Towards Fully Automated Person Re-identiﬁcation

Matteo Taiana, Dario Figueira, Athira Nambiar, Jacinto Nascimento and Alexandre Bernardino

Institute for Systems and Robotics, IST, Lisboa, Portugal

Keywords:

Re-Identiﬁcation, Pedestrian Detection, Camera Networks, Video Surveillance.

Abstract:

In this work we propose an architecture for fully automated person re-identiﬁcation in camera networks. Most

works on re-identiﬁcation operate with manually cropped images both for the gallery (training) and the probe

(test) set. However, in a fully automated system, re-identiﬁcation algorithms must work in series with person

detection algorithms, whose output may contain false positives, detections of partially occluded people and

detections with bounding boxes misaligned to the people. These effects, when left untreated, may signiﬁcantly

jeopardise the performance of the re-identiﬁcation system. To tackle this problem we propose modiﬁcations to

classical person detection and re-identiﬁcation algorithms, which enable the full system to deal with occlusions

and false positives. We show the advantages of the proposed method on a fully labelled video data set acquired

by 8 high-resolution cameras in a typical ofﬁce scenario at working hours.

1 INTRODUCTION

This paper tackles the problem of person re-

identiﬁcation (RE-ID) in camera networks. Given a

set of pictures of previously observed persons, a prac-

tical RE-ID system must locate and recognise such

people in the stream of images ﬂowing from the cam-

era network. Its classical applications are in the ﬁeld

of video surveillance and security systems as well as

in human-machine interfaces, robotics, gaming and

smart spaces. In this work we consider a set of cam-

eras with low overlap covering our research institute

facilities. When a person enters the space or passes on

some key locations where her/his identity can be ver-

iﬁed (face recognition, access control), pictures are

acquired and stored in a gallery, associated to the re-

spective identity. Such gallery provides the training

data to develop methods to recognise the person on

the other images of the camera network. The full RE-

ID system can thus be used to update the location of

a person on the covered space along time, supporting

research on several topics of our interest (modelling

activities, mining physical social networks, human-

robot interaction).

The problem we tackle is very challenging due to

a multitude of factors. The placement of a camera in-

ﬂuences the perspective and the amount of occlusion

a person is observed under, and the range of distances

people are imaged at, eventually deﬁning their height

in the image. Different illumination and camera op-

tics change the brightness and colour characteristics

of the measurements. Finally, persons may change

Figure 1: A typical re-identiﬁcation algorithm is based on

a gallery set: a database that contains the persons to be re-

identiﬁed at evaluation time. People detected in other im-

ages (probes) are matched to such database with the intent

of recognising their identities. Classically, re-identiﬁcation

algorithms are evaluated with manually cropped probes. In

this work we study the effect of using automatic probe de-

tection in the full re-identiﬁcation system.

their clothing and other appearance traits along time

(possibly for disguise purposes), making the RE-ID

problem particularly difﬁcult even for humans.

Most recent state-of-the-art algorithms on RE-ID

are developed and evaluated as a matching problem.

The appearance information of persons stored in the

gallery (training set) is matched against manually

cropped images of persons in the network cameras

(probe data). See Fig. 1 for an illustration of the pro-

cess. However, in most RE-ID applications of inter-

est, it is necessary to detect the probe person’s bound-

ing boxes (BB) in an automated way. This is com-

140

Taiana M., Figueira D., Nambiar A., Nascimento J. and Bernardino A..

Towards Fully Automated Person Re-identiﬁcation.

DOI: 10.5220/0004682301400147

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 140-147

ISBN: 978-989-758-009-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

monly accomplished using one of two approaches:

background subtraction and pattern recognition. We

focus on methods based on the latter, because of their

applicability to a wider range of scenarios, including

moving/shaking cameras and single frame images.

