.48 0.46

GOG 0.68 0.65 0.66 0.60 0.59 0.58 0.54 0.54 0.53 0.51 0.50

GOG+TriNet 0.76 0.74 0.73 0.67 0.65 0.64 0.60 0.60 0.58 0.57 0.56

RAiD

method features 1 5 10 15 20 25 30 35 40 45 50

KNFST

TriNet 0.62 0.66 0.58 0.51 0.48 0.45 0.44 0.42 0.39 - -

GOG 0.42 0.48 0.56 0.57 0.56 0.54 0.53 0.52 0.51 - -

GOG+TriNet 0.82 0.65 0.68 0.67 0.65 0.60 0.59 0.58 0.55 - -

OSVM

TriNet 0.63 0.68 0.58 0.51 0.48 0.44 0.43 0.41 0.39 - -

GOG 0.64 0.67 0.62 0.57 0.53 0.51 0.49 0.48 0.45 - -

GOG+TriNet 0.62 0.77 0.70 0.64 0.61 0.58 0.56 0.55 0.52 - -

TriNet 0.92 0.71 0.62 0.55 0.51 0.48 0.47 0.44 0.42 - -

GOG 0.60 0.63 0.58 0.55 0.51 0.49 0.47 0.45 0.43 - -

GOG+TriNet 0.80 0.75 0.68 0.63 0.60 0.57 0.55 0.53 0.50 - -

methods, we present its performance comparatively

to (Zhu et al., 2017), on the Market1501 dataset on

which (Zhu et al., 2017) has been trained and evalua-

ted.

Table 5 provides the results of this comparison.

(Zhu et al., 2017) utilizes 500 random persons for trai-

ning, which is in contrast to the training set provided

by the authors of the Market1501 dataset (751 trai-

ning persons). TriNet features are learned using the

standard training set, thus, to avoid bias, and inline

with our previous experiments, we utilized only the

testing part of the dataset. To compensate for rand-

omness, we repeat for 300 trials. The rest parameters

were as in (Zhu et al., 2017), i.e., |L

| = 2 and c = 10.

Under these settings, v

is small; as discussed ear-

lier, this is unfavorable to OW-REID which performs

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

408

Figure 2: Qualitative results on the RAiD dataset. Rows 1-2: gallery images. Rows 3-6: correct (green) and incorrect (red)

novelty detection results (see text for details).

Table 5: True Target Rate (TTR, %) at varying False Target

Rates (FTRs, %).

Method 1% 5% 10% 20% 30%

(Zhu et al., 2017) 26.81 52.73 66.47 79.66 86.16

OW-REID with

GOG+TriNet (proposed)

13.53 46.20 66.66 82.10 87.96

much better when a larger number of views per person

is available (v

≥ 10). Nevertheless, we observe that

OW-REID performs better than (Zhu et al., 2017) for

larger FTR. Moreover, we note that OW-REID is di-

rectly applicable to new datasets, without the need for

learning features or metrics under the new conditions.

Qualitative Results. Figure 2 shows qualitative re-

sults from the application of OW-REID on the RAiD

dataset. The ﬁrst two rows show representative ima-

ges in a gallery of 10 persons. Two images per person

are shown, each from a different view. Subsequent

rows show indicative results of (a) known persons cor-

rectly identiﬁed as such (true negative), (b) known

persons incorrectly identiﬁed as unknown (false posi-

tive), (c) unknown persons correctly identiﬁed as such

(true positive) and (d) unknown persons incorrectly

identiﬁed as known (false negative).

Overall, we observe that our method was success-

ful in identifying known persons as such, even under

different pose and/or lighting settings (third row). Ne-

vertheless, some false positives occur as depicted in

the fourth row. Such false positives never occurred

in cases where the person’s clothing color is promi-

nent, e.g. orange and striped cyan t-shirts (4th and

8th column of gallery images, respectively), due to

Novelty Detection for Person Re-identiﬁcation in an Open World

409

less ambiguity. A common cause of false positives

is the intervention of another person (cases 5 and 6).

Similarly, we observe that a variety of unknown per-

sons were identiﬁed as such (ﬁfth row), while most of

false negatives (sixth row) share similar appearance

to gallery persons. Some obvious mistakes, such as

the fourth case in sixth row, are considered as short-

comings of the method and require further investiga-

tion. Finally, we should note that there exist corner

cases, where a person is identiﬁed as known due to

some images, while as unknown due to others. The-

refore, if multiple images are available (e.g. video se-

quence), we should consider aggregation of multiple

novelty identiﬁcations to a single dominant result.

5 SUMMARY AND FUTURE

WORK

We presented a new approach to the problem of open-

world person re-identiﬁcation. The method is based

on a novelty detection technique appied to person des-

criptors that do not require re-training when the sy-

stem is required to operate in a new setting (new ca-

meras, illumination conditions, etc). Extensive ex-

periments indicated the improved performance of the

proposed approach over baseline and state of the art

methods. Ongoing work includes testing of the ap-

proach when it operates in an online fashion, i.e., in a

setup that involves the incremental learning (Liu et al.,

2017) and incorporation of novel persons in the gal-

lery of known persons. Future work aims at the uti-

lization of multiple query images per person (MvsM

scenario (Lisanti et al., 2015)) and further person des-

cription models, to further increase novelty detection

accuracy.

