See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/215458901

Machine Learning for High-Speed Corner Detection

Conference Paper · July 2006

DOI: 10.1007/11744023_34 · Source: DBLP

CITATIONS READS

2,726 4,272

2 authors, including:

Edward Rosten
University of Cambridge
60 PUBLICATIONS   5,617 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Edward Rosten on 04 June 2014.

The user has requested enhancement of the downloaded file.

Machine learning for high-speed corner detection

Edward Rosten and Tom Drummond

Department of Engineering, Cambridge University, UK

{er258, twd20}@cam.ac.uk

Abstract Where feature points are used in real-time frame-rate appli-

cations, a high-speed feature detector is necessary. Feature detectors such
as SIFT (DoG), Harris and SUSAN are good methods which yield high
quality features, however they are too computationally intensive for use
in real-time applications of any complexity. Here we show that machine
learning can be used to derive a feature detector which can fully process
live PAL video using less than 7% of the available processing time. By
comparison neither the Harris detector (120%) nor the detection stage
of SIFT (300%) can operate at full frame rate.
Clearly a high-speed detector is of limited use if the features produced
are unsuitable for downstream processing. In particular, the same scene
viewed from two different positions should yield features which corre-
spond to the same real-world 3D locations[1]. Hence the second contri-
bution of this paper is a comparison corner detectors based on this crite-
rion applied to 3D scenes. This comparison supports a number of claims
made elsewhere concerning existing corner detectors. Further, contrary
to our initial expectations, we show that despite being principally con-
structed for speed, our detector significantly outperforms existing feature
detectors according to this criterion.

1 Introduction
Corner detection is used as the first step of many vision tasks such as tracking,
SLAM (simultaneous localisation and mapping), localisation, image matching
and recognition. Hence, a large number of corner detectors exist in the litera-
ture. With so many already available it may appear unnecessary to present yet
another detector to the community; however, we have a strong interest in real-
time frame rate applications such as SLAM in which computational resources
are at a premium. In particular, it is still true that when processing live video
streams at full frame rate, existing feature detectors leave little if any time for
further processing, even despite the consequences of Moore’s Law.
Section 2 of this paper demonstrates how a feature detector described in ear-
lier work can be redesigned employing a machine learning algorithm to yield a
large speed increase. In addition, the approach allows the detector to be gen-
eralised, producing a suite of high-speed detectors which we currently use for
real-time tracking [2] and AR label placement [3].
To show that speed can been obtained without necessarily sacrificing the
quality of the feature detector we compare our detector, to a variety of well-
known detectors. In Section 3 this is done using Schmid’s criterion [1], that
2 Edward Rosten and Tom Drummond

when presented with different views of a 3D scene, a detector should yield (as
far as possible) corners that correspond to the same features in the scene. Here
we show how this can be applied to 3D scenes for which an approximate surface
model is known.

1.1 Previous work

The majority of feature detection algorithms work by computing a corner re-

sponse function (C) across the image. Pixels which exceed a threshold cornerness
value (and are locally maximal) are then retained.
Moravec [4] computes the sum-of-squared-differences (SSD) between a patch
around a candidate corner and patches shifted a small distance in a number of
directions. C is then the smallest SSD so obtained, thus ensuring that extracted
corners are those locations which change maximally under translations.
Harris[5] builds on this by computing an approximation to the second deriva-
tive of the SSD with respect to the shift The approximation is:
" #
Ibx2 Id
x Iy
H= b2 , (1)
Id
x Iy Iy

where b denotes averaging performed over the image patch (a smooth circular
window can be used instead of a rectangle to perform the averaging resulting in
a less noisy, isotropic response). Harris then defines the corner response to be

C = |H| − k(trace H)2 . (2)

This is large if both eigenvalues of H are large, and it avoids explicit computation
of the eigenvalues. It has been shown[6] that the eigenvalues are an approximate
measure of the image curvature.
Based on the assumption of affine image deformation, a mathematical anal-
ysis led Shi and Tomasi[7] conclude that it is better to use the smallest eigen
value of H as the corner strength function:

C = min (λ1 , λ2 ). (3)

A number of suggestion have [5,7,8,9] been made for how to compute the corner
strength from H and these have been all shown [10] to be equivalent to various
matrix norms of H
Zheng et al.[11] perform an analysis of the computation of H, and find some
suitable approximations which allow them to obtain a speed increase by com-
puting only two smoothed images, instead of the three previously required.
Lowe [12] obtains scale invariance by convolving the image with a Difference
of Gaussians (DoG) kernel at multiple scales, retaining locations which are op-
tima in scale as well as space. DoG is used because it is good approximation
for the Laplacian of a Gaussian (LoG) and much faster to compute. An √ approx-
imation to DoG has been proposed which, provided that scales are 2 apart,
 Machine learning for high-speed corner detection 3

speeds up computation by a factor of about two, compared to the striaghtforward

implementation of Gaussian convolution [13].
It is noted in [14] that the LoG is a particularly stable scale-space kernel.
Scale-space techniques have also been combined with the Harris approach
in [15] which computes Harris corners at multiple scales and retains only those
which are also optima of the LoG response across scales.
Recently, scale invariance has been extended to consider features which are
invariant to affine transformations [14,16,17].
An edge (usually a step change in intensity) in an image corresponds to the
boundary between two regions. At corners of regions, this boundary changes di-
rection rapidly. Several techniques were developed which involved detecting and
chaining edges with a view to finding corners in the chained edge by analysing the
chain code[18], finding maxima of curvature [19,20,21], change in direction [22]
or change in appearance[23]. Others avoid chaining edges and instead look for
maxima of curvature[24] or change in direction [25] at places where the gradient
is large.
Another class of corner detectors work by examining a small patch of an im-
age to see if it “looks” like a corner. Since second derivatives are not computed, a
noise reduction step (such as Gaussian smoothing) is not required. Consequently,
these corner detectors are computationally efficient since only a small number
of pixels are examined for each corner detected. A corollary of this is that they
tend to perform poorly on images with only large-scale features such as blurred
images. The corner detector presented in this work belongs to this category.
The method presented in [26] assumes that a corner resembles a blurred
wedge, and finds the characteristics of the wedge (the amplitude, angle and
blur) by fitting it to the local image. The idea of the wedge is generalised in [27],
where a method for calculating the corner strength is proposed which computes
self similarity by looking at the proportion of pixels inside a disc whose intensity
is within some threshold of the centre (nucleus) value. Pixels closer in value to
the nucleus receive a higher weighting. This measure is known as the USAN (the
Univalue Segment Assimilating Nucleus). A low value for the USAN indicates a
corner since the centre pixel is very different from most of its surroundings. A set
of rules is used to suppress qualitatively “bad” features, and then local minima
of the, SUSANs, (Smallest USAN) are selected from the remaining candidates.
Trajkovic and Hedley[28] use a similar idea: a patch is not self-similar if
pixels generally look different from the centre of the patch. This is measured by
considering a circle. fC is the pixel value at the centre of the circle, and fP and
fP ′ are the pixel values at either end of a diameter line across the circle. The
response function is defined as

C = min (fP − fC )2 + (fP ′ − fC )2 . (4)

P

This can only be large in the case where there corner. The test is performed
on a Bresenham circle. Since the circle is discretized, linear or circular interpo-
lation is used in between discrete orientations in order to give the detector a
more isotropic response. To this end, the authors present a method whereby the
4 Edward Rosten and Tom Drummond

16 1 2
15 3

14 4
13 p 5
12 6

11 7
10 9 8

Figure 1. 12 point segment test corner detection in an image patch. The highlighted
squares are the pixels used in the corner detection. The pixel at p is the centre of a
candidate corner. The arc is indicated by the dashed line passes through 12 contiguous
pixels which are brighter than p by more than the threshold.