There are two main classes of pedestrian detection

(PD) algorithms: monolithic or holistic (human as a

whole) (Doll

ar et al., 2010) and part-based (human as

a composition of parts) (Girshick et al., 2011). In this

work we use our implementation (Taiana et al., 2013)

of a monolithic algorithm based on (Doll

ar et al.,

2010) for PD and the algorithm described in (Figueira

et al., 2013) for RE-ID.

Both in background subtraction and in pattern

recognition methods, the output of PD is subject to

several forms of “noise”, in particular missed detec-

tions, false positives and BB’s misaligned with the de-

tected people. When a RE-ID algorithm is combined

with an automatic PD algorithm, the performance of

the former will suffer from the imperfections of the

data presented on its input. Furthermore, it is worth

noticing that a class of correct detections is particu-

larly difﬁcult for a RE-ID algorithm to handle: the de-

tections of partially occluded people. Currently very

few works address these problems.

To leverage the combination of PD and RE-ID al-

gorithms in a fully automated RE-ID system, we pro-

pose two important extensions. On one hand, the

PD output is used to ﬁlter detections with a large de-

gree of occlusion, which most likely contain a large

amount of data not related to the detected person. Do-

ing so prevents mis-identiﬁcations at the RE-ID stage.

On the other hand, the RE-ID module is trained to di-

rectly represent a class of false positives commonly

detected in the camera network. This minimises the

false associations of probe “noise” to actual persons

in the gallery. We evaluate our system on a recently

developed data set of high-deﬁnition cameras. Eight

cameras covering a large space of our research in-

stitute were used to acquire data simultaneously for

thirty minutes during rush hour (lunch time). Images

were fully labelled with BB locations, persons’ iden-

tities and occlusion and crowd ﬂags. We show that

the proposed add-ons to PD and RE-ID algorithms are

important features in a fully automated RE-ID sys-

tem.

2 RELATED WORK

The problem of re-identiﬁcation entails the recogni-

tion of people traversing the ﬁeld of view of differ-

ent non-overlapping sensors. State-of-the-art meth-

ods regarding this topic are often characterized by

differences in the contextual knowledge, the learning

methodology, the feature extraction, the data associa-

tion and the type of gallery.

Approaches relying on contextual knowledge from

surrounding people using human signature (Zheng

et al., 2009), or attribute weighting features (Liu et al.,

2012a) were proposed. Another class of approaches

was developed to deal with pose variation (Bak et al.,

2012) using Mean Riemannian Covariance patches

computed inside the bounding box of the pedes-

trian. Shape and appearance context to model the spa-

tial distributions of appearance relative to body parts

(Wang et al., 2007), triangular graph model (Gheis-

sari et al., 2006) and part-based models (Corvee et al.,

2012; Cheng et al., 2011) constitute valuable alterna-

tives handling pose variability. Methodologies using

supervised learning, were also suggested. They usu-

ally include discriminative models such as SVM and

boosting for feature learning (Gray and Tao, 2008;

Prosser et al., 2010; Zheng et al., 2011), in the at-

tempt to iteratively select the most reliable set of

features. Another direction concerns the learning

of task-speciﬁc distance functions with metric algo-

rithms (Zheng et al., 2011; Mignon and Jurie, 2012;

Hirzer et al., 2012; Li and Wang, 2013; Li et al.,

2012). For all these methods, training samples with

identity labels is mandatory. Approaches focusing on

direct distance metrics were also proposed in the re-

identiﬁcation context. Symmetry-Driven Accumula-

tion of Local Features was proposed in (Farenzena

et al., 2010). In (Ma et al., 2012) the BiCov de-

scriptor was presented, combining Gabor ﬁlters and

covariance descriptors to account for both illumina-

tion and background changes. In (Liu et al., 2012a)

