Evaluation of a Local Descriptor for HDR Images

Artur Santos Nascimento

, Welerson Augusto Lino de Jesus Melo

Beatriz Trinch

ao Andrade

and Daniel Oliveira Dantas

Departamento de Computac¸

ao, Universidade Federal de Sergipe, S

ao Crist

ao, SE, Brazil

Keywords:

High Dynamic Range Images, Feature Point Detection, Feature Point Description.

Abstract:

Feature point (FP) detection and description are processes that detect and extract characteristics from images.

Several computer vision applications rely on the usage of FPs. Most FP descriptors are designed to support low

dynamic range (LDR) images as input. However, high dynamic range (HDR) images can show details in bright

and shadowed areas that LDR images can not. For that reason, the interest in HDR imagery as input in the

detection and description processes has been increasing. Previous studies have explored FP detectors in HDR

images. However, none have presented FP descriptors designed for HDR images. This study compares the FP

matching performance of description vectors generated from LDR and HDR images. The FPs were detected

and described using a version of the SIFT algorithm adapted to support HDR images. The FP matching

performance of the algorithm was evaluated with the mAP metric. In all cases, using HDR images increased

the mAP values when compared to LDR images.

1 INTRODUCTION

In computer vision, characteristics are image regions

with properties such as lines, borders, or high con-

trast. These characteristics are extracted in two steps:

detection and description of feature points (FP). Ap-

plications such as object recognition, scene recon-

struction, and biometric systems rely on FP detection

and description (Andrade et al., 2012; Schmid et al.,

2000; Se et al., 2002).

Detection algorithms search for image FPs that

can be found even when the image undergoes geo-

metric or photometric transformations. Based on the

image and its FPs, description algorithms extract dis-

criminative invariant signatures from these FPs that

can identify and distinguish a given FP from an-

other (Rana et al., 2019). If a description is good

enough, it can identify the same FP in a different cap-

ture of the same scene.

Most detection and description algorithms receive

low dynamic range (LDR) images as input. An LDR

image usually has 8 bits per sample. Thus, the color

intensity values are limited to integers in the range

https://orcid.org/0000-0003-0419-6170

https://orcid.org/0000-0003-0644-2427

https://orcid.org/0000-0002-1407-8250

https://orcid.org/0000-0002-0142-891x

[0, 255]. For this reason, scenes with substantial light-

ing variation can result in underexposed and overex-

posed areas. In these cases, detectors and descriptors

may fail as details are hidden by bright or shadowed

areas.

On the other hand, HDR images use more than 8

bits per sample, allowing a higher dynamic range and

greater color accuracy in the overexposed and under-

exposed areas. Knowing that feature extraction from

a scene is highly dependent on the scene lighting at

capture time, this study hypothesizes that the FP de-

scription will detect more FPs, especially in bright or

shadowed areas.

Previous studies have evaluated FP detection and

shown that FP detection performance improves when

using HDR images as input (Melo et al., 2018; Os-

tiak, 2006). Therefore, detection algorithms adapted

for HDR images were created, such as Harris Cor-

ner for HDR and DoG for HDR (Melo et al., 2018).

However, few studies have evaluated FP descriptors

that use HDR images as input.

In this context, this study evaluates and compares

the FP description and FP matching performance in

HDR and LDR images. Using the mean average pre-

cision (mAP) metric, two datasets, and an adapted

version of the SIFT algorithm, we observed that using

HDR images as input increased the mAP by 81.80%

(from 0.22 to 0.40) with the ﬁrst dataset (3D lighting),

Nascimento, A., Melo, W., Andrade, B. and Dantas, D.

Evaluation of a Local Descriptor for HDR Images.

DOI: 10.5220/0010779700003124

In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 4: VISAPP, pages

299-306

ISBN: 978-989-758-555-5; ISSN: 2184-4321

299

and by 61.50% (from 0.39 to 0.63) with the second

one (2D lighting).

This paper is organized as follows. Section 2 de-

scribes the related works. Section 3 brieﬂy describes

the SIFT algorithm. Section 4 describes the dataset

used, the evaluation metric, and the proposed method-

ology. Section 5 shows the experimental results. The

conclusions are drawn in Section 6.

