BRISK: Binary Robust Invariant Scalable Keypoints

Stefan Leutenegger, Margarita Chli and Roland Y. Siegwart

Autonomous Systems Lab, ETH Zürich
{stefan.leutenegger, margarita.chli, and roland.siegwart}@mavt.ethz.ch

Abstract over, on top of fulfilling these properties to achieve the de-

sired quality of keypoints, the speed of detection and de-
Effective and efficient generation of keypoints from an scription needs also to be optimized to fit within the time-
image is a well-studied problem in the literature and forms constraints of the task at hand.
the basis of numerous Computer Vision applications. Es- In principle, state-of-the-art algorithms target applica-
tablished leaders in the field are the SIFT and SURF al- tions with either strict requirements in precision or speed
gorithms which exhibit great performance under a variety of computation. Lowe’s SIFT approach [9] is widely ac-
of image transformations, with SURF in particular consid- cepted as one of highest quality options currently avail-
ered as the most computationally efficient amongst the high- able, promising distinctiveness and invariance to a variety
performance methods to date. of common image transformations – however, the at the ex-
In this paper we propose BRISK1 , a novel method for pense of computational cost. On the other end of the spec-
keypoint detection, description and matching. A compre- trum, a combination of the FAST [14] keypoint detector and
hensive evaluation on benchmark datasets reveals BRISK’s the BRIEF [4] approach to description offers a much more
adaptive, high quality performance as in state-of-the-art al- suitable alternative for real-time applications. However, de-
gorithms, albeit at a dramatically lower computational cost spite the clear advantage in speed, the latter approach suf-
(an order of magnitude faster than SURF in cases). The fers in terms of reliability and robustness as it has minimal
key to speed lies in the application of a novel scale-space tolerance to image distortions and transformations, in par-
FAST-based detector in combination with the assembly of ticular to in-plane rotation and scale change. As a result,
a bit-string descriptor from intensity comparisons retrieved real-time applications like SLAM [6] need to employ prob-
by dedicated sampling of each keypoint neighborhood. abilistic methods [5] for data association to discover match-
ing consensus.
The inherent difficulty in extracting suitable features
1. Introduction from an image lies in balancing two competing goals: high-
quality description and low computational requirements.
Decomposing an image into local regions of interest or
This is where this work aims to set a new milestone with
‘features’ is a widely applied technique in Computer Vision
the BRISK methodology. Perhaps the most relevant work
used to alleviate complexity while exploiting local appear-
tackling this problem is SURF [2] which has been demon-
ance properties. Image representation, object recognition
strated to achieve robustness and speed, only, as evident in
and matching, 3D scene reconstruction and motion tracking
our results, BRISK achieves comparable quality of match-
all rely on the presence of stable, representative features in
ing at much less computation time. In a nutshell, this paper
the image, driving research and yielding a plethora of ap-
proposes a novel method for generating keypoints from an
proaches to this problem.
image, structured as follows:
The ideal keypoint detector finds salient image regions
such that they are repeatably detected despite change of • Scale-space keypoint detection: Points of interest are
viewpoint; more generally it is robust to all possible im- identified across both the image and scale dimensions us-
age transformations. Similarly, the ideal keypoint descrip- ing a saliency criterion. In order to boost efficiency of
tor captures the most important and distinctive information computation, keypoints are detected in octave layers of
content enclosed in the detected salient regions, such that the image pyramid as well as in layers in-between. The
the same structure can be recognized if encountered. More- location and the scale of each keypoint are obtained in
1 The reference implementation of BRISK can be downloaded from the continuous domain via quadratic function fitting.
http://www.asl.ethz.ch/people/lestefan/personal/
BRISK • Keypoint description: A sampling pattern consisting of