it was shown that certain appearance features can be

more important than others and selected online. In

(Mogelmose et al., 2013a) it was shown that using

RGB, depth and thermal features in a joined clas-

siﬁer are able to improve the re-identiﬁcation per-

formance. Feature extraction for fast matching was

also investigated. In (Hamdoun et al., 2008) a KD-

tree was used to store per-person signatures. Random

forests were proposed in (Liu et al., 2012b) to weight

the most informative features, while code-book rep-

resentations were used in (Jungling et al., 2011). Per-

son re-identiﬁcation can be formulated as a data as-

sociation problem, in matching observations from

pairs of cameras. Several methods were proposed

to learn this association space. Large-Margin Near-

est Neighbor (with Rejection) was proposed (Dik-

men et al., 2011) to learn the most suitable metric

to match data from two distinct cameras, while other

metrics include Probabilistic Relative Distance Com-

parison (Zheng et al., 2011). Rank-loss optimization

TowardsFullyAutomatedPersonRe-identification

141

was used to improve accuracy in re-identiﬁcation (Wu

et al., 2011) and a variant of Locally Preserving Pro-

jections (Harandi et al., 2012) was formulated over

a Riemmanian manifold. Recently Pairwise Con-

strained Component Analysis was proposed for met-

ric learning tailored to address the scenarios with a

small set of examples (Mignon and Jurie, 2012). Also

(Pedagadi et al., 2013) suggested a metric learning

for re-identiﬁcation, where a Local Fisher Discrimi-

nant Analysis is deﬁned by a training set. Most of the

aforementioned works use the VIPeR (Gray and Tao,

2008), ETHZ (Ess et al., 2007) and i-LIDS (Zheng

et al., 2009) data sets, focusing only on the matching

problem, thus neglecting the automatic detection of

people.

Methods that actively integrate pedestrian detec-

tion and re-identiﬁcation are still scarce in the liter-

ature. Yet, few examples are available. In (Corvee

et al., 2012; Bak et al., 2012), pedestrian detection

and re-identiﬁcation are used, but not actually inte-

grated. The above frameworks rely on the assump-

tion that the people must be accurately detected. The

work that relates the most with ours is that intro-

duced in (Mogelmose et al., 2013b). In that work,

the system full ﬂow (i.e. pedestrian detection and

re-identiﬁcation) is presented with a transient gallery

to tackle open scenarios. However, important issues

such as how re-identiﬁcation performance is penal-

ized when pedestrian detection or tracking failures ex-

ist are not explored. The goal of our paper is precisely

to explore how to enhance the link between pedestrian

detection and re-identiﬁcation algorithms to improve

the overall performance.

3 ARCHITECTURE

For the algorithms in the state of the art, the data

for the Re-Identiﬁcation (RE-ID) problem is provided

in the shape of hand-cropped Bounding Boxes (BB),

rather than in the shape of full image frames. Such

BB’s are centered around fully visible, upright per-

sons. The data is usually partitioned into training and

test set and the focus of the RE-ID algorithms is on

feature extraction and BB classiﬁcation. However,

the purpose of an automated RE-ID system is that of

re-identifying people directly in images, without re-

quiring manual intervention to produce the BB’s. The

natural candidate for substituting the hand-made an-

notation process is a Pedestrian Detection (PD) algo-

rithm, which takes one image as input and produces

a set of BB’s as output. The resulting architecture is

depicted in Figure 2.

Integrating PD and RE-ID poses several issues.

Detecting people in images is a hard task, in fact even

the best detectors in the state of the art are subject to

the production of at least two types of errors: False

Positives (FP) and Missed Detections (MD). Such er-

rors have an impact on the performance of the com-

pounded system: FP’s generate BB’s which is impos-

sible for the system to correctly classify as one of the

persons in the training set. MD’s, on the other hand,

cause an individual to simply go undetected, and thus,

unclassiﬁed. Even the correctly detected persons may

give rise to some difﬁculties: (1) the PD algorithm

can generate a BB not centred around the person or at

a non-optimal scale – this might hinder the feature

extraction phase, prior to the classiﬁcation, (2) the

detected person may be partially occluded, yet again

hampering feature extraction, and ﬁnally (3) there can

be the case of detecting people who are not part of the

RE-ID training set, posing an issue similar to that of

FP’s: there is no correct class that the system can at-

tribute to them. To limit the complexity of the prob-

lem, we constrain the scenario with a closed-space as-

sumption: we require the access to the surveilled area

to be granted exclusively to people listed in the train-

ing set. In the rest of this section we propose an ar-

chitecture which integrates the PD and RE-ID stages,

solving some of the aforementioned issues.

