Conference PaperPDF Available

ORB: an efficient alternative to SIFT or SURF


Abstract and Figures

Feature matching is at the base of many computer vision problems, such as object recognition or structure from motion. Current methods rely on costly descriptors for detection and matching. In this paper, we propose a very fast binary descriptor based on BRIEF, called ORB, which is rotation invariant and resistant to noise. We demonstrate through experiments how ORB is at two orders of magnitude faster than SIFT, while performing as well in many situations. The efficiency is tested on several real-world applications, including object detection and patch-tracking on a smart phone.
Content may be subject to copyright.
ORB: an efficient alternative to SIFT or SURF
Ethan Rublee Vincent Rabaud Kurt Konolige Gary Bradski
Willow Garage, Menlo Park, California
Feature matching is at the base of many computer vi-
sion problems, such as object recognition or structure from
motion. Current methods rely on costly descriptors for de-
tection and matching. In this paper, we propose a very fast
binary descriptor based on BRIEF, called ORB, which is
rotation invariant and resistant to noise. We demonstrate
through experiments how ORB is at two orders of magni-
tude faster than SIFT, while performing as well in many
situations. The efficiency is tested on several real-world ap-
plications, including object detection and patch-tracking on
a smart phone.
1. Introduction
The SIFT keypoint detector and descriptor [17], al-
though over a decade old, have proven remarkably success-
ful in a number of applications using visual features, in-
cluding object recognition [17], image stitching [28], visual
mapping [25], etc. However, it imposes a large computa-
tional burden, especially for real-time systems such as vi-
sual odometry, or for low-power devices such as cellphones.
This has led to an intensive search for replacements with
lower computation cost; arguably the best of these is S URF
[2]. There has also been research aimed at speeding up the
computation of SIFT, most notably with GPU devices [26].
In this paper, we propose a computationally-efficient re-
placement to SIFT that has similar matching performance,
is less affected by image noise, and is capable of being used
for real-time performance. Our main motivation is to en-
hance many common image-matching applications, e.g., to
enable low-power devices without GPU acceleration to per-
form panorama stitching and patch tracking, and to reduce
the time for feature-based object detection on standard PCs.
Our descriptor performs as well as SIFT on these tasks (and
better than SURF), while being almost two orders of mag-
nitude faster.
Our proposed feature builds on the well-known FAST
keypoint detector [23] and the recently-developed BRIEF
descriptor [6]; for this reason we call it ORB (Oriented
Figure 1. Typical matching result using ORB on real-world im-
ages with viewpoint change. Green lines are valid matches; red
circles indicate unmatched points.
FAST and Rotated BRIEF). Both these techniques are at-
tractive because of their good performance and low cost.
In this paper, we address several limitations of these tech-
niques vis-a-vis SIFT, most notably the lack of rotational
invariance in BRIEF. Our main contributions are:
The addition of a fast and accurate orientation compo-
nent to FAST.
The efficient computation of oriented BRIEF features.
Analysis of variance and correlation of oriented
BRIEF features.
A learning method for de-correlating BRIEF features
under rotational invariance, leading to better perfor-
mance in nearest-neighbor applications.
To validate ORB, we perform experiments that test the
properties of ORB relative to SIFT and SURF, for both
raw matching ability, and performance in image-matching
applications. We also illustrate the efficiency of ORB
by implementing a patch-tracking application on a smart
phone. An additional benefit of ORB is that it is free from
the licensing restrictions of SIFT and SURF.
2. Related Work
Keypoints FAST and its variants [23,24] are the method
of choice for finding keypoints in real-time systems that
match visual features, for example, Parallel Tracking and
Mapping [13]. It is efficient and finds reasonable corner
keypoints, although it must be augmented with pyramid
schemes for scale [14], and in our case, a Harris corner filter
[11] to reject edges and provide a reasonable score.
Many keypoint detectors include an orientation operator
(SIFT and SURF are two prominent examples), but FAST
does not. There are various ways to describe the orientation
of a keypoint; many of these involve histograms of gradient
computations, for example in SIFT [17] and the approxi-
mation by block patterns in SURF [2]. These methods are
either computationally demanding, or in the case of SURF,
yield poor approximations. The reference paper by Rosin
[22] gives an analysis of various ways of measuring orienta-
tion of corners, and we borrow from his centroid technique.