2 RELATED WORKS

The number of studies using HDR and tone-mapped

images as input for detection and description algo-

rithms is signiﬁcantly smaller than those using LDR

images as input. In this section, we classify stud-

ies that use HDR and tone-mapped images into four

groups: studies that propose datasets with complex

lighting conﬁgurations and that offer captures in both

LDR and HDR formats; studies focused on FP de-

tection algorithms; studies focused on FP description

algorithms; and studies that adapt FP detection or de-

scription algorithms for speciﬁc applications.

2.1 Datasets

ribyl (P

ribyl et al., 2012) studied the use of tone

mapping (TM) in HDR images to improve FP detec-

tion. The main idea is to capture more details in re-

gions that conventional LDR images can not correctly

represent due to the lack of details in bright and shad-

owed areas of the scene. There is no need to change

the detection algorithms in this approach as the TM

algorithm maps the tones from the HDR image into

an LDR image.

With that in mind, P

ribyl generated a dataset

whose images have abrupt lighting changes. Two

scenes were captured: a planar scene (2D) contain-

ing three posters in A4 sheet format; and a 3D scene

containing several non-planar rigid objects. Both

scenes were placed in a dark room with controlled

lighting. The scenes were captured in three different

sequences, consisting of changing viewpoints, dis-

tances, and lighting.

• Viewpoint Changing Sequence: the camera is

moved following a circular trajectory with its cen-

ter in the scene with a step of 2.5

◦

. Thereby, 21

captures were made, resulting in a 50

◦

total tra-

jectory;

• Distance Changing Sequence: the scene was

captured seven times and the distance between

the camera and the scene increased progressively,

yielding the distance sequence of 100 cm, 103 cm,

109 cm, 122 cm, 147 cm, 197 cm, and 297 cm;

• Lighting Changing Sequence: the scene was

also captured seven times, each time with a dif-

ferent combination of three light sources being on

or off, with at least one of them on.

They conclude that images generated from local

TM algorithms give better results than images from

global TM algorithms when used as input to the fea-

ture detection process (P

ribyl et al., 2012).

In order to produce a series of studies exploring

the use of TM in HDR images to improve the detec-

tion and description processes, Rana et al. (Rana et al.,

2015) proposed a new dataset. They also evaluated

TM algorithms not considered by P

ribyl et al. (P

ribyl

et al., 2012).

The dataset proposed by Rana et al. consists of

two capture sequences, named project room sequence

(PRS) and light room sequence (LRS). PRS is com-

posed of eight different lighting conﬁgurations cre-

ated by blocking light coming from a projector with

the help of different objects. The scene is composed

of several bright and dark-colored objects arranged

to create sharp shadows and bright areas. LRS com-

prises seven different natural lighting conditions cre-

ated by changes in global lighting through the open-

ing and closing of window blinds, room ambient illu-

mination, and diffuse lighting from a tungsten lamp.

In this scene conﬁguration, there are also dark and

light objects with different types of surfaces.

2.2 FP Detection Algorithms

ribyl et al. (P

ribyl et al., 2016) use the dataset pro-

posed in their previous studies (P

ribyl et al., 2012)

to expand their evaluation of the impact of TM dur-

ing FP detection. They include more TM algorithms

and the DoG detection algorithm in their experiments.

The conclusion is that the TM algorithms proposed

by Fattal et al. (Fattal et al., 2002) and Mantiuk et

al. (Mantiuk et al., 2006) detect more FPs than other

ones.

Two studies by Rana et al. focus on FP detec-

tion. In the ﬁrst study (Rana et al., 2016b), Rana

et al. propose a TM operator to optimize FP detec-

tion. As a result, using their TM operator improves

the correlation coefﬁcient (CC) and repeatability rate

(RR) compared to other TM operators. In the second

study (Rana et al., 2017a), Rana et al. develop a new

adaptive local TM operator that uses support vector

regression (SVR) to predict optimal modulation maps

to improve FP detection. As a result, the TM operator

by Rana et al. showed better RR when compared to

other TM operators.