1
 points lying on appropriately scaled concentric circles is Probably the most appealing features at the moment are
applied at the neighborhood of each keypoint to retrieve the SURF [2], which have been demonstrated to be signif-
gray values: processing local intensity gradients, the fea- icantly faster than SIFT. SURF detection uses the determi-
ture characteristic direction is determined. Finally, the nant of the Hessian matrix (blob detector), while the de-
oriented BRISK sampling pattern is used to obtain pair- scription is done by summing Haar wavelet responses at the
wise brightness comparison results which are assembled region of interest. While demonstrating impressive timings
into the binary BRISK descriptor. with respect to the state-of-the-art, SURF are, in terms of
speed, still orders of magnitude away from the fastest, yet
Once generated, the BRISK keypoints can be matched limited quality features currently available.
very efficiently thanks to the binary nature of the descriptor. In this paper, we present a novel methodology dubbed
With a strong focus on efficiency of computation, BRISK ‘BRISK’ for high-quality, fast keypoint detection, descrip-
also exploits the speed savings offered in the SSE instruc- tion and matching. As suggested by the name, the method
tion set widely supported on today’s architectures. is rotation as well as scale invariant to a significant extent,
achieving performance comparable to the state-of-the-art
2. Related Work while dramatically reducing computational cost. Follow-
ing a description of the approach, we present experimen-
Identifying local interest points to be used for image
tal results performed on the benchmark datasets and using
matching can be traced a long way back in the literature,
the standardized evaluation method of [12, 13]. Namely,
with Harris and Stephens [7] proposing one of the earli-
we present evaluation of BRISK with respect to SURF and
est and probably most well-known corner detectors. The
SIFT which are widely accepted as a standard of compari-
seminal work of Mikolajzyk et al.[13] presented a compre-
son under common image transformations.
hensive evaluation of the most competent detection meth-
ods at the time, which revealed no single all-purpose de- 3. BRISK: The Method
tector but rather the complementary properties of the differ-
ent approaches depending on the context of the application. In this section, we describe the key stages in BRISK,
The more recent FAST criterion [14] for keypoint detection namely feature detection, descriptor composition and key-
has become increasingly popular in state-of-the-art methods point matching to the level of detail that the motivated
with hard real-time constraints, with AGAST [10] extend- reader can understand and reproduce. It is important to
ing this work for improved performance. note that the modularity of the method allows the use of
Amongst the best quality features currently in the litera- the BRISK detector in combination with any other keypoint
ture is the SIFT [9]. The high descriptive power and robust- descriptor and vice versa, optimizing for the desired perfor-
ness to illumination and viewpoint changes has rated the mance and the task at hand.
SIFT descriptor at the top of the rankings list in the survey
3.1. Scale-Space Keypoint Detection
in [11]. However, the high dimensionality of this descriptor
makes SIFT prohibitively slow. PCA-SIFT [8] reduced the With the focus on efficiency of computation, our detec-
descriptor from 128 to 36 dimensions, compromising how- tion methodology is inspired by the work of Mair et al.[10]
ever its distinctiveness and increasing the time for descrip- for detecting regions of interest in the image. Their AGAST
tor formation which almost annihilates the increased speed is essentially an extension for accelerated performance of
of matching. The GLOH descriptor [12] is also worth not- the now popular FAST, proven to be a very efficient basis
ing here, as it belongs to the family of SIFT-like methods for feature extraction. With the aim of achieving invariance
and has been shown to be more distinctive but also more to scale which is crucial for high-quality keypoints, we go
expensive to compute than SIFT. a step further by searching for maxima not only in the im-
The growing demand for high-quality, high-speed fea- age plane, but also in scale-space using the FAST score s as
tures has led to more research towards algorithms able to a measure for saliency. Despite discretizing the scale axis
process richer data at higher rates. Notable is the work at coarser intervals than in alternative high-performance de-
of Agrawal et al.[1] who apply a center-symmetric local tectors (e.g. the Fast-Hessian [2]), the BRISK detector es-
binary pattern as an alternative to SIFT’s orientation his- timates the true scale of each keypoint in the continuous
tograms approach. The most recent BRIEF [4] is designed scale-space.
for super-fast description and matching and consists of a In the BRISK framework, the scale-space pyramid lay-
binary string containing the results of simple image inten- ers consist of n octaves ci and n intra-octaves di , for
sity comparisons at random pre-determined pixel locations. i = {0, 1, . . . , n − 1} and typically n = 4. The oc-
Despite the simplicity and efficiency of this approach, the taves are formed by progressively half-sampling the orig-
method is very sensitive to image rotation and scale changes inal image (corresponding to c0 ). Each intra-octave di is lo-
restricting its application to general tasks. cated in-between layers ci and ci+1 (as illustrated in Figure
 log2(t) t: scale
1). The first intra-octave d0 is obtained by downsampling
octave ci+1
the original image c0 by a factor of 1.5, while the rest of
i+1
the intra-octave layers are derived by successive halfsam-
pling. Therefore, if t denotes scale then t(ci ) = 2i and
intra-octave di
t(di ) = 2i · 1.5.

