Performance Evaluation of BRISK Algorithm on Mobile Devices

Alexander Gularte, Camila Thomasi, Rodrigo de Bem and Diana Adamatti

Centro de Ci

encias Computacionais (C3), Universidade Federal do Rio Grande (FURG),

Av. It

alia, km 8, 96203-900, Rio Grande, RS, Brazil

Keywords:

Performance Evaluation, Local Features Extraction, Mobile Imaging.

Abstract:

The great number of researches about local features extraction algorithms in the last years, allied to the

popularization of mobile devices, makes desirable efﬁcient and accurate algorithms suitable to run on such

devices. Despite this, there are few approaches adequate to run efﬁciently on the complexity-, cost- and

power-constrained mobile environments. The main objective of this work is to evaluate the performance of

the recently proposed BRISK algorithm on mobile devices. In this way, a mobile implementation, named

M-BRISK, is proposed. Some implementation strategies are considered and successful applied to execute the

algorithm in a real-world mobile device. As evaluation criterion repeatability, recall, precision and running

time metrics are used, as well as the comparison with the classical well established algorithm SURF and also

with the more recently proposed ORB. The results conﬁrm that proposed mobile implementation of BRISK

(M-BRISK) performs well and it is adequate to mobile devices.

1 INTRODUCTION

The widespread use of digital cameras is responsible

for the generation of millions of images daily. Mobile

digital devices, such as mobile phones and tablet PCs,

extend the digital still cameras functionalities, allow-

ing real-time exchange of images through their com-

munication modems. These kind of devices, usually

capable of capturing images and videos, are nowa-

days available to millions of people over the world.

This fact makes them important tools for image ac-

quisition, allowing anyone to easily capture images

of products, advertisements, people’s faces and bod-

ies, pieces of art or any other visual data from the en-

vironment where he or she is inserted. Extracting use-

ful information from all these data, through efﬁcient

computational algorithms, is an important goal of the

Computer Vison research community.

The use of distinctive visual features to accom-

plish this goal has been highly explored in the last

decade and applied to many problems such as content-

based image retrieval (CBIR) (Smeulders et al.,

2000), markerless augmented reality (MAR) (Skryp-

nyk and Lowe, 2004) and 3D scene reconstruction

(Brown and Lowe, 2005). When a feature extraction

algorithm employs an unique descriptor to an entire

image it is called global. Otherwise, if it uses descrip-

tors to many image points, it is called local. The lo-

cal feature extraction algorithms are widely used be-

cause their invariance to many visual transformations.

These extractors are able to detect salient regions into

an image, which are called keypoints. Each keypoint

is described by a feature vector, that represents nu-

merically the image properties on the neighborhood

of such point. These sets of keypoints descriptors

can be used to perform matching between images,

through the calculation of similarity metrics among

them.

Despite the great profusion of researches about

features extraction algorithms in the last years, there

are few approaches suitable to run efﬁciently into

mobile devices. Some algorithms, developed to

have the best performance into unconstrained envi-

ronments, do not perform well in the complexity-,

cost- and power-constrained environment of such de-

vices (Budagavi, 2006).

In this context, the main contribution of this work

is to evaluate the performance of the recently pro-

posed BRISK algorithm (Leutenegger et al., 2011), a

local feature extractor, on a mobile device. To achieve

it, we did a consistent analysis of the BRISK method,

as well as the identiﬁcation of its critical points con-

cerning the computational cost. Some implemen-

tation strategies were considered and successful ap-

plied to execute the algorithm in a real-world mobile

device. As evaluation criterion repeatability, recall,

precision and running time metrics (Mikolajczyk and

Schmid, 2005) were used, as well as the comparison

Gularte A., Thomasi C., de Bem R. and Adamatti D..

Performance Evaluation of BRISK Algorithm on Mobile Devices.

DOI: 10.5220/0004288400050011

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2013), pages 5-11

ISBN: 978-989-8565-48-8

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

with the classical well established algorithm SURF

(Bay et al., 2006) and also with the more recently pro-

posed ORB (Rublee et al., 2011). The results conﬁrm

the hypothesis that proposed mobile implementation

of BRISK (M-BRISK) performs well and it is ade-

quated to mobile devices.

This paper is organized as follows. In Section 2

we present the main considered algorithms for local

features extraction. In Section 3 the BRISK method

is reviewed, while their mobile implementation is pre-

sented in Section 4. The experiments and results are

shown in Section 5. And ﬁnally, Section 6 presents

the conclusions and future work.