minimum response function at all interpolated positions between two pixels can
be efficiently computed. Computing the response function requires performing
a search over all orientations, but any single measurement provides an upper
bound on the response. To speed up matching, the response in the horizontal
and vertical directions only is checked. If the upper bound on the response is
too low, then the potential corner is rejected. To speed up the method further,
this fast check is first applied at a coarse scale.
A fast radial symmetry transform is developed in [29] to detect points. Points
have a high score when the gradient is both radially symmetric, strong, and of
a uniform sign along the radius. The scale can be varied by changing the size of
the area which is examined for radial symmetry.
An alternative method of examining a small patch of an image to see if it
looks like a corner is to use machine learning to classify patches of the image as
corners or non-corners. The examples used in the training set determine the type
of features detected. In [30], a three layer neural network is trained to recognise
corners where edges meet at a multiple of 45◦ , near to the centre of an 8 × 8
window. This is applied to images after edge detection and thinning. It is shown
how the neural net learned a more general representation and was able to detect
corners at a variety of angles.

2 High-speed corner detection

2.1 FAST: Features from Accelerated Segment Test
The segment test criterion operates by considering a circle of sixteen pixels
around the corner candidate p. The original detector [2,3] classifies p as a corner
 Machine learning for high-speed corner detection 5

if there exists a set of n contiguous pixels in the circle which are all brighter
than the intensity of the candidate pixel Ip plus a threshold t, or all darker
than Ip − t, as illustrated in Figure 1. n was chosen to be twelve because it
admits a high-speed test which can be used to exclude a very large number of
non-corners: the test examines only the four pixels at 1, 5, 9 and 13 (the four
compass directions). If p is a corner then at least three of these must all be
brighter than Ip + t or darker than Ip − t. If neither of these is the case, then p
cannot be a corner. The full segment test criterion can then be applied to the
remaining candidates by examining all pixels in the circle. This detector in itself
exhibits high performance, but there are several weaknesses:

1. The high-speed test does not generalise well for n < 12.
2. The choice and ordering of the fast test pixels contains implicit assumptions
about the distribution of feature appearance.
3. Knowledge from the first 4 tests is discarded.
4. Multiple features are detected adjacent to one another.

2.2 Machine learning a corner detector

Here we present an approach which uses machine learning to address the first
three points (the fourth is addressed in Section 2.3). The process operates in
two stages. In order to build a corner detector for a given n, first, corners are
detected from a set of images (preferably from the target application domain)
using the segment test criterion for n and a convenient threshold. This uses a
slow algorithm which for each pixel simply tests all 16 locations on the circle
around it.
For each location on the circle x ∈ {1..16}, the pixel at that position relative
to p (denoted by p → x) can have one of three states:

 d, Ip→x ≤ Ip − t (darker)
Sp→x = s, Ip − t < Ip→x < Ip + t (similar) (5)

b, Ip + t ≤ Ip→x (brighter)

Choosing an x and computing Sp→x for all p ∈ P (the set of all pixels in all train-
ing images) partitions P into three subsets, Pd , Ps , Pb , where each p is assigned
to PSp→x .
Let Kp be a boolean variable which is true if p is a corner and false otherwise.
Stage 2 employs the algorithm used in ID3[31] and begins by selecting the x
which yields the most information about whether the candidate pixel is a corner,
measured by the entropy of Kp .
The entropy of K for the set P is:

H(P ) = (c + c̄) log2 (c + c̄) − c log2 c − c̄ log2 c̄ (6)

¯ ¯
where c = ¯{p|Kp is true}¯ (number of corners)
¯ ¯
and c̄ = ¯{p|Kp is false}¯ (number of non corners)
6 Edward Rosten and Tom Drummond

The choice of x then yields the information gain:

H(P ) − H(Pd ) − H(Ps ) − H(Pb ) (7)
Having selected the x which yields the most information, the process is applied
recursively on all three subsets i.e. xb is selected to partition Pb in to Pb,d , Pb,s ,
Pb,b , xs is selected to partition Ps in to Ps,d , Ps,s , Ps,b and so on, where each
x is chosen to yield maximum information about the set it is applied to. The
process terminates when the entropy of a subset is zero. This means that all
p in this subset have the same value of Kp , i.e. they are either all corners or
all non-corners. This is guaranteed to occur since K is an exact function of the
learning data.
This creates a decision tree which can correctly classify all corners seen in the
training set and therefore (to a close approximation) correctly embodies the rules
of the chosen FAST corner detector. This decision tree is then converted into
C-code, creating a long string of nested if-then-else statements which is compiled
and used as a corner detector. For full optimisation, the code is compiled twice,
once to obtain profiling data on the test images and a second time with arc-
profiling enabled in order to allow reordering optimisations. In some cases, two
of the three subtrees may be the same. In this case, the boolean test which
separates them is removed.
Note that since the data contains incomplete coverage of all possible corners,
the learned detector is not precisely the same as the segment test detector. It
would be relatively straightforward to modify the decision tree to ensure that it
has the same results as the segment test algorithm, however, all feature detectors
are heuristic to some degree, and the learned detector is merely a very slightly
different heuristic to the segment test detector.

2.3 Non-maximal suppression

Since the segment test does not compute a corner response function, non max-
imal suppression can not be applied directly to the resulting features. Conse-
quently, a score function, V must be computed for each detected corner, and
non-maximal suppression applied to this to remove corners which have an adja-
cent corner with higher V . There are several intuitive definitions for V :
1. The maximum value of n for which p is still a corner.
2. The maximum value of t for which p is still a corner.
3. The sum of the absolute difference between the pixels in the contiguous arc
and the centre pixel.
Definitions 1 and 2 are very highly quantised measures, and many pixels share
the same value of these. For speed of computation, a slightly modified version
of 3 is used. V is given by:
 
X X
V = max  |Ip→x − Ip | − t , |Ip − Ip→x | − t (8)
x∈Sbright x∈Sdark
 Machine learning for high-speed corner detection 7

Detector Opteron 2.6GHz Pentium III 850MHz

ms % ms %
Fast n = 9 (non-max suppression) 1.33 6.65 5.29 26.5
Fast n = 9 (raw) 1.08 5.40 4.34 21.7
Fast n = 12 (non-max suppression) 1.34 6.70 4.60 23.0
Fast n = 12 (raw) 1.17 5.85 4.31 21.5
Original FAST n = 12 (non-max suppression) 1.59 7.95 9.60 48.0
Original FAST n = 12 (raw) 1.49 7.45 9.25 48.5
Harris 24.0 120 166 830
DoG 60.1 301 345 1280
SUSAN 7.58 37.9 27.5 137.5
Table 1. Timing results for a selection of feature detectors run on fields (768 × 288)
of a PAL video sequence in milliseconds, and as a percentage of the processing budget
per frame. Note that since PAL and NTSC, DV and 30Hz VGA (common for web-
cams) have approximately the same pixel rate, the percentages are widely applicable.
Approximately 500 features per field are detected.

with

Sbright ={x|Ip→x ≥ Ip + t}
Sdark ={x|Ip→x ≤ Ip − t}
(9)

2.4 Timing results

Timing tests were performed on a 2.6GHz Opteron and an 850MHz Pentium III
processor. The timing data is taken over 1500 monochrome fields from a PAL
video source (with a resolution of 768 × 288 pixels). The learned FAST detectors
for n = 9 and 12 have been compared to the original FAST detector, to our
implementation of the Harris and DoG (difference of Gaussians—the detector
used by SIFT) and to the reference implementation of SUSAN[32].
As can be seen in Table 1, FAST in general offers considerably higher perfor-
mance than the other tested feature detectors, and the learned FAST performs
up to twice as fast as the handwritten version. Importantly, it is able to gen-
erate an efficient detector for n = 9, which (as will be shown in Section 3) is
the most reliable of the FAST detectors. On modern hardware, FAST consumes
only a fraction of the time available during video processing, and on low power
hardware, it is the only one of the detectors tested which is capable of video rate
processing at all.
Examining the decision tree shows that on average, 2.26 (for n = 9) and 2.39
(for n = 12) questions are asked per pixel to determine whether or not it is a
feature. By contrast, the handwritten detector asks on average 2.8 questions.
Interestingly, the difference in speed between the learned detector and the
original FAST are considerably less marked on the Opteron processor compared
to the Pentium III. We believe that this is in part due to the Opteron having
8 Edward Rosten and Tom Drummond