It is important to note here that both FAST and AGAST

provide different alternatives of mask shapes for keypoint
octave ci
detection. In BRISK, we mostly use the 9-16 mask, which i
essentially requires at least 9 consecutive pixels in the 16-
interpolated position
pixel circle to either be sufficiently brighter or darker than intra-octave di-1
the central pixel for the FAST criterion to be fulfilled.

Initially, the FAST 9-16 detector is applied on each oc-

tave and intra-octave separately using the same threshold T
octave ci-1
to identify potential regions of interest. Next, the points be- i-1
longing to these regions are subjected to a non-maxima sup-
pression in scale-space: firstly, the point in question needs
to fulfill the maximum condition with respect to its 8 neigh- FAST score s

boring FAST scores s in the same layer. The score s is Figure 1. Scale-space interest point detection: a keypoint (i.e. saliency
maximum) is identified at octave ci by analyzing the 8 neighboring
defined as the maximum threshold still considering an im- saliency scores in ci as well as in the corresponding scores-patches in
age point a corner. Secondly, the scores in the layer above the immediately-neighboring layers above and below. In all three layers
and below will need to be lower as well. We check inside of interest, the local saliency maximum is sub-pixel refined before a 1D
parabola is fitted along the scale-axis to determine the true scale of the
equally sized square patches: the side-length is chosen to be keypoint. The location of the keypoint is then also re-interpolated between
2 pixels in the layer with the suspected maximum. Since the the patch maxima closest to the determined scale.
neighboring layers (and therefore its FAST scores) are rep-
resented with a different discretization, some interpolation
is applied at the boundaries of the patch. Figure 1 depicts
an example of this sampling and the maxima search.

The detection of maxima across the scale axis at octave

c0 is a special case: in order to obtain the FAST scores for
a virtual intra-octave d−1 below c0 , we apply the FAST 5-8
mask on c0 . However, the scores in patch of d−1 are in this
case not required to be lower than the score of the examined (a) Boat image 1 (b) Boat image 2
point in octave c0 . Figure 2. Close-up of a BRISK detection example on images 1 and 2 of
the Boat sequence exhibiting small zoom and in-plane rotation. The size
Considering image saliency as a continuous quantity not of the circles denote the scale of the detected keypoints while the radials
denote their orientation. For clarity, the detection threshold is set here to a
only across the image but also along the scale dimension, stricter value than in the typical setup, yielding slightly lower repeatability.
we perform a sub-pixel and continuous scale refinement for
each detected maximum. In order to limit complexity of the
refinement process, we first fit a 2D quadratic function in
3.2. Keypoint Description
the least-squares sense to each of the three scores-patches
(as obtained in the layer of the keypoint, the one above, and Given a set of keypoints (consisting of sub-pixel refined
the one below) resulting in three sub-pixel refined saliency image locations and associated floating-point scale values),
maxima. In order to avoid resampling, we consider a 3 by the BRISK descriptor is composed as a binary string by con-
3 score patch on each layer. Next, these refined scores are catenating the results of simple brightness comparison tests.
used to fit a 1D parabola along the scale axis yielding the This idea has been demonstrated in [4] to be very efficient,
final score estimate and scale estimate at its maximum. As a however here we employ it in a far more qualitative man-
final step, we re-interpolate the image coordinates between ner. In BRISK, we identify the characteristic direction of
the patches in the layers next to the determined scale. An each keypoint to allow for orientation-normalized descrip-
example of the BRISK detection in two images of the Boat tors and hence achieve rotation invariance which is key to
sequence (defined in Section 4) is shown up-close in Figure general robustness. Also, we carefully select the brightness
2. comparisons with the focus on maximizing descriptiveness.
 The threshold distances are set to δmax = 9.75t and
δmin = 13.67t (t is the scale of k). Iterating through the
15
point pairs in L, we estimate the overall characteristic pat-
10 tern direction of the keypoint k to be:
 
gx 1 X
5
g= = · g(pi , pj ). (4)
gy L
(pi ,pj )∈L
0

−5
The long-distance pairs are used for this computation, based
on the assumption that local gradients annihilate each other
−10 and are thus not necessary in the global gradient determina-
tion – this was also confirmed by experimenting with varia-
−15
tion of the distance threshold δmin .
−15 −10 −5 0 5 10 15