Melo et al. propose a modiﬁcation for FP detec-

tors in order to improve FP detection on HDR images.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

300

A local mask based on the coefﬁcient of variation

(CoV) of sets of pixels is proposed as an additional

step in FP detection. Also, the Uniformity metric is

introduced as a new criterion to evaluate FP detection

in areas under different lighting intensities. The ob-

tained results show better uniformity and repeatabil-

ity rate in most tested HDR images when compared

to standard FP detectors. Moreover, they indicate that

HDR images have great potential to be explored in

applications that rely on FP detection (Melo et al.,

2018).

2.3 FP Description Algorithms

There are several studies from Rana et al. (Rana et al.,

2016a; Rana et al., 2017b; Rana et al., 2019) fo-

cused on image description based on HDR images.

The ﬁrst one (Rana et al., 2016a) compared image

matching using TM in HDR images and LDR images

using SIFT (Lowe, 2004), SURF (Bay et al., 2008),

FREAK (Alahi et al., 2012), and BRISK (Leuteneg-

ger et al., 2011) to describe the FPs. By using sev-

eral global and local TM algorithms, it is observed

that all combinations that used tone-mapped images

as input performed better than with LDR images. In

their second study (Rana et al., 2017b), Rana et al.

proposed an adaptive TM operator that uses SVR to

predict optimal modulation maps to improve FPs de-

scription, speciﬁcally to make image matching invari-

ant to day or night scene illumination. Finally, in the

third study (Rana et al., 2019), Rana et al. used the

adaptive TM operator to improve FP description in

image matching.

Khwildi and Zaid (Khwildi and Zaid, 2018) in-

troduce a descriptor based on LDR image expansion.

In this approach, the LDR image is converted to an

HDR image using a reverse TM operator (TM-HDR

image). Then, the resulting TM-HDR image is tone

mapped back to LDR image (TM-LDR) and SIFT de-

scriptor is used to describe FPs. As a result, it is

demonstrated that features described from the TM-

LDR image are more descriptive than those extracted

from LDR and TM-HDR images. In future works,

they consider exploring local TMs and improving the

efﬁciency of local descriptors directly in HDR im-

ages.

2.4 Applications

Ige et al. (Ige et al., 2016) developed a facial expres-

sion recognition algorithm using support vector ma-

chines (SVMs) and local binary patterns (LBP). First,

a TM algorithm is used to convert HDR into LDR

images. Then, the resulting tone-mapped image is

used as input to the SURF algorithm. Finally, the

tone-mapped images and the LDR images are com-

pared. As a result, the approach using tone-mapped

images showed better results than LDR images. Tone-

mapped images reach 79.8% accuracy, while the tra-

ditional methods that use LDR images range between

31.3% and 70.8% accuracy.

Ostiak et al. (Ostiak, 2006) used HDR images

to execute an image stitching process to generate a

panorama. In their study, a tone-mapped panorama is

generated to make the shadowed and bright areas of

the image visible. They mention a modiﬁcation of the

SIFT algorithm to describe FPs in HDR images with-

out giving further details. The discussion is based on

a visual analysis of the generated panoramas, and the

results are subjective. The algorithm presents a better

performance in static scenes than in dynamic scenes,

when stitching shadowed and bright areas that are not

visible in LDR images.

The related works show that FP detection and de-

scription using HDR images as input is a ﬁeld yet

to be explored. Most HDR-related studies use tone-

mapped images generated from HDR. Tone-mapped

images bring more details in shadowed and bright ar-

eas than LDR images and there is no need to adjust

detection and description algorithms.

On the other hand, few studies using HDR im-

ages as input were found and just Melo et al. (Melo

et al., 2018) detailed the modiﬁcations made in their

FP detection algorithm to support HDR images. Us-

ing HDR images brings more information but requires

adaptations in detection and description algorithms to

support ﬂoating-point values used to represent pixels.