a diminishing cost per pixel queried that is less well modelled by our system
(which assumes equal cost for all pixel accesses), compared to the Pentium III.

3 A comparison of detector repeatability

Although there is a vast body of work on corner detection, there is much less
on the subject of comparing detectors. Mohannah and Mokhtarian[33] evaluate
performance by warping test images in an affine manner by a known amount.
They define the ‘consistency of corner numbers’ as

CCN = 100 × 1.1−|nw −no | ,

where nw is the number of features in the warped image and no is the number
of features in the original image. They also define accuracy as
na na
no + ng
ACU = 100 × ,
2
where ng are the number of ‘ground truth’ corners (marked by humans) and na
is the number of matched corners compared to the ground truth. This unfortu-
nately relies on subjectively made devisions.
Trajkovic and Hedley[28] define stability to be the number of ‘strong ’ matches
(matches detected over three frames in their tracking algorithm) divided by the
total number of corners. This measurement is clearly dependent on both the
tracking and matching methods used, but has the advantage that it can be
tested on the date used by the system.
When measuring reliability, what is important is if the same real-world fea-
tures are detected from multiple views [1] This is the definition which will be
used here. For an image pair, a feature is ‘detected’ if is is extracted in one
image and appears in the second. It is ‘repeated’ if it is also detected nearby in
the second. The repeatability is the ratio of repeated features detected features.
In [1], the test is performed on images of planar scenes so that the relationship
between point positions is a homography. Fiducial markers are projected on to
the planar scene to allow accurate computation of this.
By modelling the surface as planar and using flat textures, this technique
tests the feature detectors’ ability to deal with mostly affine warps (since image
features are small) under realistic conditions. This test is not so well matched
to our intended application domain, so instead, we use a 3D surface model to
compute where detected features should appear in other views (illustrated in
Figure 2). This allows the repeatability of the detectors to be analysed on features
caused by geometry such as corners of polyhedra, occlusions and T-junctions.
We also allow bas-relief textures to be modelled with a flat plane so that the
repeatability can be tested under non-affine warping.
A margin of error must be allowed because:

1. The alignment is not perfect.

 Machine learning for high-speed corner detection 9

Detect features in frame 1 Detect features in frame 2

Warp frame 1 compare

to match frame 2 warped feature
positions to detected
features in frame 2

Figure 2. Repeatability is tested by checking if the same real-world features are de-
tected in different views. A geometric model is used to compute where the features
reproject to.

2. The model is not perfect.

3. The camera model (especially regarding radial distortion) is not perfect.
4. The detector may find a maximum on a slightly different part of the corner.
This becomes more likely as the change in viewpoint and hence change in
shape of the corner become large.
Instead of using fiducial markers, the 3D model is aligned to the scene by hand
and this is then optimised using a blend of simulated annealing and gradient
descent to minimise the SSD between all pairs of frames and reprojections.
To compute the SSD between frame i and reprojected frame j, the position
of all points in frame j are found in frame i. The images are then bandpass
filtered. High frequencies are removed to reduce noise, while low frequencies are
removed to reduce the impact of lighting changes. To improve the speed of the
system, the SSD is only computed using 1000 random points (as opposed to
every point).
The datasets used are shown in Figure 3, Figure 4 and Figure 5. With these
datasets, we have tried to capture a wide range of corner types (geometric and
textural).
The repeatability is computed as the number of corners per frame is varied.
For comparison we also include a scattering of random points as a baseline mea-
sure, since in the limit if every pixel is detected as a corner, then the repeatability
is 100%.
To test robustness to image noise, increasing amounts of Gaussian noise were
added to the bas-relief dataset. It should be noted that the noise added is in
10 Edward Rosten and Tom Drummond

Figure 3. Box dataset: photographs taken of a test rig (consisting of photographs

pasted to the inside of a cuboid) with strong changes of perspective, changes in scale and
large amounts of radial distortion. This tests the corner detectors on planar textures.