3.1 Occlusion Filter

We devise the Occlusion Filter – a ﬁltering block be-

tween the PD and the RE-ID modules – with the intent

of improving the RE-ID performance. The Occlu-

sion Filter uses geometrical reasoning to reject BB’s

which can harm the performance of the RE-ID stage,

the harmful BB’s being the ones depicting partially

occluded people. A BB including a person appear-

ing under partial occlusion generates features differ-

ent from a BB including the same person under full

visibility conditions. When the partial occlusion is

caused by a second person standing between the cam-

era and the original person, the extracted features can

be a mixture of those generated by the two people,

making the identity classiﬁcation especially hard (see

illustration in Figure 3). For this reason, it would be

advantageous for the RE-ID module to receive only

BB’s depicting fully visible people. We deﬁne a ﬁrst,

aggressive, operation mode for the Occlusion Filter

which rejects all the detections which overlap with

others, the “RejectAll” mode. Though the visibil-

ity information is not available to the system, it can

be estimated quite accurately with a heuristic based

on scene geometry: in a typical scenario the cam-

era’s perspective projection makes pedestrians closer

to it extend to relatively lower regions of the image.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

142

Pre-trained Pedestrian

Detector

Body-Part

Detector

Camera Images

Re-Identification Classifier

FP Class

ON OFF

Training Set

FP class

TP classes

ON OFF

Occlusion

Filter

FAR Mode

Ovlp %

ALL Mode

Generic Large Training Set

Geometric

Rules

Filter

Figure 2: Architecture of the proposed fully automated Re-Identiﬁcation system. The images acquired by a camera network

are processed by a Pedestrian Detection algorithm to extract candidate Bounding Boxes. The Bounding Boxes are optionally

processed by the Occlusion Filter, which can operate either in the “RejectAll” or the “RejectFarther” mode (see explanation

in text). Lastly, the Re-Identiﬁcation module computes the features corresponding to each Bounding Box and classiﬁes it.

The classiﬁcation can optionally take into account a “False Positive” class.

(a) (b)

Figure 3: Example of body part detection for feature extrac-

tion in two instances: (a) a person appearing with full visi-

bility and (b) under partial occlusion, with detection Bound-

ing Boxes overlap, and feature extraction on the occluded

person mistakenly extracting features from the occluding

pedestrian. The contrast of both images was enhanced for

visualization purposes.

Thus, we design the “RejectFarther” operation mode

for the Occlusion Filter. In this mode the ﬁlter com-

putes the overlap among all pairs of detections in one

image and rejects the one in each overlapping pair

for which the lower side of the BB is higher (as il-

lustrated in Figure 4). Considering the mismatch be-

tween the shape of the pedestrians’ bodies and that of

the BB’s, it is clear that an overlap between BB’s does

not always imply an overlap between the correspond-

ing pedestrians’ projections on the image. We deﬁne

an overlap threshold for the “RejectFarther” mode of

the ﬁlter, considering as overlapping only detections

whose overlap is above such threshold. The impact

of the overlap threshold on the RE-ID performance is

analysed in the following section.

Figure 4: An example of geometrical reasoning: two detec-

tion Bounding Boxes overlap. The comparison between the

lower sides of the two Bounding Boxes leads to the con-

clusion that the person marked with the red, dashed Bound-

ing Box is occluded by the person in the green, continuous

Bounding Box.