2 LOCAL FEATURES

DETECTION AND

DESCRIPTION

Lowe (Lowe, 2004) proposed the Scale Invariant Fea-

ture Transform (SIFT) detector and descriptor algo-

rithm, for extracting high distinctive features from

images. The SIFT method, one of the most reliable

features extractors, is invariant to scale and rotation

transformations, and partially invariant to brightness

and viewpoint changes. This algorithm is composed

by four main stages: scale-space extrema detection,

keypoints localization, keypoints orientation and key-

points description.

Due to the high dimensionality of the SIFT de-

scriptor and its poor computing performance, other

less time-consuming alternatives began to be pro-

posed in the literature, with the purpose of enabling

their use in real-time applications. In PCA-SIFT (Ke

and Sukthankar, 2004) algorithm, the Principal Com-

ponents Analysis (PCA) was applied to decrease the

dimension of SIFT descriptors and to improve the

matching time.

Bay et al. (Bay et al., 2006) presented a novel ap-

proach, partly based on SIFT, called Speed Up Robust

Features (SURF). For some cases, the SURF algo-

rithm is more robust than SIFT, taking less computing

time. SURF employs integral images to build a fast

detection method called Fast-Hessian. The descrip-

tion is based in a ﬁrst order Haar Wavelet distribu-

tion. The SURF proves to be a very efﬁcient method,

however it has limited performance when compared

to recent binary description approaches.

Calonder et al. (Calonder et al., 2010) built a bi-

nary descriptor, called Binary Robust Independent El-

ementary Features (BRIEF), with low dimensionality

and more suitable for real-time applications. In this

approach, the descriptors are formed by binary strings

which makes comparisons more efﬁcient. However,

despite the high performance, the algorithm suffers in

terms of robustness and reliability, because it has low

tolerance in terms of distortion and image transforma-

tions, such as scale and rotation.

In order to overcome the BRIEF limitations, in

terms of rotation invariance, the Oriented FAST and

Rotated BRIEF (ORB) (Rublee et al., 2011) algorithm

was proposed. This method uses Features from Ac-

celerated Segment Test (FAST) (Rosten and Drum-

mond, 2006) detector to provide a complete detection

and description algorithm. It has a good robustness

and a good performance, even concerning mobile de-

vices.

Leutenegger et al. (Leutenegger et al., 2011) pro-

posed the Binary Robust Invariant Scalable Keypoints

(BRISK), a new approach that joins the fast and ef-

ﬁcient Adaptative and Generic Accelerated Segment

Test (AGAST) (Mair et al., 2010) detector to a bi-

nary descriptor inspired by BRIEF. This algorithm,

that uses a sampling pattern resembling the one used

in DAISY dense descriptor (Tola et al., 2010), is in-

variant to many transformations, such as scale and ro-

tation. BRISK allies high-quality descriptor to low

computational requirements, what makes it proper to

real-time applications.

3 BRISK REVISITED

The BRISK method can be divided in three steps:

keypoints detection, keypoints description and de-

scriptors matching

. These stages are executed se-

quentially.

3.1 Keypoints Detection

It is performed at the ﬁrst moment. The goal at this

step is to identify salient points in the image that, ide-

ally, could be uniquely differentiated from any other

point. To do so, these points must be searched across

the image and scale dimensions using a saliency cri-

terion (Leutenegger et al., 2011). The search through

scale space is fundamental to achieve scale invariance

and it is realized in BRISK by the use of a pyramid

of images. This pyramid is formed by many layers

which correspond to resamples of the original image.

In BRISK method, these layers are divided in n oc-

taves c

and n intra-octaves d

, with i = {0, 1, ...., n}

and usually n = 4. The intra-octaves d

are located

between octaves c

and c

i+1

. These intra-octaves

The matching step can also be seen as an independent

stage.

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

and octaves are formed by the original image sub-

sampled by factors

(half-sampling) and

(two-

third-sampling). The original image is progressively

half-sampled to form the octaves, while the intra-

octaves are formed by progressives half-sampling of

the ﬁrst intra-octave d

, which in turn is composed by

a two-third-sampling of the original image.

After the octaves and intra-octaves calculation, the

AGAST corner detector (Mair et al., 2010) is applied

to each one of them. All the points of each layer

will be evaluated as a keypoint candidate, and to be

elected a point must be salient among their intra-layer

and inter-layers neighbors. It means that each point

is compared with its neighbors in the same layer and

also with those in the above and bellow layers. Af-

ter the analysis of all octaves and intra-octaves, the

detection stages ﬁnally ends, producing a set of key-

points with space and scale coordinates, what means

that each one of them can be exactly located.

