Evaluating the Robustness of Global Appearance Descriptors in a Visual

Localization Task, under Changing Lighting Conditions

Vicente Rom

an, Luis Pay

a and

Oscar Reinoso

Engineering Systems and Automation Department, Miguel Hernandez University, Elche (Alicante), Spain

Keywords:

Mobile Robots, Global Appearance Descriptors, Omnidirectional Images, Mapping, Localization.

Abstract:

Map building and localization are two important abilities that autonomous mobile robots must develop. This

way, much research has been carried out on these topics, and researchers have proposed many approaches to

address these problems. This work presents some possibilities to solve these problems robustly using global

appearance descriptors. In this way, robots capture visual information from the environment and obtain,

usually by means of a transformation, a global appearance description for each whole image. Using these

descriptors, the robot is able to estimate its position in a previously built map, which is also composed of

a set of global appearance descriptors. In this work, the global appearance descriptors performance at the

localization task inside a known environment has been studied. In the assignment a map is already built

and the goal is to evaluate the descriptors’ robustness to perform localization tasks when the environment

visual appearance changes substantially. To achieve this objective, a comparative evaluation of several global

appearance descriptors is carried out.

1 INTRODUCTION

Over the last few years, the presence of mobile robots

in industrial and domestic environments has increased

substantially. This increase has been facilitated by the

development of their abilities in perception and com-

putation, which allow them to improve the operation

in large and heterogeneous areas without the neces-

sity of introducing changes into the robot or environ-

ment structure. However, the mobile robot boom will

not be deﬁnitive until the level of autonomy and the

adaptability to different conditions enhance substan-

tially. An autonomous robot has to ﬁnd a good solu-

tion to two critical problems: building a model of the

environment (mapping) and estimating the position of

the robot within this model (localization). Both have

to be solved with an acceptable accuracy and compu-

tational cost.

Nowadays, one of the most extended systems to

obtain information from the environment are vision

sensors, which permit extracting local information

from the scenes. The use of local appearance descrip-

tors has been the classical approach for obtaining rel-

evant information from the images. This is a mature

and very developed alternative and many researchers

make use of it in mapping and localization. For ex-

ample Valgren and Lilienthal (Valgren and Lilienthal,

2010) have proposed using these descriptors to per-

form topological localization. Murillo et al. (Murillo

et al., 2007) present a comparative study for the local-

ization task in indoor environments. A comparative

evaluation of this kind of local appearance methods

was made by Gil et al. (Gil et al., 2010). In this paper

they evaluated, the repeatability of these detectors, as

well as the invariance and distinctiveness of the de-

scriptors under different perceptual conditions. More

recently, an alternative has emerged, which consists

in representing each image as a whole, without ex-

tracting local features. This method could offer com-

parative results and it also simpliﬁes the structure of

the map. Additionally, localization can be carried out

with a simple process, based on the pairwise compar-

ison of global descriptors. As a disadvantage, it has

to work a large amount of data, so is necessary to use

a compression technique that minimize the computa-

tional cost.

Some researchers have proposed different meth-

ods to describe the global appearance of the scenes,

which maintain the necessary information to do the

localization and mapping tasks. The discrete Fourier

Transform is one alternative. (Menegatti et al., 2004)

submit the Fourier Signature (FS) concept and deﬁne

a method to carry out mapping and localization. Other

studies support a method based on the Histogram of

258

Román, V., Payá, L. and Reinoso, Ó.

Evaluating the Robustness of Global Appearance Descriptors in a Visual Localization Task, under Changing Lighting Conditions.

DOI: 10.5220/0006837802580265

In Proceedings of the 15th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2018) - Volume 2, pages 258-265

ISBN: 978-989-758-321-6

Oriented Gradients (HOG). For example (Dalal and

Triggs, 2005) use this method to detect pedestrians

with a good performance and (Pay

a et al., 2016) prove

the robustness of the descriptor in mapping and lo-

calization tasks.Furthermore, a method know as gist,

which tries to imitate the human vision system, can be

used. Oliva and Torralba were precursors of this idea

(Oliva and Torralba, 2001). Other authors, like Sia-

gian and Itti, have used this method in their work (Sia-

gian and Itti, 2009) where they have described this bi-

ological inspired vision localization system and have

tested it in different outdoor environments. Finally,

the Radon Transform could be used to compress the

information too, as (Berenguer et al., 2016) show in

their studies .