All studies showed improvements when using

tone-mapped images as input to the detection and de-

scription algorithms. Speciﬁcally, Melo et al. (Melo

et al., 2018) showed that using HDR images as input

resulted in better detection performance in terms of

repeatability rate and uniformity, especially in shad-

owed areas of the image.

3 SIFT

The Scale Invariant Feature Transform (SIFT) de-

tects and describes FPs from an image. SIFT is one

of the most popular FP extractors, invariant to scale

and orientation (Rey Otero and Delbracio, 2014), and

shows good performance in detection and extraction,

as shown in various studies (P

ribyl et al., 2016; Rana

et al., 2016a; Rana et al., 2017a; Rana et al., 2017b;

Rana et al., 2019).

To detect the FPs, SIFT creates an image represen-

tation from the input image called scale-space, which

Evaluation of a Local Descriptor for HDR Images

301

consists of a collection of increasingly blurred im-

ages. Each FP consists of a blob-like structure whose

coordinates (x, y) and characteristic scale σ are lo-

cated with subpixel accuracy. SIFT calculates the

dominant orientation Θ of each FP detected. In this

way, the tuple (x, y, σ, Θ) deﬁnes the center, size, and

orientation of a normalized patch where the descrip-

tion vector is computed. The SIFT detector is known

as Difference of Gaussian (DoG).

The SIFT description vector has 128 dimensions

and is built from a histogram of local gradient direc-

tions around the FP. The local neighborhood size is

normalized to obtain scale-invariance in the descrip-

tion. Next, a dominant orientation in the FP neigh-

borhood is calculated and used to orient the grid over

which the histogram is determined to make the de-

scription rotation-invariant. The SIFT FP description

vector is theoretically invariant to image translation,

rotation and scale changes (Lowe, 2004; Rey Otero

and Delbracio, 2014).

4 METHODS

To evaluate and compare FP description when using

LDR and HDR images, we chose datasets that pro-

vide scenes with complex lighting and scene captures

(i.e. images with bright and shadowed areas) in both

LDR and HDR formats (Section 4.1). We performed

the FP detection and the description of the images in

the LDR and HDR datasets using an adapted version

of SIFT (Section 4.2). Finally, we matched FPs from

pairs of images and evaluated the matching perfor-

mance (Section 4.3). Figure 1 illustrates the pipeline

of the experiment.

4.1 Dataset

The datasets by Rana et al. (Rana et al., 2015) and

ribyl (P

ribyl et al., 2012) were the only datasets

found that explore complex lighting conﬁgurations

and are publicly available (Section 2.1). Those

datasets are composed of a series of captures, each

one made with a different shutter speed. The faster

is the shutter, the darker the obtained photo is, as

the camera sensor is less stimulated. The algorithm

by Debevec and Malik (Debevec and Malik, 2008)

uses this collection of captures with different shutter

speeds to generate the HDR image of the scene. In

each collection, the LDR sample image is acquired

with automatic exposure.

We chose to use the dataset by P

ribyl et al. (P

ribyl

et al., 2012), which consists of two scenes: a planar

scene (2D) containing three posters in A4 sheet for-

mat; and a volumetric (3D) scene containing several

non-planar rigid objects. Both scenes were placed in

a dark room with controlled lighting. As we are in-

terested in investigating the effect of lighting changes

in images with different dynamic ranges, in this study,

we use the lighting changing sequence of both 2D and

3D datasets by P

ribyl (P

ribyl et al., 2012).

4.2 FP Detection and Description

We used as FP detector the SIFT detection step,

known as DoG. Minor changes were made in DoG

to support 32-bit ﬂoating point pixel values. In

this study, we adopt the approach used by Rana et

al. (Rana et al., 2015) and Melo et al. (Melo et al.,

2018), i.e., to select the 500 FPs with strongest re-

sponses of each detector.

To describe the FPs, the SIFT description step

was adapted to receive both LDR and HDR images.