3.2 Keypoints Description

In this stage, around each keypoint, which has its co-

ordinates found in the previous step, a sampling pat-

tern is positioned. The BRISK algorithm uses a sam-

pling pattern similar to the one in DAISY descriptor

(Tola et al., 2010). However, it is important to note

that the use of this pattern is quite different. The

pattern relies on 60 equally spaced points located in

four concentric rings. In order to produce a scale

and rotation invariant description, the sampling pat-

tern is exactly scaled and rotated according to each

keypoint. The BRISK descriptor is composed as a bi-

nary string of length 512. This string concatenates

the results of simple brightness comparisons tests be-

tween each keypoint and its 60 neighbors in the pat-

tern (Leutenegger et al., 2011). This approach was in-

spired by the BRIEF method (Calonder et al., 2010).

3.3 Matching

Finally, to perform the comparison between two or

more keypoint descriptors, the Hamming distance is

used. This distance measures the number of differ-

ent bits between two binary strings and it represents

the degree of inequality of the descriptors being com-

pared.

4 M-BRISK

This section presents the mobile implementation of

BRISK algorithm and the main aspects concern-

ing to the environment where it was developed and

tested. This mobile implementation is called M-

BRISK (Mobile-BRISK).

4.1 Implementation

In order to implement the algorithm to be executed

into a mobile device, some issues had to be overcame.

The ﬁrst of them was the fact that SSE2 and SSE3

instructions (Intel, 2006), the Intel’s SIMD instruc-

tions set (Flynn, 1972), were employed by Leuteneg-

ger et al. (Leutenegger et al., 2011) in their original

implementation of BRISK

. It was used as a crucial

resource to improve the BRISK performance at criti-

cal time-consuming routines.

SIMD instructions are available in ARM architec-

ture (ARM, 2012b), the most popular architecture of

mobile devices processors, just since the ARMv7 ver-

sion, through the NEON instructions (ARM, 2012a).

As these instructions are not available into all devices,

specially into the low-end ones, it was decided not to

use them in the present implementation. This choice

allows M-BRISK to be a more generic implementa-

tion, avoiding the dependence of speciﬁc mobile pro-

cessors and instructions sets. This approach also con-

stitutes a baseline assessment of the algorithm once it

is being evaluated in a most ordinary condition.

The sub-sampling routines, originally imple-

mented trough the SSE instructions, was replaced by

a bilinear interpolation function implemented using

the OpenCV library (OpenCV, 2011). This approach

showed to be efﬁcient and applicable to a mobile en-

vironment while it kept robustness.

Another critical point is the sampling pattern rou-

tine. In the original BRISK implementation, as a

manner to obtain this pattern efﬁciently, 65536 sam-

ples are generated to each one of the 60 pattern points

during the algorithm initilization and stored in a look-

up table. These samples corresponds to all possible

discretized values of scale and rotation. There are 64

and 1024 possible values to each dimension, respec-

tively. The goal of this approach is to save running

time in the description construction of each keypoint.

However, such process is so time-consuming that its

use is prohibitive on mobile devices. Thus, to avoid

this problem, we chose to previously save a binary

representation of the look-up table in mobile device

memory card. Doing so, at the time of descriptors

construction the table is already completely loaded

into device memory.

Finally, during the matching routine, the SSE in-

structions were also used to improve the performance

The original BRISK implementation is available in:

http://www.asl.ethz.ch/people/lestefan/personal/BRISK

PerformanceEvaluationofBRISKAlgorithmonMobileDevices

of original BRISK implementation. These instruc-

tions were used in this stage to optimize the calcula-

tion of the Hamming distance, employed to compare

two keypoint descriptors. In M-BRISK the Hamming

distance was also implemented through the OpenCV

library. Unfortunately this change represented con-

siderable addition in the running time of matching.

4.2 Mobile Execution Environment

The mobile platform chosen to support the imple-

mentation of M-BRISK algorithm was the Android

(Google, 2012). This decision was motivated mainly

because it is a well documented open development

platform. Another distinguishing feature of Android

is its ability to run very efﬁciently C and C++ li-

braries, allowing the reuse of previously implemented

code. In order to execute these libraries into Android,

the Native Development Kit (NDK) (Google, 2011)

was used. Since Android applications are written in

Java, to enable interfacing between the different lan-

guages, the NDK relies on the Java Native Interface

(JNI) (Liang, 2003). The Android 2.2 version was

used.