Figure 4. Maze dataset: photographs taken of a prop used in an augmented reality

application. This set consists of textural features undergoing projective warps as well
as geometric features. There are also significant changes of scale.

addition to the significant amounts of camera noise already present (from thermal
noise, electrical interference, and etc).

4 Results and Discussion

Shi and Tomasi [7], derive their result for better feature detection on the as-
sumption that the deformation of the features is affine. In the box and maze
datasets, this assumption holds and can be seen in Figure 6B and Figure 6C the
detector outperforms the Harris detector. In the bas-relief dataset, this assump-
tion does not hold, and interestingly, the Harris detector outperforms Shi and
Tomasi detector in this case.
Mikolajczyk and Schmid [15] evaluate the repeatability of the Harris-Laplace
detector evaluated using the method in [34], where planar scenes are examined.
The results show that Harris-Laplace points outperform both DoG points and
Harris points in repeatability. For the box dataset, our results verify that this
is correct for up to about 1000 points per frame (typical numbers, probably
commonly used); the results are somewhat less convincing in the other datasets,
where points undergo non-projective changes.
In the sample implementation of SIFT[35], approximately 1000 points are
generated on the images from the test sets. We concur that this a good choice for
the number of features since this appears to be roughly where the repeatability
curve for DoG features starts to flatten off.
Smith and Brady[27] claim that the SUSAN corner detector performs well
in the presence of noise since it does not compute image derivatives, and hence,
does not amplify noise. We support this claim: although the noise results show
 Machine learning for high-speed corner detection 11

Figure 5. Bas-relief dataset: the model is a flat plane, but there are many objects with
significant relief. This causes the appearance of features to change in a non affine way
from different viewpoints.

that the performance drops quite rapidly with increasing noise to start with, it
soon levels off and outperforms all but the DoG detector.
The big surprise of this experiment is that the FAST feature detectors, despite
being designed only for speed, outperform the other feature detectors on these
images (provided that more than about 200 corners are needed per frame). It can
be seen in Figure 6A, that the 9 point detector provides optimal performance,
hence only this and the original 12 point detector are considered in the remaining
graphs.
The DoG detector is remarkably robust to the presence of noise. Since convo-
lution is linear, the computation of DoG is equivalent to convolution with a DoG
kernel. Since this kernel is symmetric, this is equivalent to matched filtering for
objects with that shape. The robustness is achieved because matched filtering is
optimal in the presence of additive Gaussian noise[36].
FAST, however, is not very robust to the presence of noise. This is to be
expected: Since high speed is achieved by analysing the fewest pixels possible,
the detector’s ability to average out noise is reduced.

5 Conclusions
In this paper, we have used machine learning to derive a very fast, high quality
corner detector. It has the following advantages:
– It is many times faster than other existing corner detectors.
– High levels of repeatability under large aspect changes and for different kinds
of feature.
However, it also suffers from a number of disadvantages:
– It is not robust to high levels noise.
– It can respond to 1 pixel wide lines at certain angles, when the quantisation
of the circle misses the line.
– It is dependent on a threshold.
12 Edward Rosten and Tom Drummond

We were also able to verify a number of claims made in other papers using the
method for evaluating the repeatability of corners and have shown the impor-
tance of using more than just planar scenes in this evaluation.
The corner detection code is made available from
http://mi.eng.cam.ac.uk/˜er258/work/fast.html
and
http://savannah.nongnu.org/projects/libcvd
and the data sets used for repeatability are available from
http://mi.eng.cam.ac.uk/˜er258/work/datasets.html