In our experiments we use omnidirectional im-

ages, which give a 360 range vision so they offer more

information than other visual systems. With this sys-

tem, features are more time in the camera range vision

so it allows to obtain stabler ones. Moreover, omnidi-

rectional systems contains information enough to cre-

ate a model if a person or an object occlude the image

during the route.

In this area, previously works like (Pay

a et al.,

2016) present a comparative evaluation of these

global-appearance techniques in topological mapping

tasks. (Pay

a et al., 2017) present an evaluation of

these descriptors in compacting visual maps using FS,

HOG and gist descriptors. The goal of this paper is to

extend these previous works. On the one hand, we in-

clude a new scene descriptor called Radon transform

and; on the other hand, we compare the methods on a

real and changing environment.

We have studied the global-appearance descrip-

tors strength, verifying their performance when the

environment visual appearance changes substantially.

In real-operating conditions the robot has to cope with

different events: not the same lighting conditions be-

tween the model and the new localization; scenes oc-

cluded by people or other mobile robots and changes

in the scene like the furniture position.

2 REVIEW OF GLOBAL

APPEARANCE DESCRIPTORS

In this section we summarize some alternatives to ex-

tract global information from panoramic images try-

ing to keep relevant data with the minimum memory.

As the images used in this work are given in an om-

nidirectional form, the ﬁrst step is to transform them

into a panoramic ones. This process is fairly slow and

it will affect to the computational cost. The starting

point is a panoramic image i(x,y) ∈ R

×N

and after

using the global appearance method the result is a de-

scriptor

d ∈ R

l×1

The robot has a planar movement and to cap-

ture the images a catadioptric vision system is used

(Pronobis and Caputo, 2009). This system is mounted

vertically on a robot, a camera and a convex mir-

ror were aligned and mounted together on a portable

bracket to build the vision system. The movement of

this robot is restricted to the ﬂoor plane.

2.1 Fourier Signature

(Menegatti et al., 2004) show that when we calcu-

late the Discrete Fourier Transform of each row of

a panoramic image f(x,y), we obtain a new complex

matrix F(u,v) that can be split in two matrices: one

with the modules A(u,v) and other with the arguments

Φ(u,v).

Two images that have been taken from the same

position but with different orientation would have the

same modules’ matrix but a change in the arguments’

matrix is produced, according to the shift property of

the Fourier Transform. The relative rotation between

images could be calculated (Pay

a et al., 2009). So we

can use A(u,v) as a localization descriptor and Φ(u,v)

as an orientation estimator.

Furthermore, in the Fourier domain, the most rele-

vant information is concentrated in the low frequency

components so we can discard a number of high fre-

quency components that are usually affected by noise.

It allows us to minimize the computational cost in the

subsequent comparisons. In this work we use the pa-

rameter k

which indicates the number of frequencies

we retain. N

is the panoramic image number of rows.

The image is reduced into a global appearance de-

scriptor

d ∈ R

·k

×1

2.2 Histogram of Oriented Gradient

As (Dalal and Triggs, 2005) describes, the HOG tech-

nique works with the orientation of the gradient in

speciﬁc areas. Hofmeister et al. use a weighted his-

togram of oriented gradients in small and controlled

environments (Hofmeister et al., 2009).

Essentially it consists in calculating the gradient

of the panoramic image obtaining module and orien-

tation of each pixel to localize mobile robots. Af-

terwards, the image is divided in a set of cells and

the orientation histogram of each cell is calculated,

weighting the each bin with the module of the gra-

dient of each pixel. Although the image is rotated,

two panoramic images contain the same information

in each row, but a shift of the pixels. As a result, if

we divide the image into horizontal cells, a descriptor

Evaluating the Robustness of Global Appearance Descriptors in a Visual Localization Task, under Changing Lighting Conditions

259

will be obtained which is invariant to rotations of the

robot in the ﬂoor plane and can be used as a localiza-

tion descriptor.

Moreover, the whole image can be reduced to a

vector whose size will depend on the number of cells

and bins. In this work we use b

, which is the number

of bins per histogram and k

, which is the number of

horizontal cells in which the image has been divided.

This method transform the panoramic image into a

descriptor

d ∈ R

·k

×1

2.3 Gist

This descriptor was initially described in (Oliva and

Torralba, 2006), and further developed in other stud-

ies like (Siagian and Itti, 2009) where gist descriptor

had been tested in three outdoor environments.

The method offers rotational invariance when it is

applied to panoramic images. The descriptor is built

from orientation information, obtained after exposing

some Gabor ﬁlters with different orientations to the

image in several resolution levels. Finally, the volume

of data is reduced for an average; like when we use

HOG, we divide the image with horizontal cells and

we calculate the mean intensity of each cell.