The employed device was a low-end mobile phone

with Mediatek MTK6516 416MHz ARMv5 proces-

sor, with 256MB of RAM memory. As mentioned

earlier, the NEON instructions are not available on

this processor architecture version.

5 EXPERIMENTS AND RESULTS

Leutenegger et al. (Leutenegger et al., 2011) com-

pared BRISK with SIFT and SURF algorithms, show-

ing the high performance of the former. The purpose

of the present experiments was to evaluate the perfor-

mance of M-BRISK. To accomplish this, the mobile

implementation is compared with the original BRISK

implementation. Tests were performed to compare

the repeatability rate and to measure the relation be-

tween execution time in the steps of detection, de-

scription and matching of the both approaches. More-

over, M-BRISK it is compared with SURF and ORB

using the recall and precision metrics and also on a

real mobile device, used to compare the running time

of such algorithms.

5.1 Image Dataset

The image dataset

used to perform the experiments

is a well established benchmark, largely employed to

The image dataset is available in: http://www.robots.

ox.ac.uk/∼vgg/research/afﬁne/.

(a) Wall. (b) Bikes. (c) Trees.

(d) Leuven. (e) Grafﬁti. (f) Boat.

(g) Ubc.

Figure 1: Image dataset used for the tests.

evaluate many methods in this ﬁeld (Calonder et al.,

2010), (Mikolajczyk and Schmid, 2005), including

the BRISK algorithm (Leutenegger et al., 2011). The

transformations covered by the dataset, shown in the

Figure 1, are: viewpoint change (Grafﬁti and Wall),

blur (Bikes and Trees), brightness change (Leuven),

zoom and rotation (Boat) and JPEG compression

(Ubc).

5.2 Repeatability Test

The repeatability test was used to identify the ability

of the detector to ﬁnd the same keypoint into distinct

images modiﬁed by some kind of transformation. The

goal of this test is to evaluate if the modiﬁcations in

the sub-sampling methods (half-sampling and two-

third-sampling) affected the quality of the detection

in the M-BRISK implementation.

The results in Figure 2 show the repeatability for

viewpoint change (Boat images) and zoom and rota-

tion transformations (Wall images). The percentual of

repeatability measured for each algorithm is shown on

top of each bar. Based in the values, we can conclude

that M-BRISK has repeatability similar to the BRISK,

proving that the mobile implementation of the detec-

tion step did not degrade the quality of the algorithm.

In some cases, M-BRISK has even slightly better re-

sults than BRISK.

5.3 Recall × 1-Precision Evaluation

Recall and 1-precision are two relevant metrics used

to show the effectiveness of a descriptor method.

Recall measures the proportion of correct positives

matches, considering all the possible correct posi-

tive matches between two images. While 1-precision

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

(a) Wall (b) Boat

Figure 2: Repeatability test for images of Wall and Boat

sets, with different transformations. The degree of changes

of viewpoint, zoom and rotation gradually increases from

the image 2 to image 6 in each image set. The image 1 is

considered as the original image in both sets.

(a) Recall × 1-precision curves for

the images Wall 1 and 4.

(b) Recall × 1-precision curves for

the images Boat 1 and 4.

the images Leuven 1 and 4.

(d) Recall × 1-precision curves for

the images Ubc 1 and 4.

Figure 3: Results shown recall × 1-precision curves for M-

BRISK, BRISK and ORB. Values of recall and 1-precision

were calculated for different matching thresholds (Ham-

ming distance limit bellow which two keypoints descriptors

are considered equal).

measures the proportion of false positives matches,

considering all the performed matches between those

images.

These curves represent the algorithm capacity to

keep a small number of false positives (1-precision)

while it associates the maximum number of keypoints

between two images (recall).

The resulting to recall × 1-precision curves,

shown in Figure 3, allow to compare the effectiveness

of M-BRISK, BRISK and ORB. It enables to eval-

uate the quality of the algorithms in all steps. The

curves allow us to observe that M-BRISK and BRISK

have similar robustness to viewpoint and brightness

changes. However, BRISK has slightly better results

to JPG compression, zoom and rotation transforma-

tions. Besides this, it is clear that ORB presents worst

results for almost all threshold values.

5.4 Running Time Test

These tests compare the time variations in the dif-

ferent stages of BRISK and M-BRISK implementa-

tions. Moreover, the comparison between the mobile

device running time of M-BRISK, SURF and ORB

are shown.