3.2 False Positives Class

The second contribution of this work is to adapt the

RE-ID module so that it can deal with the FP’s pro-

duced by the PD. The standard RE-ID module cannot

deal properly with FP’s: each FP turns into a wrongly

classiﬁed instance for the RE-ID. Observing that the

appearance of the FP’s in a given scenario is not com-

pletely aleatory, but is worth modelling (see Fig. 5),

we introduce a FP class for the RE-ID module. In

these conditions a correct output for when a FP is pre-

sented on the RE-ID’s input exists: the FP class. This

change allows us to coherently evaluate the perfor-

mance of the integrated system.

TowardsFullyAutomatedPersonRe-identification

143

Figure 5: Example False Positive samples in the False Pos-

itive Class training set.

4 EXPERIMENTAL RESULTS

We work with the recently proposed high-deﬁnition

data set described in Section 1. For our experiments

we consider the closed-space assumption. A set of

images with pedestrians authorised in the surveilled

space is collected in a training stage and stored in a

gallery set associated to the pedestrians identities. In

our experiment we simulate the closed-space assump-

tion selecting the best images of 7 out of the 8 cameras

sequences for training, and using the last sequence as

a test set. The training set is built hand-picking one

detection per image sequence, per pedestrian, leading

to a total of 230 detections for 76 pedestrians. The

False Positive (FP) class is built with the detections

from the training sequences that have no overlap with

a Ground Truth (GT) Bounding Box (BB), for a total

of 3972 detections. The test image sequence contains

1182 GT BB’s, centered on 20 different people. Such

people are fully visible in 416 occurrences and appear

with some degree of occlusion by other BB’s or trun-

cated by the image border in 766 occasions.

Pedestrian Detection. In this work we use a state-

of-the-art PD system: our implementation (Taiana

et al., 2013) of Doll

ar’s Fastest Pedestrian Detector

in the West (Doll

ar et al., 2010) (FPDW). This mod-

ule generates 1182 detections on the test camera se-

quence. The initial detections are ﬁltered based on

their size, removing the ones whose height is unrea-

sonable given the geometric constraints of the scene

(under 68 pixels). This rejects 159 detections and al-

lows 1023 of them pass. Considering that three pedes-

trians who appear in the test set are not present in the

training set, we remove the corresponding 59 detec-

tions in the test set from the detections’ pool. This

leads to the 964 elements that form the base set of de-

tections. Being FPDW a monolithic detector, it is con-

strained to generate detections which lie completely

inside the image boundary. This naturally generates a

detection set without persons truncated by the image

boundary, facilitating the re-identiﬁcation (RE-ID).

Re-identiﬁcation. We use the state-of-the-art RE-

ID algorithm from (Figueira et al., 2013). Given a

BB provided by the GT or the PD module, the al-

gorithm ﬁrst detects body parts using Pictorial Struc-

tures (Andriluka et al., 2009) (see Figure 3). The al-

gorithm uses a body model consisting of head, torso,

two thighs and two shins during the detection phase.

It then merges the areas corresponding to both thighs

and both shins, respectively. Subsequently, the al-

gorithm extracts color histograms and texture his-

tograms from the distinct image regions. Finally it

performs the classiﬁcation based on the extracted fea-

tures with the Multi-View algorithm (Quang et al.,

2013).

The necessary GT for the RE-ID task is obtained

by processing the original GT and the detections gen-

erated by the PD module. Each detection is associated

with the label of a person or the special label for the

FP class. The assignment is done associating each de-

tection with the label of the GT BB that has the most

overlap with it. The Pascal VOC criterion (Evering-

ham et al., 2010) is used to determine FP’s: when

the intersection between a detection BB and the cor-

responding BB from the original GT is smaller than

half the union of the two, the detection is marked as a

FP.