References

1. Schmid, C., Mohr, R., Bauckhage, C.: Evaluation of interest point detectors. In-
ternational Journal of Computer Vision 37 (2000) 151–172
2. Rosten, E., Drummond, T.: Fusing points and lines for high performance tracking.
In: 10th IEEE International Conference on Computer Vision. Volume 2., Beijing,
China, Springer (2005) 1508–1515
3. Rosten, E., Reitmayr, G., Drummond, T.: Real-time video annotations for aug-
mented reality. In: International Symposium on Visual Computing. (2005)
4. Moravec, H.: Obstacle avoidance and navigation in the real world by a seeing robot
rover. In: tech. report CMU-RI-TR-80-03, Robotics Institute, Carnegie Mellon Uni-
versity & doctoral dissertation, Stanford University. Carnegie Mellon University
(1980) Available as Stanford AIM-340, CS-80-813 and republished as a Carnegie
Mellon University Robotics Institue Technical Report to increase availability.
5. Harris, C., Stephens, M.: A combined corner and edge detector. In: Alvey Vision
Conference. (1988) 147–151
6. Noble, J.A.: Finding corners. Image and Vision Computing 6 (1988) 121–128
7. Shi, J., Tomasi, C.: Good features to track. In: 9th IEEE Conference on Computer
Vision and Pattern Recognition, Springer (1994)
8. Noble, A.: Descriptions of image surfaces. PhD thesis, Department of Engineering
Science, University of Oxford. (1989)
9. Kenney, C.S., Manjunath, B.S., Zuliani, M., Hewer, M.G.A., Nevel, A.V.: A con-
dition number for point matching with application to registration and postreg-
istration error estimation. IEEE Transactions on Pattern Analysis and Machine
Intelligence 25 (2003) 1437–1454
10. Zuliani, M., Kenney, C., Manjunath, B.: A mathematical comparison of point de-
tectors. In: Second IEEE Image and Video Registration Workshop (IVR), Wash-
ington DC, USA (2004)
11. Zheng, Z., Wang, H., Teoh, E.K.: Analysis of gray level corner detection. Pattern
Recognition Letters 20 (1999) 149–162
12. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. Interna-
tional Journal of Computer Vision 60 (2004) 91–110
13. James L. Crowley, O.R.: Fast computation of characteristic scale using a half
octave pyramid. In: Scale Space 03: 4th International Conference on Scale-Space
theories in Computer Vision, Isle of Skye, Scotland, UK (2003)
14. Mikolajczyk, K., Schmid, C.: An affine invariant interest point detector. In: Eu-
ropean Conference on Computer Vision, Springer (2002) 128–142 Copenhagen.
 Machine learning for high-speed corner detection 13