5.4.1 Distribution of Running Time between

BRISK and M-BRISK Stages

In Figure 4, the mean time per point spent in each

step of BRISK and M-BRISK is presented. The pur-

pose is to analyse the relation between the running

time of each algorithm step aiming to evaluate the ef-

fects of the modiﬁcations over BRISK. The tests were

performed using 100 samples of images for the sets

Grafﬁti, Boat and Leuven. The execution was host in

a computer with a Core2Duo T8100 2.10Ghz proces-

sor, 3GB of RAM memory and Windows 7 (64-bits)

operating system.

The mean times and the standard deviations are

showed at the top of the bars. As can be observed,

in the average, M-BRISK had performances in the

detection and matching stages, respectively, around

15% and 75% worst than BRISK. Even without the

use of SSE instructions, what was the reason of the

M-BRISK worst results, the time consumed in the de-

tection stage was not affected so signiﬁcantly. In the

description stage the performance of both implemen-

tations was similar.

Another important difference between the two

approaches is concerning the look-up table usage.

While it is calculated during the execution in BRISK,

in M-BRISK the look-up table is calculated previ-

ously. The mean time consumed in the BRISK look-

up table calcutation is 1.269s, while the mean time

spent to load it from memory in M-BRISK is 0.088s.

The latter amount of time corresponds to less than

10% of the former.

5.4.2 Running Time Test on a Mobile Device

The purpose of this test is to provide a reference for

running times of M-BRISK, SURF and ORB in a real

mobile device. These tests are performed using the

images Grafﬁti 1-3 with 50 samples.

Tables 1 and 2 show the execution running times

in the mobile device. The times in the tables represent

an average of all running time samples with their re-

spective standard deviation between parenthesis. As

we can observe, the M-BRISK is about twice faster

than ORB and more than 20 times faster than SURF

in detection and description steps. In the matching

PerformanceEvaluationofBRISKAlgorithmonMobileDevices

(a) Detection (b) Description

Figure 4: Comparison between the running time of BRISK

and M-BRISK steps. The mean amount of time spent per

point in each step is shown at the top of the bars, as well as

the standard deviations in parenthesis.

step, the time difference is not so signiﬁcant compar-

ing ORB and M-BRISK. However, SURF had worst

results than M-BRISK in all execution steps.

The Table 1 shows the detection and description

times. The time per point represents the average time

for detection and description for each keypoint. This

value represents a fairness comparison between the

algorithms. The Table 2 shows the matching times,

including the average time spent during each descrip-

tor comparison. The serial implementation of Ham-

ming distance decreased the matching performance,

however remained very fast when compared with

SURF descriptors matching. Even compared with

ORB, M-BRISK implementation still demonstrates

pretty competitive timings.

Table 1: Detection and description timings for the Grafﬁti

image 1.

SURF ORB M-BRISK

Number of points 3201 1000 1957

Detection time (s) 65.87 (1.56) 3.17 (1.13) 1.83 (1.81)

Description time (s) 140.66 (0.50) 2.85 (1.25) 3.59 (3.53)

Total time (s) 206.54 (1.71) 6.02 (1.89) 5.42 (5.35)

Time per point (ms) 64.52 (0.53) 6.02 (1.89) 2.77 (2.78)

Table 2: Matching timings for the Grafﬁti images 1 and 3.

SURF ORB M-BRISK

Points in ﬁrst image 3201 1000 1957

Points in second image 3820 1000 2529

Total time (s) 370.66 (17.90) 5.99 (2.44) 26.59 (9.66)

Time per comparison (ns) 30.31 (1.46) 5.99 (2.44) 5.37 (1.95)

6 CONCLUSIONS

The wide use of local features extractors applied to

many computer vision problems, allied to the popu-

larization of mobile devices, makes desirable efﬁcient

and accurate algorithms suitable to run on such de-

vices. These algorithms must join robustness and low

computational cost, and still there are few methods

efﬁciently adequate to the constrained mobile envi-

ronments. Thus, the main objective of this work was

to evaluate the performance of the recently proposed

BRISK algorithm on a mobile device. To do so, a

mobile implementation, named M-BRISK, was pro-

posed.

M-BRISK is focused to run in the most com-

mon ARM processors architecture. The implemen-

tation successfully overcame some mobile environ-

ment constraints, as processing power and limited in-

structions set, conﬁrmed through the analysis of the

obtained results. These showed that M-BRISK is

so efﬁcient and accurate as the original BRISK im-

plementation, which presents excellent results even

when compared with the well established SIFT and

SURF. Moreover, M-BRISK presented a overall run-

ning time approximately twice smaller than ORB and

20 times smaller than SURF, the other mobile ap-

proaches evaluated. Thus, it is possible to point M-

BRISK as a very suitable approach for detection and

description on mobile applications.