On this way, we use m

which indicates the num-

ber of orientations of Gabor ﬁlters and k

that desig-

nates the number of horizontal cells in which the im-

age has been split. With these parameters the image

can be reduced to a descriptor whose size depends on

, k

and the number of different resolution models

used, but in the experiments we maintain this parame-

ter constant in 2 levels. So ﬁnally the result is a global

appearance descriptor

d ∈ R

2·m

·k

×1

2.4 Radon Transform

The Radon Transform can also be used to describe

globally an omnidirectional image. This transforma-

tion was described in (Radon, 2005). It was used in

some computer vision activities, such as (Hoang and

Tabbone, 2010) where the Radon transform (RT) was

used as a geometric shape descriptor. In the robotic

ﬁeld we can ﬁnd the study (Berenguer et al., 2016)

where RT was employed to ﬁnd the nearest neighbor

in a position estimation.

Radon method can be used directly with omnidi-

rectional images so they have not to be transformed

into panoramic. It will presume a reduction in the

computational cost. The method consists in apply-

ing the Radon Transform along several sets of straight

lines, with a variety of orientations between consecu-

tive sets. The difference of orientation is considered

a variable parameter named p

(deg.). This way, an

omnidirecctional image data can be transformed to

a vector that fulﬁlls some interesting properties. (a)

It reduces the information; ﬁgure 1 shows an image

× N

which could be reduced into a

360

× 0.5 · N

matrix; (b) invariability of the information if against

changes of the robot orientation and (c) possibility of

calculate the orientation with a column shift.

(a) (b) (c)

Figure 1: Radon transform. (a) omnidireccional image

× N

), (b) Radon transform of a the image (

360

× N

)

and (c) ﬁnal Radon transform (

360

× 0.5 · N

When the Radon transform has been calculated,

the localization process could be addressed with some

methods. In this paper we use a method based on FS.

We apply the Fourier Signature to transform the data

we obtain after we have calculated the Radon Trans-

form. We get two matrices (magnitudes and argu-

ments) and the descriptors are built as described in

section 2.1.

In this case, the descriptor size will depend on p

and the number of columns we retain once we apply

the FS. This parameter is called k

. The descriptor

obtained is

d ∈ R

360

·0.5·N

×1

3 SETS OF IMAGES

In this work, the COLD database (Pronobis and Ca-

puto, 2009) is used. This database provides omnidi-

reccional image sequences taken under three different

lighting conditions: a sunny day, a cloudy one and at

night. The images were acquired across several days

under illumination variations as well as changes intro-

duced by human activity in the environment (people

appearing in the rooms, objects and furniture being

relocated or moved and so on). These different con-

ditions can be seen in ﬁgure 2

Therefore, the COLD database is an ideal test

bench to evaluate the localization and mapping al-

gorithms’ robustness because it offers variations that

might occur in real-world environments. COLD

database offers diverse routes captured in indoor envi-

ronments in several universities. Among those possi-

bilities we use Freiburg dataset and the longest route,

called Part A, Path 2, size 3 (Pronobis and Caputo,

2009). Freiburg route supposes an especial challeng-

ing one because many walls between rooms are made

of glass and there are lots of windows so the extern

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

260

(a) Image from the sunny dataset

(b) Image from the night dataset

Figure 2: Panoramic images from COLD database.

light conditions affect substantially the localization

task.

The sequences of this set of images were recorded

using a mobile robot and both perspective and om-

nidirectional cameras. The catadioptric vision sys-

tem was made using a hyperbolic mirror was mounted

with the cameras on a portable bracket. The robot was

also equipped with a laser range scan and wheels en-

coders to detect the odometry.

4 EXPERIMENTS

In this section we test the global-appearance descrip-

tors effectiveness. First we outline the distances and

measurement error models used, then we present the

localization results considering different illumination

(Cloudy weather, sunny weather and at night) and the

computational cost of these experiments.

4.1 Method to Create a Model of the

Environment

As explained in the previous section, COLD offers

different databases captured in diverse buildings and

conditions. We use Freiburg cloudy dataset to build

the model. COLD database took an image every

0.06m approximately. To build our model we make a

reduction taking one every ﬁve images, so our model

follow the same route but the distance between im-

ages is around 0.30m. This method allows us to have

a model with 556 images instead of 2778 images that

cloudy dataset (Part A, Path 2, size 3) have.