15. Mikolajczyk, K., Schmid, C.: Indexing based on scale invariant interest points.
In: 8th IEEE International Conference on Computer Vision, Vancouver, Canada,
Springer (2001) 525–531
16. Brown, M., Lowe, D.G.: Invariant features from interest point groups. In: 13th
British Machine Vision Conference, Cardiff, British Machine Vision Assosciation
(2002) 656–665
17. Schaffalitzky, F., Zisserman, A.: Multi-view matching for unordered image sets, or
How do I organise my holiday snaps? In: 7th Euproean Conference on Computer
Vision, Springer (2002) 414–431
18. Rutkowski, W.S., Rosenfeld, A.: A comparison of corner detection techniques for
chain coded curves. Technical Report 623, Maryland University (1978)
19. Langridge, D.J.: Curve encoding and detection of discontinuities. Computer Vision,
Graphics and Image Processing 20 (1987) 58–71
20. Medioni, G., Yasumoto, Y.: Corner detection and curve representation using cubic
b-splines. Computer Vision, Graphics and Image Processing 39 (1987) 279–290
21. Mokhtarian, F., Suomela, R.: Robust image corner detection through curvature
scale space. IEEE Transactions on Pattern Analysis and Machine Intelligence 20
(1998) 1376–1381
22. Haralick, R.M., Shapiro, L.G.: Computer and robot vision. Volume 1. Adison-
Wesley (1993)
23. Cooper, J., Venkatesh, S., Kitchen, L.: Early jump-out corner detectors. IEEE
Transactions on Pattern Analysis and Machine Intelligence 15 (1993) 823–828
24. Wang, H., Brady, M.: Real-time corner detection algorithm for motion estimation.
Image and Vision Computing 13 (1995) 695–703
25. Kitchen, L., Rosenfeld, A.: Gray-level corner detection. Pattern Recognition Let-
ters 1 (1982) 95–102
26. Guiducci, A.: Corner characterization by differential geometry techniques. Pattern
Recognition Letters 8 (1988) 311–318
27. Smith, S.M., Brady, J.M.: SUSAN - a new approach to low level image processing.
International Journal of Computer Vision 23 (1997) 45–78
28. Trajkovic, M., Hedley, M.: Fast corner detection. Image and Vision Computing
16 (1998) 75–87
29. Loy, G., Zelinsky, A.: A fast radial symmetry transform for detecting points of
interest. In: 7th Euproean Conference on Computer Vision, Springer (2002) 358–
368
30. Dias, P., Kassim, A., Srinivasan, V.: A neural network based corner detection
method. In: IEEE International Conference on Neural Networks. Volume 4., Perth,
WA, Australia (1995) 2116–2120
31. Quinlan, J.R.: Induction of decision trees. Machine Learning 1 (1986) 81–106
32. Smith, S.M.: http://www.fmrib.ox.ac.uk/~steve/susan/susan2l.c (Accessed
2005)
33. Cootes, T.F., Taylor, C., eds.: Performenace Evaluation of Corner Detection Al-
gorithms under Affine and Similarity Transforms. In Cootes, T.F., Taylor, C.,
eds.: 12th British Machine Vision Conference, Manchester, British Machine Vision
Assosciation (2001)
34. Schmid, C., Mohr, R., Bauckhage, C.: Comparing and evaluating interest points.
In: 6th IEEE International Conference on Computer Vision, Bombay, India,
Springer (1998) 230–235
35. Lowe, D.G.: Demo software: Sift keypoint detector.
http://www.cs.ubc.ca/~lowe/keypoints/ (Accessed 2005)
36. Sklar, B.: Digital Communications. Prentice Hall (1988)
 14 Edward Rosten and Tom Drummond

A
Comparison of FAST detectors Legend for B–E
80 Fast 9
Fast 12
70
Harris
60 Shi & Tomasi
Repeatability %

50
DoG
Harris−Laplace
Fast 9
40
Fast 10 SUSAN
30 Fast 11 Random
Fast 12
20 Fast 13
Fast 14
10 Fast 15
Fast 16
0
0 500 1000 1500 2000
Corners per frame
B C
Box dataset Maze dataset
90 80
80 70
70
60
Repeatability %

Repeatability %

60
50
50
40
40
30
30
20 20

10 10

0 0
0 500 1000 1500 2000 0 500 1000 1500 2000
Corners per frame Corners per frame
D E
Bas-relief dataset Additive noise
80 70

70 60

60 50
Repeatability %
Repeatability %

50
40
40
30
30
20
20

10 10

0 0
0 500 1000 1500 2000 0 10 20 30 40 50
Corners per frame Noise σ

Figure 6. A: A comparison of the FAST detectors shown that n = 9 is the most

repeatable. For n ≤ 8, the detector starts to respond strongly to edges. B, C, D:
Repeatability results for the three datasets as the number of features per frame is
varied. D: repeatability results for the bas-relief data set (500 features per frame) as
the amount of Gaussian noise added to the images is varied. For FAST and SUSAN,
the number of features can not be chosen arbitrarily; the closest approximation to 500
features per frame achievable is used.

View publication stats