HDR images were converted to a common interval of

[0.0, 255.0], using ﬂoating-point instead of unsigned

char, and the description of the best 500 FPs was gen-

erated. The implementation is available at GitHub

Dataset

DetectionDescriptionEvaluation

Adapted SIFT detector (DoG)

Adapted SIFT descriptor

Matching and evaluation

LDR

images

HDR

images

Figure 1: Flowchart representing the executed pipeline.

https://github.com/welersonMelo/TCC UFS/

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

302

4.3 Matching and Evaluation

Our aim is to evaluate if the description using HDR

images is better than the description generated using

conventional LDR images. The most common way to

do that is by matching description vectors from differ-

ent captures of the same scene. To do that, the 500 FPs

with strongest responses from each detector were se-

lected. This approach was used in other studies (Rana

et al., 2016a; Melo et al., 2018).

To evaluate the description matching, we com-

puted a precision-recall (PR) curve (Mikolajczyk and

Schmid, 2005; Rana et al., 2016a). The PR curve is

based on the true and false matches between FPs from

a pair of images. The description of a FP is usually

represented by a vector that encapsulate its character-

istics (description vector). Given a description vector,

we have a match if the nearest neighbor distance ra-

tio (NNDR) criteria is satisﬁed. In this criteria, a good

match occurs when the ratio between its distance from

the ﬁrst best match (d

BM1

) and the second best match

BM2

) is less than a given threshold th (Equation 1).

The threshold is deﬁned by testing the value that best

ﬁts the dataset used (Lowe, 2004). We use 0.7 as

threshold and the Euclidean distance to compute the

distance between two description vectors.

NNDR =

BM1

BM2

< th (1)

Two description vectors are considered to have a

true positive match if they correspond to two FPs that

are repeated in the reference and test images. A test

FP is considered repeated if it lies in a circle of ra-

dius r centered on the projection of the reference FP

onto the test image. In our case, we considered r = 5.

Analogously, a match is considered a false positive if

the corresponding FPs are not repeated.

To compute the PR curve, we need the recall and

precision values. Recall is deﬁned by Equation 3 and

precision by Equation 2, where tp, fp, and fn refer to

numbers of true positives, false positives, and false

negatives respectively. By varying the threshold th,

we draw a PR curve and measure the area under this

curve. The value of the area under the curve is also

known as average precision (AP). The mean of APs

for all possible image pairs from a dataset is called

mean average precision (mAP) (Rana et al., 2016a).

P =

tp + fp

(2)

R =

tp + fn

(3)

5 RESULTS

The ﬁrst step was to detect the FPs of the LDR and

HDR databases using the DoG detector. Then, the

description vectors of the FPs was calculated using

the adapted SIFT descriptor. For each dataset, FPs

from every possible pair of images were matched. As

each dataset has seven captures, 21 pairs of images

were considered for each dataset.

Afterwards, we obtained the PR curve and AP val-

ues of each pair of images. Table 1 shows the AP val-

ues of the DoG detector applied to LDR and HDR im-

ages from the 2D lighting dataset. Table 2 shows the

AP values of the adapted SIFT algorithm applied to

LDR and HDR images from the 3D lighting dataset.

With the AP values in hand, we calculated the

mAP of the FP matching. The last lines of Tables 1

and 2 show the calculated mAP values from descrip-

tion vectors generated by the adapted SIFT algorithm.

The mAP values are in the interval [0.0, 1.0]. Higher

values are better. Matching results show an improve-

ment of 61.50% in 2D lighting dataset and 81.80% in

3D lighting dataset when using HDR instead of LDR

images.

Figures 2, 3, 4 and 5 show examples of FP match-

ing using LDR and HDR images respectively from the

2D lighting dataset. Green lines are correct matches

and red lines are incorrect matches.

6 CONCLUSIONS

In this study, we evaluated the use of HDR images as

input to an FP extraction pipeline. Although previous

studies have explored FP detectors in HDR images,

none have presented FP descriptors designed for HDR

images. This study compares the FP matching per-

formance of description vectors generated from LDR

and HDR images. The code of the proposed SIFT de-

tector and descriptor are available at GitHub. Using

ribyl et al. dataset (P

ribyl et al., 2012), the SIFT

algorithm adapted to support HDR images, and the

mAP metric, we evaluated if the usage of HDR im-

ages improves the performance of FP matching.