3.2.2 Building the Descriptor

Figure 3. The BRISK sampling pattern with N = 60 points: the small
blue circles denote the sampling locations; the bigger, red dashed circles For the formation of the rotation- and scale-normalized de-
are drawn at a radius σ corresponding to the standard deviation of the scriptor, BRISK applies the sampling pattern rotated by
Gaussian kernel used to smooth the intensity values at the sampling points. α = arctan2 (gy , gx ) around the keypoint k. The bit-vector
The pattern shown applies to a scale of t = 1.
descriptor dk is assembled by performing all the short-
distance intensity comparisons of point pairs (pα α
i , pj ) ∈ S
(i.e. in the rotated pattern), such that each bit b corresponds
3.2.1 Sampling Pattern and Rotation Estimation
to:
1, I(pα α

The key concept of the BRISK descriptor makes use of j , σj ) > I(pi , σi )
b=
a pattern used for sampling the neighborhood of the key- 0, otherwise (5)
point. The pattern, illustrated in Figure 3, defines N loca- ∀(pi , pα
α
) ∈ S
j
tions equally spaced on circles concentric with the keypoint.
While this pattern resembles the DAISY descriptor [15], it While the BRIEF descriptor is also assembled via bright-
is important to note that its use in BRISK is entirely dif- ness comparisons, BRISK has some fundamental differ-
ferent, as DAISY was built specifically for dense matching, ences apart from the obvious pre-scaling and pre-rotation
deliberately capturing more information and thus resulting of the sampling pattern. Firstly, BRISK uses a determinis-
to demanding speed and storage requirements. tic sampling pattern resulting in a uniform sampling-point
In order to avoid aliasing effects when sampling the im- density at a given radius around the keypoint. Consequently,
age intensity of a point pi in the pattern, we apply Gaus- the tailored Gaussian smoothing will not accidentally dis-
sian smoothing with standard deviation σi proportional to tort the information content of a brightness comparison by
the distance between the points on the respective circle. Po- blurring two close sampling-points in a comparison. Fur-
sitioning and scaling the pattern accordingly for a partic- thermore, BRISK uses dramatically fewer sampling-points
ular keypoint k in the image, let us consider one of the than pairwise comparisons (i.e. a single point participates
N · (N − 1)/2 sampling-point pairs (pi , pj ). The smoothed in more comparisons), limiting the complexity of looking-
intensity values at these points which are I(pi , σi ) and up intensity values. Finally, the comparisons here are re-
I(pj , σj ) respectively, are used to estimate the local gra- stricted spatially such that the brightness variations are only
dient g(pi , pj ) by required to be locally consistent. With the sampling pat-
tern and the distance thresholds as shown above, we obtain
I(pj , σj ) − I(pi , σi ) a bit-string of length 512. The bit-string of BRIEF64 also
g(pi , pj ) = (pj − pi ) · 2 . (1)
kpj − pi k contains 512 bits, thus the matching for a descriptor pair
will be performed equally fast by definition.
Considering the set A of all sampling-point pairs:
3.3. Descriptor Matching
A = (pi , pj ) ∈ R2 × R2 | i < N ∧ j < i ∧ i, j ∈ N