ACKNOWLEDGMENTS

This work is partially funded by the H2020 projects

CONNEXIONs (GA 786731) and Co4Robots (GA

731869).

REFERENCES

(2016). MARS: A Video Benchmark for Large-Scale Person

Re-identiﬁcation.

Bendale, A. and Boult, T. E. (2016). Towards Open Set

Deep Networks. In CVPR.

Bodesheim, P., Freytag, A., Rodner, E., Kemmler, M., and

Denzler, J. (2013). Kernel null space methods for no-

velty detection. In CVPR.

Brun, L., Conte, D., Foggia, P., and Vento, M. (2011). Pe-

ople re-identiﬁcation by graph kernels methods. In

International Workshop, GbRPR.

Chang, X., Hospedales, T. M., and Xiang, T. (2018). Multi-

level factorisation net for person re-identiﬁcation. In

CVPR.

Chen, J., Zhang, Z., and Wang, Y. (2015). Relevance metric

learning for person re-identiﬁcation by exploiting lis-

twise similarities. IEEE Transactions on Image Pro-

cessing, 24(12):4741–4755.

Chen, W., Chen, X., Zhang, J., and Huang, K. (2017).

Beyond triplet loss: a deep quadruplet network for

person re-identiﬁcation. In CVPR.

Das, A., Chakraborty, A., and Roy-Chowdhury, A. K.

(2014). Consistent re-identiﬁcation in a camera net-

work. In ECCV.

Farenzena, M., Bazzani, L., Perina, A., Murino, V., and

Cristani, M. (2010). Person re-identiﬁcation by

symmetry-driven accumulation of local features. In

CVPR.

Gou, M., Camps, O., and Sznaier, M. (2017). moM: Mean

of Moments Feature for Person Re-identiﬁcation. In

CVPR.

unther, M., Cruz, S., Rudd, E. M., and Boult, T. E.

(2017). Toward Open-Set Face Recognition. CoRR,

abs/1705.01567.

Hermans, A., Beyer, L., and Leibe, B. (2017). In Defense

of the Triplet loss for Person Re-Identiﬁcation. CoRR,

abs/1703.07737.

Jain, A. K. and Li, S. Z. (2011). Handbook of face recogni-

tion. Springer.

Jose, C. and Fleuret, F. (2016). Scalable metric learning via

weighted approximate rank component analysis. In

ECCV.

Kliger, M. and Fleishman, S. (2018). Novelty Detection

with GAN. cite arxiv:1802.10560.

Li, W., Zhu, X., and Gong, S. (2018a). Harmonious atten-

tion network for person re-identiﬁcation. In CVPR.

Li, X., Wu, A., and Zheng, W.-S. (2018b). Adversarial

open-world person re-identiﬁcation. In ECCV.

Liao, S., Hu, Y., Zhu, X., and Li, S. Z. (2015). Person re-

identiﬁcation by local maximal occurrence represen-

tation and metric learning. In CVPR.

Lisanti, G., Masi, I., Bagdanov, A. D., and Del Bimbo,

A. (2015). Person re-identiﬁcation by iterative re-

weighted sparse ranking. PAMI, 37(8):1629–1642.

Liu, J., Lian, Z., Wang, Y., and Xiao, J. (2017). Incremental

Kernel Null Space Discriminant Analysis for Novelty

Detection. In CVPR.

Maaten, L. v. d. and Hinton, G. (2008). Visualizing data

using t-SNE. Journal of machine learning research,

9(Nov):2579–2605.

Martinel, N., Micheloni, C., and Foresti, G. L. (2015).

Kernelized saliency-based person re-identiﬁcation

through multiple metric learning. IEEE Transactions

on Image Processing, 24(12):5645–5658.

Masana, M., Ruiz, I., Serrat, J., van de Weijer, J., and Lo-

pez, A. M. (2018). Metric learning for novelty and

anomaly detection. arXiv preprint arXiv:1808.05492.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

410

Matsukawa, T., Okabe, T., Suzuki, E., and Sato, Y.

(2016a). Hierarchical gaussian descriptor for person

re-identiﬁcation. In CVPR.

Matsukawa, T., Okabe, T., Suzuki, E., and Sato, Y.

(2016b). Hierarchical Gaussian Descriptor for Person

Re-identiﬁcation. In CVPR, pages 1363–1372.

Matsukawa, T. and Suzuki, E. (2016). Person re-

identiﬁcation using CNN features learned from com-

bination of attributes. In ICPR.

Perera, P. and Patel, V. M. (2018). Learning deep fe-

atures for one-class classiﬁcation. arXiv preprint

arXiv:1801.05365.