Using HDR images to describe FPs signiﬁcantly

improved the FP matching. The adapted SIFT al-

gorithm, when applied to HDR instead of LDR im-

ages, increased the mAP by 61.50% in the 2D light-

ing dataset and by 81.80% in the 3D lighting dataset.

This advocates that using HDR images can improve

description performance.

The usage of HDR images as input in FP descrip-

tion algorithms is an area yet to be explored, as there

are few studies about this topic in the literature. Fu-

Evaluation of a Local Descriptor for HDR Images

303

Table 1: Average precision in 2D lighting dataset using pairs of LDR and HDR images.

2D lighting LDR average Precision 2D lighting HDR average precision

Image Label 2 3 4 5 6 7 2 3 4 5 6 7

1 0.38 0.14 0.39 0.74 0.53 0.72 0.80 0.21 0.75 0.88 0.88 0.87

2 – 0.10 0.16 0.43 0.60 0.41 – 0.19 0.65 0.73 0.88 0.77

3 – – 0.39 0.29 0.12 0.28 – – 0.31 0.24 0.36 0.25

4 – – – 0.40 0.22 0.38 – – – 0.75 0.70 0.66

5 – – – – 0.42 0.72 – – – – 0.82 0.89

6 – – – – – 0.45 – – – – – 0.79

mAP 0.39 0.63

Table 2: Average precision in 3D lighting dataset using pairs of LDR and HDR images.

3D lighting LDR average Precision 3D lighting HDR average precision

Image Label 2 3 4 5 6 7 2 3 4 5 6 7

1 0.16 0.17 0.04 0.07 0.18 0.25 0.58 0.67 0.03 0.16 0.55 0.68

2 – 0.48 0.03 0.05 0.58 0.53 – 0.84 0.03 0.17 0.85 0.80

3 – – 0.01 0.04 0.59 0.75 – – 0.03 0.17 0.77 0.87

4 – – – 0.06 0.04 0.08 – – – 0.03 0.12 0.08

5 – – – – 0.06 0.02 – – – – 0.10 0.12

6 – – – – – 0.61 – – – – – 0.80

mAP 0.22 0.40

Figure 2: Example of FP matching using 2D lighting dataset and LDR images.

Figure 3: Example of FP matching using 2D lighting dataset and HDR images.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

304

Figure 4: Example of FP matching using 3D lighting dataset

and LDR images.

Figure 5: Example of FP matching using 3D lighting dataset

and HDR images.

ture works may include testing and evaluating the

performance of alternative detectors and descriptors

applied to HDR images and comparing the adapted

SIFT algorithm with methods that use tone mapping.

Other metrics may also be used, such as repeatability

rate and uniformity.

REFERENCES

Alahi, A., Ortiz, R., and Vandergheynst, P. (2012). Freak:

Fast retina keypoint. In 2012 IEEE Conference on

Computer Vision and Pattern Recognition, pages 510–

517.

Andrade, B. T., Mendes, C. M., de Oliveira Santos Jr, J.,

Bellon, O. R. P., and Silva, L. (2012). 3d preserv-

ing xviii century barroque masterpiece: Challenges

and results on the digital preservation of Aleijadinho’s

sculpture of the prophet joel. Journal of Cultural Her-

itage, 13(2):210–214.

Bay, H., Ess, A., Tuytelaars, T., and Van Gool, L. (2008).

Speeded-up robust features (surf). Computer Vision

and Image Understanding, 110(3):346–359. Similar-

ity Matching in Computer Vision and Multimedia.

Debevec, P. E. and Malik, J. (2008). Recovering high dy-

namic range radiance maps from photographs. In

ACM SIGGRAPH 2008 Classes, SIGGRAPH ’08,

New York, NY, USA. Association for Computing Ma-

chinery.

Fattal, R., Lischinski, D., and Werman, M. (2002). Gradi-

ent domain high dynamic range compression. In Pro-

ceedings of the 29th annual conference on Computer

graphics and interactive techniques, pages 249–256.