(2) Matching two BRISK descriptors is a simple computa-

we define a subset of short-distance pairings S and another tion of their Hamming distance as done in BRIEF [4]: the
subset of L long-distance pairings L: number of bits different in the two descriptors is a measure
of their dissimilarity. Notice that the respective operations
S = {(pi , pj ) ∈ A | kpj − pi k < δmax } ⊆ A reduce to a bitwise XOR followed by a bit count, which can
(3)
L = {(pi , pj ) ∈ A | kpj − pi k > δmin } ⊆ A. both be computed very efficiently on today’s architectures.
3.4. Notes on Implementation
Here, we give a very brief overview of some implemen-
tation issues which contribute significantly to the overall
computational performance and the reproducibility of the (a) Graffiti (b) Wall (c) Boat (d) Ubc
method. All the BRISK functionality builds on the com-
mon 2D feature interface of OpenCV 2.2 allowing easy inte-
gration and interchangeability with existing features (SIFT,
SURF, BRIEF, etc.).
The detection process uses the AGAST implementation
[10] for computing saliency scores. The non-maxima sup- (e) Leuven (f) Bikes (g) Trees
Figure 4. Datasets used for evaluation: viewpoint change (Graffiti and
pression benefits from early termination capability limiting Wall), zoom and rotation (Boat), JPEG compression (Ubc), brightness
the saliency scores calculation to a minimum. Building change (Leuven), and blur (Bikes and Trees).
the image pyramid makes use of some SSE2 and SSSE3
commands, both concerning the halfsampling as well as the
downsampling by a factor of 1.5.
(Ubc). Since the viewpoint change scenes are planar, the
In order to efficiently retrieve gray values with the sam-
image pairs in all sequences are provided with a ground
pling pattern, we generate a look-up table of discrete ro-
truth homography used to determine the corresponding key-
tated and scaled BRISK pattern versions (consisting of the
points. In the rest of the section we present quantitative
sampling-point locations and the properties of the Gaus-
results concerning the detector and descriptor performance
sian smoothing kernel as well as the indexing of long and
of BRISK compared to SIFT (OpenCV2.2 implementation)
short distance pairings) consuming around 40MB of RAM
as well as SURF (original implementation). Our evalua-
– which is still acceptable for applications constrained to
tion uses similarity matching which considers any pair of
low computational power.
keypoints with descriptor distance below a certain thresh-
We furthermore use the integral image along with a sim- old a match – in contrast to e.g. nearest neighbor matching,
plified Gaussian kernel version inspired by [2]: the kernel where a database is searched for the match with the lowest
is scalable when changing σ without any increase in com- descriptor distance. Finally, we also demonstrate BRISK’s
putational complexity. In our final implementation we use big advantage in computational speed by listing compara-
as an approximation a simple square box mean filter with tive timings.
floating point boundaries and side length ρ = 2.6 · σ.
Thus we do not need time-consuming Gaussian smooth- 4.1. BRISK Detector Repeatability
ing of the whole image with many different kernels, but we
instead retrieve single values using an arbitrary parameter The detector repeatability score as defined in [13] is cal-
σ. culated as the ratio between the corresponding keypoints
We also integrated an improved SSE Hamming distance and the minimum total number of keypoints visible in both
calculator achieving matching at 6 times the speed of the images. The correspondences are identified by looking at
current OpenCV implementation as used for example with the overlap area of the keypoint region in one image (i.e.
BRIEF in OpenCV. the extracted circle) and the projection of the keypoint re-
gion from the other image (i.e. ellipse-like): if the region
of intersection is larger than 50% of the union of the two
4. Experiments
regions, it is considered a correspondence. Note that this
Our proposed method has been extensively tested fol- method is largely dependent on the assignment of the key-
lowing the now established evaluation method and datasets point circle radius, i.e. the constant factor between scale and
in the field first proposed by Mikolajczyk and Schmid radius. We choose this such that the average radii obtained
[12, 13]. For the sake of consistency with results presented with the BRISK detector approximately match the average
in other works, we also used their MATLAB evaluation radii obtained with the SURF and SIFT detectors.
scripts which are available online. Each of the datasets The assessment of repeatability scores (a selection of
contains a sequence of six images exhibiting an increas- results is shown in Figure 5) is performed using constant
ing amount of transformation. All comparisons here are BRISK detection thresholds across one sequence. For the
performed against the first image in each dataset. Figure sake of a fair comparison with the SURF detector, we adapt
4 shows one image for each dataset analyzed. the respective Hessian threshold such that it outputs approx-
The transformations cover view-point change (Graffiti imately the same number of correspondences in the similar-
and Wall), zoom and rotation (Boat), blur (Bikes and Trees), ity based matching setup.
brightness changes (Leuven) as well as JPEG compression As illustrated in Figure 5, the BRISK detector exhibits
 100 BRISK 100 SIFT SURF BRISK
Repeatability score [%]

Repeatability score [%]

SURF 858 752
80 80
1504 631 1 1
1284
792 421 SIFT(1292), SURF(1338), BRISK(1284) SIFT(636), SURF(631), BRISK(633)
60 60
0.8 0.8
221
40 40

Recall [−]