In future works, a version of M-BRISK with

NEON instructions will be tested against the serial

version. Also other parallel development paradigms,

such as GPGPU processing, supported by the most

advance mobile devices, must be considered. More-

over, a detailed analysis of M-BRISK computational

complexity is desirable.

REFERENCES

ARM (2012a). Arm neon instructions. Available in:

http://www.arm.com/products/processors/technologies/

neon.php, accessed on June 05, 2012.

ARM (2012b). Arm processor architecture. Available in:

http://www.arm.com/products/processors/technologie-

s/instruction-set-architectures.php, accessed on June

05, 2012.

Bay, H., Tuytelaars, T., and Gool, L. V. (2006). Surf:

Speeded up robust features. In Proceedings of

the Ninth European Conference on Computer Vision

(ECCV 2006), pages 404–417.

Brown, M. and Lowe, D. G. (2005). Unsupervised 3d object

recognition and reconstruction in unordered datasets.

In Proceedings of the Fifth International Conference

on 3-D Digital Imaging and Modeling (3DIM 2005),

pages 56–63.

Budagavi, M. (2006). Real-time image and video process-

ing in portable and mobile devices. Journal of Real-

Time Image Processing, 1(1):3–7.

Calonder, M., Lepetit, V., Strecha, C., and Fua, P. (2010).

Brief: Binary robust independent elementary features.

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

In Proceedings of the 11ht European Conference on

Computer Vision (ECCV 2010), pages 778–792.

Flynn, M. J. (1972). Some computer organizations and their

effectiveness. IEEE Transactions on Computers, C-

21(9):948–960.

Google (2011). Android ndk. Available in: http://

developer.android.com/sdk/ndk/overview.html,

accessed on November 15, 2011.

Google (2012). Android. Available in: http://

www.android.com/, accessed on February 29, 2012.

Intel (2006). Intel c++ compiler for linux intrinsics

reference. Available in: http://software.intel.com/

ﬁle/6337/, Accessed November 15, 2011.

Ke, Y. and Sukthankar, R. (2004). Pca-sift: a more dis-

tinctive representation for local image descriptors. In

Proceedings of the 2004 IEEE Conference on Com-

puter Vision and Pattern Recognition (CVPR 2004),

volume 2, pages II–506 – II–513.

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

Binary Robust Invariant Scalable Keypoints. In Pro-

ceedings of the 13ht International Conference on

Computer Vision (ICCV 2011).

Liang, S. (2003). Java Native Interface: Programmer’s

Guide and Speciﬁcation. Addison-Wesley.

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

invariant keypoints. 60(2):91–110.

Mair, E., Hager, G. D., Burschka, D., Suppa, M., and

Hirzinger, G. (2010). Adaptive and generic corner de-

tection based on the accelerated segment test. In Pro-

ceedings of 11th European Conference on Computer

Vision (ECCV 2010), pages 183–196.

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.

OpenCV (2011). Opencv v2.1 documentation. Avail-

able in http://opencv.willowgarage.com/wiki/, Ac-

cessed November 15, 2011.

Rosten, E. and Drummond, T. (2006). Machine learn-

ing for high-speed corner detection. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

3951:430–443.

Rublee, E., Rabaud, V., Konolige, K., and Bradski, G.

(2011). Orb: An efﬁcient alternative to sift or surf. In

Proceedings of 13th IEEE International Conference

on Computer Vision (ICCV 2011), pages 2564 –2571.

Skrypnyk, I. and Lowe, D. G. (2004). Scene modelling,

recognition and tracking with invariant image fea-

tures. In Proceedings of the Third IEEE and ACM In-

ternational Symposium on Mixed and Augmented Re-

ality (ISMAR 2004), pages 110–119.

Smeulders, A., Worring, M., Santini, S., Gupta, A., and

Jain, R. (2000). Content-based image retrieval at the

end of the early years. IEEE Transactions on Pat-

tern Analysis and Machine Intelligence, 22(12):1349

–1380.

Tola, E., Lepetit, V., and Fua, P. (2010). Daisy: An efﬁ-

cient dense descriptor applied to wide-baseline stereo.

IEEE Transactions on Pattern Analysis and Machine

Intelligence, 32(5):815 –830.

PerformanceEvaluationofBRISKAlgorithmonMobileDevices