Ige, E. O., Debattista, K., and Chalmers, A. (2016). To-

wards HDR based facial expression recognition under

complex lighting. In Proceedings of the 33rd Com-

puter Graphics International, CGI ’16, page 49–52,

New York, NY, USA. Association for Computing Ma-

chinery.

Khwildi, R. and Zaid, A. O. (2018). A new retrieval system

based on low dynamic range expansion and sift de-

scriptor. In 2018 IEEE 20th International Workshop

on Multimedia Signal Processing (MMSP), pages 1–6.

IEEE.

Leutenegger, S., Chli, M., and Siegwart, R. Y. (2011).

Brisk: Binary robust invariant scalable keypoints. In

2011 International Conference on Computer Vision,

pages 2548–2555.

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. International journal of computer

vision, 60(2):91–110.

Mantiuk, R., Myszkowski, K., and Seidel, H.-P. (2006). A

perceptual framework for contrast processing of high

dynamic range images. ACM Transactions on Applied

Perception (TAP), 3(3):286–308.

Melo, W. A. L. J., de Tavares, J. A. O., Dantas, D. O.,

and Andrade, B. T. (2018). Improving feature point

detection in high dynamic range images. In 2018

IEEE Symposium on Computers and Communications

(ISCC), pages 00091–00096. IEEE.

Mikolajczyk, K. and Schmid, C. (2005). A perfor-

mance evaluation of local descriptors. IEEE trans-

actions on pattern analysis and machine intelligence,

27(10):1615–1630.

Ostiak, P. (2006). Implementation of HDR panorama stitch-

ing algorithm. In Proceedings of the 10th CESCG

Conference, pages 24–26. Citeseer.

ribyl, B., Chalmers, A., and Zem

ık, P. (2012). Feature

point detection under extreme lighting conditions. In

Proceedings of the 28th Spring Conference on Com-

puter Graphics, pages 143–150.

ribyl, B., Chalmers, A., Zem

ık, P., Hooberman, L., and

Cad

ık, M. (2016). Evaluation of feature point detec-

tion in high dynamic range imagery. Journal of Visual

Communication and Image Representation, 38:141–

160.

Rana, A., Valenzise, G., and Dufaux, F. (2015). Evalua-

tion of feature detection in HDR based imaging un-

der changes in illumination conditions. In 2015 IEEE

International Symposium on Multimedia (ISM), pages

289–294.

Evaluation of a Local Descriptor for HDR Images

305

Rana, A., Valenzise, G., and Dufaux, F. (2016a). An eval-

uation of HDR image matching under extreme illumi-

nation changes. In 2016 Visual Communications and

Image Processing (VCIP), pages 1–4.

Rana, A., Valenzise, G., and Dufaux, F. (2016b). Optimiz-

ing tone mapping operators for keypoint detection un-

der illumination changes. In 2016 IEEE 18th Inter-

national Workshop on Multimedia Signal Processing

(MMSP), pages 1–6.

Rana, A., Valenzise, G., and Dufaux, F. (2017a). Learning-

based adaptive tone mapping for keypoint detection.

In 2017 IEEE International Conference on Multime-

dia and Expo (ICME), pages 337–342.

Rana, A., Valenzise, G., and Dufaux, F. (2017b). Learning-

based tone mapping operator for image matching. In

2017 IEEE International Conference on Image Pro-

cessing (ICIP), pages 2374–2378.

Rana, A., Valenzise, G., and Dufaux, F. (2019). Learning-

based tone mapping operator for efﬁcient image

matching. IEEE Transactions on Multimedia,

21(1):256–268.

Rey Otero, I. and Delbracio, M. (2014). Anatomy of the

SIFT Method. Image Processing On Line, 4:370–396.

https://doi.org/10.5201/ipol.2014.82.

Schmid, C., Mohr, R., and Bauckhage, C. (2000). Evalua-

tion of interest point detectors. International Journal

of computer vision, 37(2):151–172.

Se, S., Lowe, D., and Little, J. (2002). Global localization

using distinctive visual features. In IEEE/RSJ inter-

national conference on intelligent robots and systems,

volume 1, pages 226–231. IEEE.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

306