Recall [−]
187 0.6 0.6
20 20
0.4 0.4
0 0
20 30 40 50 60 20 30 40 50 60 0.2 0.2
Viewpoint change [deg] Viewpoint change [deg]
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) Graffiti (b) Wall 1−Precision [−] 1−Precision [−]

100 100 (a) Graffiti 1-3 (b) Wall 1-4

Repeatability score [%]

Repeatability score [%]

2412 818 622 467

80 3352 1148 80 356 228 1 1
698
SIFT(1020), SURF(1009), BRISK(1049) SIFT(1187), SURF(1147), BRISK(1148)
60 591 60 0.8 0.8

Recall [−]

Recall [−]
40 40 0.6 0.6

20 20 0.4 0.4

0 0 0.2 0.2
1 1.5 2 2.5 2 3 4 5 6
scale change [−] Second image number [−] 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(c) Boat (d) Leuven 1−Precision [−] 1−Precision [−]
Figure 5. Repeatability scores for 50% overlap error of the BRISK and the (c) Image Rotation of 60◦ on Wall 1. (d) Boat 1-4
SURF detector. The resulting similarity correspondences (approximately
matched between the detectors) are given as numbers above the bars. 1 1
SIFT(660), SURF(465), BRISK(476) SIFT(2670), SURF(2714), BRISK(2712)
0.8 0.8
Recall [−]

Recall [−]
0.6 0.6
equivalent repeatability as the SURF detector as long as the 0.4 0.4
image transformations applied are not too large. Given the
0.2 0.2
clear advantage in computational cost of the BRISK over
0 0
the SURF detector however, the proposed method consti- 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
1−Precision [−] 1−Precision [−]
tutes a strong competitor, even if the performance at larger
(e) Bikes 1-4 (f) Trees 1-4
transformations appears to be slightly inferior.
1 1
4.2. Evaluation and Comparison of the Overall 0.8
SIFT(458), SURF(467), BRISK(467)
0.8
SIFT(1555), SURF(1562), BRISK(1645)

BRISK Algorithm
Recall [−]

Recall [−]

0.6 0.6

Since our work aims at providing an overall fast as well 0.4 0.4
as robust detection, description and matching, we evaluate 0.2 0.2
the joint performance of all these stages in BRISK and com- 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
pare it to SIFT and SURF. Figure 6 shows the precision- 1−Precision [−] 1−Precision [−]
recall curves using threshold-based similarity matching for
(g) Leuven 1-4 (h) Ubc 1-4
a selection of image pairs of different datasets. Again, for Figure 6. Evaluation results showing precision-recall curves (of all detec-
this assessment we adapt the detection thresholds such that tion, extraction and matching stages jointly) for BRISK, SURF and SIFT.
they output an approximately equal number of correspon- Results are shown for viewpoint changes (a and b), pure in-plane rotation
(c), zoom and rotation (d), blur (e and f), brightness changes (g) and JPEG
dences in the spirit of fairness. Note that the evaluation compression (h). The number of similarity correspondences are indicated
results here are different from the ones in [3], where all de- in the figures per algorithm. The red dotted line in (f) shows the perfor-
scriptors are extracted on the same regions (obtained with mance of BRISK descriptors extracted from SURF regions, yielding 2274
the Fast-Hessian detector). correspondences. Overall, BRISK exhibits competitive performance in all
cases and even outperforms SIFT and SURF in some cases.
As illustrated in Figure 6, BRISK performs competi-
tively with SIFT and SURF in all datasets and even out-
performs the other two in some cases. The reduced perfor-
mance of BRISK in the Trees dataset is attributed to the de- tive to blur than blob-like detectors. We therefore also show
tector performance: while SURF detects 2606 and 2624 re- the evaluation of the BRISK descriptors extracted from the
gions in the images, respectively, BRISK only detects 2004 SURF regions for the Trees dataset, demonstrating again
regions in image 4 compared to 5949 found in image 1 that the descriptor performance is comparable to SURF.
to achieve the approximately same number of correspon- Evidently, SIFT performs significantly worse in the
dences. The same holds for the other blur dataset, Bikes: Trees, Boat, and Ubc datasets, which can be explained with
saliency as assessed with FAST is inherently more sensi- the limited detector repeatability in these cases. On the
 1 1 SIFT SURF BRISK
0.8 0.8 Detection threshold 4.4 45700 67
Number of points 1851 1557 1051
Recall [−]