Once we have built the model with known images,

the method to solve the localization task consists in

comparing each image from a test group with the im-

ages in the model. The program calculates the near-

est neighbor. To obtain the nearest neighbor differ-

ent distances can be used. In our study we compare

Table 1: Distances.

Measure Distance Mathematical expression

d1 Cityblock d

(~a,

b) =

∑

i=1

− b

d2 Euclidean d

(~a,

b) =

∑

i=1

· b

)

d3 Correlation d

(~a,

b) = 1 −

d4 Cosine d

(~a,

b) = 1 −

Where:

= [a

− a, ...,a

− a]; a =

∑

·a

= [b

− b, ..., b

− b]; b =

∑

·b

these ones: cityblock, Euclidean, correlation and co-

sine. Table 1 shows the deﬁnition of each distance.

At the mathematical expression,~a ∈ R

lx1

and

b ∈ R

lx1

are two vectors where: a

,i = 1, ...,l.

Once the nearest neighbor is obtained we calculate

the geometric distance between the test image and his

neighbor, and the result will be the error. This real

distance can be calculated because COLD database

offers the coordinates where each image had been

captured but the coordinates have been only used as

ground truth to check the error. The localization task

is carried out with pure visual information.

The experiments have been done using different

descriptors and varying the parameters that could be

modiﬁed in order to evaluate the inﬂuence of each pa-

rameter in the task. These parameters can be seen in

table 2. As shown in section 2, therse parameters de-

ﬁne the descriptor size. The larger the descriptor is,

the slower the process will be.

Table 2: Parameters that impact on the location process.

Descriptor Parameters

FS k

⇒ number of retained columns.

HOG b

⇒ number of bins per histogram.

⇒ number of horizontal cells.

gist m

⇒ number of Gabor ﬁlters.

⇒ number of horizontal blocks.

Radon p

⇒ degrees between sets of lines.

⇒ number of retained columns.

4.2 Position Estimation

Initially, the cloudy images are used to create the ref-

erence model. Afterwards, to study the robustness

of the global-appearance descriptors, the test images

used to solve the localization problem will be chosen

from the sunny and night datasets, which are com-

posed of 2231 and 2896 images, respectively. The

localization process evaluates which image in the ref-

erence model is the most similar to each test image.

The error is calculated as the geometric distance be-

tween the capture points of both images. After re-

Evaluating the Robustness of Global Appearance Descriptors in a Visual Localization Task, under Changing Lighting Conditions

261

peating this process considering all the sunny images

as test image, ﬁgures 3, 4, 5 and 6 show the average

error (m) with each descriptor method.

Figure 3: FS. Localization error (m) using test images from

the sunny dataset versus k

Figure 4: HOG. Localization error (m) using test images

from the sunny dataset versus k

and b

The behaviour of the FS changes depending on the

distance measure used, although d2 and d4 offer sim-

ilar results. The error varies between 3 and 7 meters.

The higher k

is, the better results we obtain but the

process will be slower. We obtain the best accuracy

with d3 and k

=128 where the location error is 3,56m;

when k

=32 the error is 3,91m.

About HOG, a medium-high value of k

and an

intermediate value of b

offer the best results. The

lowest error is obtained using distance d1 and b

=32

and k

=8 and it is 1,09m.

In general, using the gist descriptor, the error is

about 2m. The error decreases with medium values of

and medium-low values of m

. The best solution

is detected with d3 and k

=16 where the error is

1,69m.

At last, in the case of RT, the results are in general

worse. The error is higher than 5m in most cases. The

Figure 5: gist. Localization error (m) using test images from

the sunny dataset versus k

and m

Figure 6: Radon. Localization error (m) using test images

from the sunny dataset versus p

and k

error depends more on the k

parameter and the higher

it is, the lower the error is. The lowest error is detected

when k

=32 and p

=1. With these parameters and d1

the error is 4,93m.

Analysing globally these ﬁgures, HOG and gist

are the description algorithms with a lower error that

varies between 1 an 2 m when they use d1 and d3.

There are remarkably good results taking into account

that not only the illumination changes, but also furni-

ture movements and occlusions appear in the scenes.

After this experiment, the localization process is

repeated considering all the images in the night set as

test images. Like in the previous case, the reference

model is preciously created with cloudy images. Fig-

ures 7, 8, 9 and 10 present the results showing the

error in meters.

The results using FS are between 0,6 and 2,5m.

Low values of k

offer lower errors, obtaining the

best one with k

=4 and distance d1 where the error

is 0,597m.