We use the Cumulative Matching Characteristic

curve (CMC), the de facto standard, for evaluating the

performance of RE-ID algorithms. The CMC shows

how often, on average, the correct person ID is in-

cluded in the best K matches against the training set

for each test image. The overall performance is mea-

sured by the nAUC – the normalized Area Under the

CMC curve.

4.1 Experiments

We evaluate the performance of different conﬁgura-

tions of the RE-ID system as detailed in Table 1. Ini-

tially, we assess the performance of the RE-ID mod-

ule in the conditions that are common in the RE-ID

literature: using manually labelled GT BB’s as test

set. This is done to establish a term of comparison.

Then we evaluate the naive integration of the PD and

RE-ID modules and we contrast that with the inte-

grated system enriched with the introduction of the

FP class in the RE-ID module. Eventually, we assess

the effectiveness of two operating modes for the Oc-

clusion Filter, on the integrated system using the FP

class.

In exp. (A) we perform RE-ID on the 416 fully

visible BB’s provided in the GT, consistently with the

modus operandi of the state of the art. This means

that the RE-ID module works with unoccluded per-

sons and BB’s that are correctly centred and sized.

This provides a meaningful term of comparison when

moving towards a fully automatic RE-ID system (see

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

144

Table 1: GT indicates the use of Ground Truth in an experiment, namely the hand-labelled Bounding Boxes. In Occlusion

Filter, ALL indicates the “RejectAll” mode, and FAR the “RejectFarther” mode. NA indicates that the use of the Occlusion

Filter or the FP class is not applicable to experiments with GT data. Rank1 corresponds to the ﬁrst point of the CMC curve

and indicates how often the correct person ID is the best match against the training set. nAUC corresponds to the area under

the CMC curve. The total number of detections and the amount of corresponding false positives which are passed to the

RE-ID module are listed under Detections and False Positives, respectively.

Exp. GT Occlusion Filter FP class Rank1 (%) nAUC (%) Detections False Positives

A 1 NA NA 42.02 90.21 416 0

B 0 OFF OFF 26.45 72.68 964 155

C 0 OFF ON 31.33 87.60 964 155

D 0 ON ALL ON 39.61 88.01 563 88

E 0 ON FAR 30% ON 34.19 89.14 854 119

10 20 30 40 50 60 70

100

Rank Score

Re−identification %

Experiment A (nAUC 90.21%)

Experiment B (nAUC 72.68%)

Experiment C (nAUC 87.60%)

Experiment D (nAUC 88.01%)

Experiment E (nAUC 89.14%)

Figure 6: Cumulative Matching Characteristic curves com-

paring the performance of various conﬁgurations of the in-

tegrated Re-Identiﬁcation system, for details see Table 1.

Working points which lie comparatively higher or more to

the left of the plot correspond to better performances.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

87.5

88.5

89.5

nAUC %

Overlap %

Figure 7: Normalized area under the Cumulative Matching

Characteristic curve plotted against varying degrees of over-

lap for the Overlap Filter in ’RejectFarther’ mode. Based on

this evidence we set this parameter for exp. (E) to 30%.

Fig. 6 for the CMC curves associated to this and to

other experiments). It is to be noted that this method

of operation is not applicable in a real-world scenario,

since it requires manual annotation of every person in

the video sequence.

In exp. (B) we analyse the performance of the sys-

tem resulting from the naive integration of the PD and

RE-ID modules. The 155 FP’s generated by the detec-

tor are impossible for the RE-ID to classify correctly,

yielding a CMC curve that does not reach 100%. Ap-

plying the CMC evaluation scheme to this case leads

to an unfair comparison with the curve from exp. (A).

We provide this result to underline this point.

Exp. (C) shows that using the FP class in the RE-

ID module allows for a meaningful comparison with

the classic evaluation approach – exp. (A). The RE-

ID module is able to classify a fraction of the FP’s

as such, thus achieving a performance comparable to

that of exp. (A).