Recall [−]
0.6 0.6
BRIEF64 Detection time [ms] 1611 107.9 17.20
0.4 SU−BRISK 0.4
Description time [ms] 9784 559.1 22.08
S−BRISK
0.2 0.2
BRISK Total time [ms] 11395 667.0 39.28
0
0 0.2 0.4 0.6 0.8 1
0
0 0.2 0.4 0.6 0.8 1
Time per point (ms) 6.156 0.4284 0.03737
1−Precision [−] 1−Precision [−]
Table 1. Detection and extraction timings for the first image in the Graffiti
(a) Wall 1-2 (b) Boat 1-2 sequence (size: 800 × 640 pixels).
Figure 7. Comparison of different BRISK versions to 64 byte BRIEF.
BRIEF, as well as both SU-BRISK (single-scale, unrotated) and S-BRISK
(single-scale) are extracted from AGAST keypoints detected in the original SIFT SURF BRISK
image. Notice that the BRISK pattern was scaled such that it matches the
BRIEF patch size. The standard version of BRISK had to be extracted
Points in first image 1851 1557 1051
from our scale-invariant corner detection with adapted threshold to match Points in second image 2347 1888 1385
the number of correspondences: they are 850 in the Wall pair and 1530 in Total time [ms] 291.6 194.6 29.92
the Boat pair. Time per comparison [ns] 67.12 66.20 20.55
Table 2. Matching timings for the Graffiti image 1 and 3 setup.

other hand, SIFT and BRISK handle the important case of

pure in-plane rotation very well and better than SURF.
BRISK is easily scalable for faster execution by reducing
In order to complete the experimental section, we want
the number of sampling-points in the pattern at some ex-
to make the link to BRIEF. Figure 7 shows a comparison of
pense of matching quality – which might be affordable in
the unrotated, single-scale BRISK version (SU-BRISK) to
a particular application. Moreover, scale and/or rotation in-
64 byte BRIEF features on the same (single scale) AGAST
variance can be omitted trivially, increasing the speed as
keypoints. Also included are the rotation invariant, single-
well as the matching quality in applications where they are
scale S-BRISK, as well as the standard BRISK. The exper-
not needed.
iment is conducted with two image pairs: on the one hand,
we used the first two images in the Wall dataset proving that 4.4. An Example
SU-BRISK and BRIEF64 are exhibiting a very similar per-
formance in the absence of scale change and in-plane ro- Complementary to the extensive evaluation presented
tation. Notice that this is really the situation BRIEF was above, we also provide a real-world example demonstrat-
designed for. On the other hand, we applied the differ- ing matching using BRISK. Figure 8 shows an image pair
ent versions to the first two images of the Boat sequence: exhibiting various transformations. A similarity match with
this experiment demonstrates some advantage of the SU- a threshold of 90 was performed (out of 512 comparisons)
BRISK over BRIEF in terms of robustness against small resulting in robust matches without significant outliers.
rotation (10◦ ) and scale changes (10%). Furthermore, the
well known and intuitive price for both rotation and scale 5. Conclusions
invariance is easily observable. We have presented a novel method named BRISK, which
tackles the classic Computer Vision problem of detecting,
4.3. Timings
describing and matching image keypoints for cases with-
Timings have been recorded on a laptop with a quad- out sufficient a priori knowledge on the scene and cam-
core i7 2.67 GHz processor (only using one core, however) era poses. In contrast to well-established algorithms with
running Ubuntu 10.04 (32-bit), using the implementation proven high performance, such as SIFT and SURF, the
and setup as detailed above. Table 1 presents the results method at hand offers a dramatically faster alternative at
concerning detection on the first image of the Graffiti se- comparable matching performance – a statement which we
quence, while Table 2 shows the matching times. The val- base on an extensive evaluation using an established frame-
ues are averaged over 100 runs. Note that all matchers do work. BRISK relies on an easily configurable circular sam-
a brute-force descriptor distance computation without any pling pattern from which it computes brightness compar-
early termination optimizations. isons to form a binary descriptor string. The unique prop-
The timings show a clear advantage of BRISK. Its de- erties of BRISK can be useful for a wide spectrum of ap-
tection and descriptor computation is typically an order of plications, in particular for tasks with hard real-time con-
magnitude faster than the one of SURF, which are consid- straints or limited computation power: BRISK finally offers
ered to be the fastest rotation and scale invariant features the quality of high-end features in such time-demanding ap-
currently available. It is also important to highlight that plications.
Figure 8. BRISK matching example: a detection threshold of 70 is used and a matching Hamming distance threshold of 90. The resulting matches
are connected by the green lines showing no clear false positives. The authors provide a reference implementation of BRISK downloadable from
http://www.asl.ethz.ch/people/lestefan/personal/BRISK .