The behaviour with HOG presents an error be-

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

262

tween 0,2 and 0,6m . The best results are given with

intermediate values of k

and b

and the distance d1.

With k

=32, b

=16 and distance d1 we obtain the low-

est error, 0,189m.

About gist, the mean error is between 0,2 and 1m.

If k

or m

values are low, the error will increase and

it will not decrease signiﬁcantly when the parameters

are high. In fact it will offer relatively bad results

if m

increases a lot; so medium values are the best

choice. The lowest error has been obtained with the

distance d1, k

=32 and b

=8, when the error is 0,21m.

The maximum error in our night proofs using gist has

been 1,37m. It has been obtained with d3, k

=2 and

=256. This is a relatively low error if it is compared

with the maximum error using other descriptors.

Finally, the localization error with Radon is be-

tween 0,4 and 1,8m. The lower k

is, the lower the

error is; p

has little inﬂuence in the results. When

=32, p

=1 and d3 the localization error is 0,3281m.

Figure 7: FS. Localization error (m) using test images from

the night dataset versus k

Figure 8: HOG. Localization error (m) using test images

from the night dataset versus k

and b

It is remarkable that we obtain better results when

the localization task is made at night than when it was

made on a sunny day. This could be cause because of

Figure 9: Gist. Localization error (m) using test images

from the night dataset versus k

and m

Figure 10: Radon. Localization error (m) using test images

from the night dataset versus p

and k

the large amount of walls made of by glass in Freiburg

database; and, as we work with different lighting con-

ditions, it could cause anomalies, such as brights and

reﬂections. But taking into account the great chal-

lenge that it suppose, the localization presents notably

good results in both states.

4.3 Computational Cost

Besides having a low error, it is important to have a

reasonable computation time. We have also studied

the computational time the process spends in order to

know if the localization can be made in real time. Fig-

ures 11, 12, 13 and 14 show the runtime needed to es-

timate the robot position with each descriptor method

and distance measure. The data are given in seconds.

FS offers the quickest results. The highest time is

0,221s when k

=32 and the distance d3. The time has

an exponential growth when the variable k

increases.

Runtimes with HOG are similar to FS ones. The

optimal result is 0.235s. When the parameters in-

Evaluating the Robustness of Global Appearance Descriptors in a Visual Localization Task, under Changing Lighting Conditions

263

crease, the runtime increases too. b

has more inﬂu-

ence so it is important not to have high values of it.

However, k

has less importance in the runtime.

Gist does not offer a quick runtime if it is com-

pared with other descriptors. Using gist the runtime

could be 0,537s when m

=64 and k

=32. In this case,

the parameter m

has a lot of importance in the time.

The quickest runtime with gist is given when m

and k

=4, whose associated time is 0.2405s.

Finally, we have studied the RT runtime. We can

observe high dependence on p

and less importance

of k

. The lower p

, is the higher the runtime will be.

The highest time is 0,447s with p

=1 and k

=16 but

Radon could offer relatively reduced times like 0.059s

when p

=8 and k

=4.

The time spent while the program had to change

from omnidireccional images to panoramic ones is

the time which has more relevance in the results ob-

tained. So the compression process and the compar-

ison between descriptors have less importance. As

a result the experiments give us similar runtimes be-

tween Fourier and HOG and a slow runtime with gist.

The RT process does not have to change the images to

panoramic pictures so it reduces the cost of the pro-

cess, but when p

is low, its compression process is

slower and it has signiﬁcant relevance in the RT pro-

cess. The experiments have been carried out in a PC

with a CPU Quad-Core Intel i7-700

 at 3.60 GHz

and through Matlab

 programming.

These results are not absolute, they will depend

of the machine which runs the process. But they

are comparable because all the calculations have been

done with the same computer.

Figure 11: FS. Localization time (s) per image versus k

5 CONCLUSIONS

This work has focused on the study of different prob-

lems in navigation tasks under illumination variations

Figure 12: HOG. Localization time (s) per image versus k

and b

Figure 13: gist. Localization time (s) per image versus k

and m

Figure 14: Radon. Localization time (s) per image versus

and k

and changes introduced by human activity, and the

robot is only equipped with an omnidireccional cam-

era. We have compared four appearance-based algo-

rithms applied to this localization task. When the im-

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

264

ages have been described and the localization error

has been calculated a comparative evaluation has been

carried out and the computational cost has also been

studied.

On the one hand, the Fourier Signature and HOG

offer quicker results, so they constitute the most suit-

able option to do tasks in real time. On the other hand,

HOG and gist descriptors provide better accuracy in

localization tasks.