Rejecting all the detections affected by mutual

overlap – exp. (D) – leads to a drastic reduction of the

number of detections presented in input to the RE-

ID module. This rejection mode eliminates 42% of

the initial detections, including many depicting unoc-

cluded pedestrians. This is due to this ﬁltering method

not taking advantage of perspective information. In

Figure 4, for instance, one of the persons is fully

visible in spite of the overlap between the bounding

boxes. Because the CMC evaluation is concerned

solely with the quality of the re-identiﬁcation of the

test examples it is provided with, it does not penalise

the drastic reduction in detections. A practical appli-

cation of RE-ID, though, would certainly be adversely

affected by the Missed Detections.

Finally, in exp. (E) we run the Overlap Filter in the

the “RejectFarther” mode. Figure 7 shows the impact

of varying the ﬁlter’s overlap threshold. We empir-

ically choose 30% as the best value of this parame-

ter, and illustrate again in Figure 6 the corresponding

CMC curve. The “RejectFarther” mode proves to be

better than the “RejectAll” both in terms of classiﬁca-

tion performance (showing a 1% increase in nAUC)

and of the number of Missed Detections (rejecting

only 12% of the initial detections).

TowardsFullyAutomatedPersonRe-identification

145

5 CONCLUSIONS

In this work we studied how to combine person detec-

tors and re-identiﬁcation algorithms towards the de-

velopment of fully automated surveillance systems.

Whereas current work on re-identiﬁcation focuses on

matching “clean data”, we analyse the effect of the

unavoidable “noise” produced by automatic detec-

tions methods and devise ways to deal with it. In

particular, a correct treatment of false positives and

occlusions can recover performance levels that are

lost if the modules are naively combined. This re-

quires some additional characterisation of the acqui-

sition process, in terms of collecting false positives

in the environment where the system is deployed and

performing a geometrical analysis to deﬁne the over-

lap ﬁlter criteria. This effort is largely compensated

by the obtained improvements. In future work we aim

at extending our methodology and evaluation criteria

to drop the closed-space assumption and address the

effect of missed detections in the quality and usability

of fully automatic re-identiﬁcation systems.

ACKNOWLEDGEMENTS

This work was partially supported by the FCT

project [PEst-OE/EEI/LA0009/2013], the European

Commission project POETICON++ (FP7-ICT-

288382), the FCT project VISTA [PTDC/EIA-

EIA/105062/2008] and the project High Deﬁnition

Analytics (HDA), QREN - I&D em Co-Promoc¸

13750.

REFERENCES

Andriluka, M., Roth, S., and Schiele, B. (2009). Picto-

rial Structures Revisited: People Detection and Artic-

ulated Pose Estimation. CVPR.

Bak, S., Corv

ee, E., Br

emond, F., and Thonnat, M. (2012).

Boosted human re-identiﬁcation using Riemannian

manifolds. ImaVis.

Cheng, D. S., Cristani, M., Stoppa, M., Bazzani, L., and

Murino, V. (2011). Custom pictorial structures for re-

identiﬁcation. In BMVC.

Corvee, E., Bak, S., and Bremond, F. (2012). People detec-

tion and re-identiﬁcation for multi surveillance cam-

eras. VISAPP.

Dikmen, M., Akbas, E., Huang, T., and Ahuja, N. (2011).

Pedestrian recognition with a learned metric. In

ACCV.

Doll

ar, P., Belongie, S., and Perona, P. (2010). The Fastest

Pedestrian Detector in the West. BMVC.

Ess, A., Leibe, B., and Van Gool, L. (2007). Depth and

Appearance for Mobile Scene Analysis. ICCV.

Everingham, M., Van Gool, L., Williams, C., Winn, J., and

Zisserman, A. (2010). The Pascal visual object classes

(VOC) challenge. IJCV.

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

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

symmetry-driven accumulation of local features. In

CVPR.