Amongst avenues for further research into BRISK, we [6] A. J. Davison, N. D. Molton, I. Reid, and O. Stasse.
aim to explore alternatives to the scale-space maxima search MonoSLAM: Real-time single camera SLAM. IEEE
of saliency scores to yield higher repeatability whilst main- Transactions on Pattern Analysis and Machine Intelligence
taining speed. Furthermore, we aim at analyzing both theo- (PAMI), 29(6):1052–1067, 2007. 1
retically and experimentally the BRISK pattern and the con- [7] C. Harris and M. Stephens. A combined corner and edge de-
figuration of comparisons, such that the information content tector. In Proceedings of 4th Alvey Vision Conference, pages
147–151, 1988. 2
and/or robustness of the descriptor is maximized.
[8] Y. Ke and R. Sukthankar. PCA-SIFT: A More Distinctive
Representation for Local Image Descriptors. 2004. 2
6. Acknowledgements
[9] D. G. Lowe. Distinctive image features from scale-invariant
This research was supported by the Autonomous Sys- keypoints. International Journal of Computer Vision (IJCV),
tems Lab, ETH Zurich and the EC’s 7th Framework 60(2):91–110, 2004. 1, 2
Programme (FP7/2001-2013) under grant agreement no. [10] E. Mair, G. D. Hager, D. Burschka, M. Suppa, and
231855 (sFly). We are grateful to Simon Lynen and Davide G. Hirzinger. Adaptive and generic corner detection based
on the accelerated segment test. In Proceedings of the Eu-
Scaramuzza for their valuable inputs, as well as to many
ropean Conference on Computer Vision (ECCV), 2010. 2,
other colleagues at ETH Zurich for very helpful discussions.
5
[11] K. Mikolajczyk and C. Schmid. A performance evaluation
References of local descriptors. In Proceedings of the IEEE Conference
[1] M. Agrawal, K. Konolige, and M. R. Blas. CenSurE: Center on Computer Vision and Pattern Recognition (CVPR), 2003.
surround extremas for realtime feature detection and match- 2
ing. In Proceedings of the European Conference on Com- [12] K. Mikolajczyk and C. Schmid. A performance evaluation
puter Vision (ECCV), 2008. 2 of local descriptors. IEEE Transactions on Pattern Analysis
[2] H. Bay, A. Ess, T. Tuytelaars, and L. V. Gool. SURF: and Machine Intelligence (PAMI), 2:1115–1125, 2005. 2, 5
Speeded up robust features. Computer Vision and Image Un- [13] K. Mikolajczyk, T. Tuytelaars, C. Schmid, A. Zisserman,
derstanding (CVIU), 110(3):346–359, 2008. 1, 2, 5 J. Matas, F. Schaffalitzky, T. Kadir, and L. Gool. A com-
[3] H. Bay, T. Tuytelaars, and L. Van Gool. SURF: Speeded up parison of affine region detectors. International Journal of
robust features. In Proceedings of the European Conference Computer Vision (IJCV), 65(1):43–72, 2005. 2, 5
on Computer Vision (ECCV), 2006. 6 [14] E. Rosten and T. Drummond. Machine learning for high-
[4] M. Calonder, V. Lepetit, C. Strecha, and P. Fua. BRIEF: speed corner detection. In Proceedings of the European Con-
Binary Robust Independent Elementary Features. In Pro- ference on Computer Vision (ECCV), 2006. 1, 2
ceedings of the European Conference on Computer Vision [15] E. Tola, V. Lepetit, and P. Fua. Daisy: an Efficient Dense
(ECCV), 2010. 1, 2, 3, 4 Descriptor Applied to Wide Baseline Stereo. IEEE Transac-
[5] M. Chli and A. J. Davison. Active Matching. In Proceedings tions on Pattern Analysis and Machine Intelligence (PAMI),
of the European Conference on Computer Vision (ECCV), 32(5):815–830, 2010. 4
2008. 1