This work opens the door to new applications of

the appearance-based methods in mobile robots. As

we have shown, the appearance-based descriptors are

a suitable method to do navigation tasks. A prob-

lem is the high requirements of memory and compu-

tation time to do the database and make the neces-

sary calculations to compute the position. Once we

have tested the methods’ robustness under human ac-

tivity and changes in illumination and in the position

of some objects, the next step should be to create a

method that continuously renews the database adapt-

ing it to new lighting conditions. It can also evolve

to a system that creates more sophisticated maps to

make it possible an autonomous navigation system.

ACKNOWLEDGEMENTS

This work has been supported by the Spanish Gov-

ernment through the project DPI 2016-78361-R

(AEI/FEDER, UE): “Creaci

on de mapas mediante

etodos de apariencia visual para la navegaci

on de

robots”.

REFERENCES

Berenguer, Y., Pay

a, L., Peidr

o, A., Gil, A., and Reinoso,

O. (2016). Nearest position estimation using omni-

directional images and global appearance descriptors.

In Robot 2015: Second Iberian Robotics Conference,

pages 517–529. Springer.

Dalal, N. and Triggs, B. (2005). Histograms of oriented gra-

dients for human detection. In 2005 IEEE Computer

Society Conference on Computer Vision and Pattern

Recognition (CVPR’05), volume 1, pages 886–893

vol. 1.

Gil, A., Mozos, O. M., Ballesta, M., and Reinoso, O.

(2010). A comparative evaluation of interest point de-

tectors and local descriptors for visual slam. Machine

Vision and Applications, 21(6):905–920.

Hoang, T. V. and Tabbone, S. (2010). A geometric invari-

ant shape descriptor based on the radon, fourier, and

mellin transforms. In Pattern Recognition (ICPR),

2010 20th International Conference on, pages 2085–

2088. IEEE.

Hofmeister, M., Liebsch, M., and Zell, A. (2009). Vi-

sual self-localization for small mobile robots with

weighted gradient orientation histograms. In 40th In-

ternational Symposium on Robotics (ISR), pages 87–

91. Barcelona.

Menegatti, E., Maeda, T., and Ishiguro, H. (2004). Image-

based memory for robot navigation using properties

of omnidirectional images. Robotics and Autonomous

Systems, 47(4):251 – 267.

Murillo, A. C., Guerrero, J. J., and Sagues, C. (2007). Surf

features for efﬁcient robot localization with omnidi-

rectional images. In Robotics and Automation, 2007

IEEE International Conference on, pages 3901–3907.

IEEE.

Oliva, A. and Torralba, A. (2001). Modeling the shape

of the scene: A holistic representation of the spatial

envelope. International journal of computer vision,

42(3):145–175.

Oliva, A. and Torralba, A. (2006). Building the gist of a

scene: The role of global image features in recogni-

tion. Progress in brain research, 155:23–36.

Pay

a, L., Fern

andez, L., Reinoso,

O., Gil, A., and

Ubeda, D. (2009). Appearance-based dense maps

creation-comparison of compression techniques with

panoramic images. In ICINCO-RA, pages 250–255.

Pay

a, L., Mayol, W., Cebollada, S., and Reinoso, O. (2017).

Compression of topological models and localization

using the global appearance of visual information. In

Robotics and Automation (ICRA), 2017 IEEE Interna-

tional Conference on, pages 5630–5637. IEEE.

Pay

a, L., Reinoso, O., Berenguer, Y., and

Ubeda, D. (2016).

Using omnidirectional vision to create a model of

the environment: A comparative evaluation of global-

appearance descriptors. Journal of Sensors, 2016.

Pronobis, A. and Caputo, B. (2009). COLD: COsy Lo-

calization Database. The International Journal of

Robotics Research (IJRR), 28(5):588–594.

Radon, J. (2005). 1.1

uber die bestimmung von funktio-

nen durch ihre integralwerte l

angs gewisser mannig-

faltigkeiten. Classic papers in modern diagnostic ra-

diology, 5:21.

Siagian, C. and Itti, L. (2009). Biologically inspired mo-

bile robot vision localization. IEEE Transactions on

Robotics, 25(4):861–873.

Valgren, C. and Lilienthal, A. J. (2010). Sift, surf & sea-

sons: Appearance-based long-term localization in out-

door environments. Robotics and Autonomous Sys-

tems, 58(2):149–156.

Evaluating the Robustness of Global Appearance Descriptors in a Visual Localization Task, under Changing Lighting Conditions

265