Qian, X., Fu, Y., Wang, W., Xiang, T., Wu, Y., Jiang, Y.-

G., and Xue, X. (2017). Pose-Normalized Image Ge-

neration for Person Re-identiﬁcation. arXiv preprint

arXiv:1712.02225.

Ruff, L. et al. (2018). Deep one-class classiﬁcation. In

ICML.

Sarfraz, M. S., Schumann, A., Eberle, A., and Stiefelhagen,

R. (2018). A pose-sensitive embedding for person re-

identiﬁcation with expanded cross neighborhood re-

ranking. In CVPR.

Sch

olkopf, B., Smola, A. J., Williamson, R. C., and Bartlett,

P. L. (2000). New support vector algorithms. Neural

computation, 12(5):1207–1245.

Shi, Z., Hospedales, T. M., and Xiang, T. (2015). Trans-

ferring a semantic representation for person re-

identiﬁcation and search. In CVPR.

Song, C., Huang, Y., Ouyang, W., and Wang, L. (2018).

Mask-guided contrastive attention model for person

re-identiﬁcation. In CVPR.

Su, C., Li, J., Zhang, S., Xing, J., Gao, W., and Tian, Q.

(2017). Pose-driven Deep Convolutional Model for

Person Re-identiﬁcation. In ICCV, pages 3980–3989.

IEEE.

Su, C., Zhang, S., Xing, J., Gao, W., and Tian, Q.

(2016). Deep attributes driven multi-camera person

re-identiﬁcation. In ECCV.

Wang, H., Zhu, X., Xiang, T., and Gong, S. (2016). To-

wards unsupervised open-set person re-identiﬁcation.

In ICIP.

Wang, J., Wang, Z., Gao, C., Sang, N., and Huang,

R. (2017). Deeplist: Learning deep features with

adaptive listwise constraint for person reidentiﬁcation.

TCSVT, 27(3):513–524.

Wang, J., Zhou, S., Wang, J., and Hou, Q. (2018). Deep

ranking model by large adaptive margin learning for

person re-identiﬁcation. Pattern Recognition, 74:241–

252.

Wang, T., Gong, S., Zhu, X., and Wang, S. (2014). Person

re-identiﬁcation by video ranking. In ECCV.

Wu, S., Chen, Y.-C., Li, X., Wu, A.-C., You, J.-J., and

Zheng, W.-S. (2016). An enhanced deep feature re-

presentation for person re-identiﬁcation. In WACV.

Xu, J., Zhao, R., Zhu, F., Wang, H., and Ouyang, W. (2018).

Attention-aware compositional network for person re-

identiﬁcation. In CVPR.

Yang, Y., Yang, J., Yan, J., Liao, S., Yi, D., and Li,

S. Z. (2014). Salient color names for person re-

identiﬁcation. In ECCV.

Ye, M., Ma, A. J., Zheng, L., Li, J., and Yuen, P. C. (2017).

Dynamic label graph matching for unsupervised video

re-identiﬁcation. In ICCV.

Yu, H.-X., Wu, A., and Zheng, W.-S. (2017). Cross-view

asymmetric metric learning for unsupervised person

re-identiﬁcation. In ICCV.

Zhang, L., Xiang, T., and Gong, S. (2016). Learning a dis-

criminative null space for person re-identiﬁcation. In

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 Fe-

ature Decomposition and Fusion. In CVPR.

Zheng, L., Huang, Y., Lu, H., and Yang, Y. (2017a). Pose

invariant embedding for deep person re-identiﬁcation.

arXiv preprint arXiv:1701.07732.

Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., and Tian,

Q. (2015a). Scalable Person Re-identiﬁcation: A Ben-

chmark. In ICCV.

Zheng, L., Wang, S., Tian, L., He, F., Liu, Z., and Tian, Q.

(2015b). Query-adaptive late fusion for image search

and person re-identiﬁcation. In CVPR.

Zheng, L., Yang, Y., and Hauptmann, A. G. (2016a). Person

Re-identiﬁcation: Past, Present and Future. CoRR,

abs/1610.02984.

Zheng, W.-S., Gong, S., and Xiang, T. (2016b). To-

wards open-world person re-identiﬁcation by one-shot

group-based veriﬁcation. PAMI, 38(3):591–606.

Zheng, Z., Zheng, L., and Yang, Y. (2017b). Unlabeled

Samples Generated by GAN Improve the Person Re-

identiﬁcation Baseline in vitro. In ICCV.

Zhong, Z., Zheng, L., Cao, D., and Li, S. (2017). Re-

ranking person re-identiﬁcation with k-reciprocal en-

coding. In CVPR.

Zhou, S., Wang, J., Wang, J., Gong, Y., and Zheng, N.

(2017). Point to set similarity based deep feature lear-

ning for person re-identiﬁcation. In CVPR.

Zhu, X., Wu, B., Huang, D., and Zheng, W.-S. (2017). Fast

Open-World Person Re-Identiﬁcation. IEEE Tran-

sactions on Image Processing.

Novelty Detection for Person Re-identiﬁcation in an Open World

411