Figueira, D., Bazzani, L., Minh, H. Q., Cristani, M.,

Bernardino, A., and Murino, V. (2013). Semi-

supervised multi-feature learning for person re-

identiﬁcation. AVSS.

Gheissari, N., Sebastian, T., and Hartley, R. (2006). Person

reidentiﬁcation using spatiotemporal appearance. In

CVPR.

Girshick, R., Felzenszwalb, P., and McAllester, D. (2011).

Object detection with grammar models. PAMI.

Gray, D. and Tao, H. (2008). Viewpoint invariant pedestrian

recognition with an ensemble of localized features. In

ECCV.

Hamdoun, O., Moutarde, F., Stanciulescu, B., and Steux.,

B. (2008). Person re-identiﬁcation in multi-camera

system by signature based on interest point descriptors

collected on short video sequences. In ICDSC.

Harandi, M. T., Sanderson, C., Wiliem, A., and Lovell,

B. C. (2012). Kernel analysis over Riemannian mani-

folds for visual recognition of actions, pedestrians and

textures. In WACV.

Hirzer, M., Roth, P., Kostinger, M., and Bischof, H.

(2012). Relaxed pairwise learned metric for person

re-identiﬁcation. In ECCV.

Jungling, K., Bodensteiner, C., and Arens, M. (2011). Per-

son re-identiﬁcation in multi-camera networks. In

CVPRW.

Li, W. and Wang, X. (2013). Locally aligned feature trans-

forms across views. In CVPR.

Li, W., Zhao, R., and Wang, X. (2012). Human reidentiﬁ-

cation with transferred metric learning. In ACCV.

Liu, C., Gong, S., Loy, C., and Lin, X. (2012a). Person re-

identiﬁcation: What features are important? In ECCV.

Liu, C., Wang, G., Lin, X., and Li, L. (2012b). Person re-

identiﬁcation by spatial pyramid color representation

and local region matching. IEICE.

Ma, B., Su, Y., and Jurie., F. (2012). Bicov: a novel

image representation for person re-identiﬁcation and

face veriﬁcation. In BMVC.

Mignon, A. and Jurie, F. (2012). Pcca: A new approach for

distance learning from sparse pairwise constraints. In

CVPR.

Mogelmose, A., Bahnsen, C., and Moeslung, T. B. (2013a).

Tri-modal person re-identiﬁcation with RGB, depth

and thermal features. In IEEE WPBVS.

Mogelmose, A., Moeslund, T. B., and Nasrollahi, K.

(2013b). Multimodal person re-identiﬁcation us-

ing RGB-D sensors and a transient identiﬁcation

database. In IWBF.

Pedagadi, S., Orwell, J., Velastin, S., and Boghossian, B.

(2013). Local ﬁsher discriminant analysis for pedes-

trian re-identiﬁcation. In CVPR.

Prosser, B., Zheng, W., Gong, S., Xiang, T., and Mary.,

Q. (2010). Person re-identiﬁcation by support vector

ranking. In BMCV.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

146

Quang, M. H., Bazzani, L., and Murino, V. (2013). A uni-

fying framework for vector-valued manifold regular-

ization and multi-view learning. In ICML.

Taiana, M., Nascimento, J., and Bernardino, A. (2013). An

improved labelling for the INRIA person data set for

pedestrian detection. IbPRIA.

Wang, X., Doretto, G., Sebastian, T., Rittscher, J., and Tu,

P. (2007). Shape and appearance context modeling. In

ICCV.

Wu, Y., Mukunoki, M., Funatomi, T., Minoh, M., and Lao,

S. (2011). Optimizing mean reciprocal rank for person

re-identiﬁcation. In AVSS.

Zheng, W., Gong, S., and Xiang, T. (2009). Associating

groups of people. In BMVC.

Zheng, W., Gong, S., and Xiang, T. (2011). Person re-

identiﬁcation by probabilistic relative distance com-

parison. In CVPR.

TowardsFullyAutomatedPersonRe-identification

147