Unlike the orientation operator in SIFT, which can have
multiple value on a single keypoint, the centroid operator
gives a single dominant result.
Descriptors BRIEF [6] is a recent feature descriptor that
uses simple binary tests between pixels in a smoothed image
patch. Its performance is similar to SIFT in many respects,
including robustness to lighting, blur, and perspective dis-
tortion. However, it is very sensitive to in-plane rotation.
BRIEF grew out of research that uses binary tests to
train a set of classification trees [4]. Once trained on a set
of 500 or so typical keypoints, the trees can be used to re-
turn a signature for any arbitrary keypoint [5]. In a similar
manner, we look for the tests least sensitive to orientation.
The classic method for finding uncorrelated tests is Princi-
pal Component Analysis; for example, it has been shown
that PCA for SIFT can help remove a large amount of re-
dundant information [12]. However, the space of possible
binary tests is too big to perform PCA and an exhaustive
search is used instead.
Visual vocabulary methods [21,27] use offline clustering
to find exemplars that are uncorrelated and can be used in
matching. These techniques might also be useful in finding
uncorrelated binary tests.
The closest system to ORB is [3], which proposes a
multi-scale Harris keypoint and oriented patch descriptor.
This descriptor is used for image stitching, and shows good
rotational and scale invariance. It is not as efficient to com-
pute as our method, however.
3. oFAST: FAST Keypoint Orientation
FAST features are widely used because of their compu-
tational properties. However, FAST features do not have an
orientation component. In this section we add an efficiently-
computed orientation.
3.1. FAST Detector
We start by detecting FAST points in the image. FAST
takes one parameter, the intensity threshold between the
center pixel and those in a circular ring about the center.
We use FAST-9 (circular radius of 9), which has good per-
FAST does not produce a measure of cornerness, and we
have found that it has large responses along edges. We em-
ploy a Harris corner measure [11] to order the FAST key-
points. For a target number Nof keypoints, we first set the
threshold low enough to get more than Nkeypoints, then
order them according to the Harris measure, and pick the
top Npoints.
FAST does not produce multi-scale features. We employ
a scale pyramid of the image, and produce FAST features
(filtered by Harris) at each level in the pyramid.
3.2. Orientation by Intensity Centroid
Our approach uses a simple but effective measure of cor-
ner orientation, the intensity centroid [22]. The intensity
centroid assumes that a corner’s intensity is offset from its
center, and this vector may be used to impute an orientation.
Rosin defines the moments of a patch as:
mpq =X
xpyqI(x, y),(1)
and with these moments we may find the centroid:
m00 (2)
We can construct a vector from the corner’s center, O, to the
centroid, ~
OC. The orientation of the patch then simply is:
θ=atan2(m01, m10 ),(3)
where atan2 is the quadrant-aware version of arctan. Rosin
mentions taking into account whether the corner is dark or
light; however, for our purposes we may ignore this as the
angle measures are consistent regardless of the corner type.
To improve the rotation invariance of this measure we
make sure that moments are computed with xand yre-
maining within a circular region of radius r. We empirically
choose rto be the patch size, so that that xand yrun from
[r, r]. As |C|approaches 0, the measure becomes unsta-
ble; with FAST corners, we have found that this is rarely the
We compared the centroid method with two gradient-
based measures, BIN and MAX. In both cases, Xand
Ygradients are calculated on a smoothed image. MAX
chooses the largest gradient in the keypoint patch; BIN
forms a histogram of gradient directions at 10 degree inter-
vals, and picks the maximum bin. BIN is similar to the SIFT
algorithm, although it picks only a single orientation. The
variance of the orientation in a simulated dataset (in-plane
rotation plus added noise) is shown in Figure 2. Neither of
the gradient measures performs very well, while the cen-
troid gives a uniformly good orientation, even under large
image noise.
0 5 10 15 20 25
Standard Deviation (degrees)
Image Intensity Noise (gaussian noise for an 8 bit image)
Standard Deviation in Angle Error
Figure 2. Rotation measure. The intensity centroid (IC) per-
forms best on recovering the orientation of artificially rotatednoisy
patches, compared to a histogram (BIN) and MAX method.
4. rBRIEF: Rotation-Aware Brief
In this section, we first introduce a steered BRIEF de-
scriptor, show how to compute it efficiently and demon-
strate why it actually performs poorly with rotation. We
then introduce a learning step to find less correlated binary
tests leading to the better descriptor rBRIEF, for which we
offer comparisons to SIFT and SURF.
4.1. Efficient Rotation of the BRIEF Operator
Brief overview of BRIEF
The BRIEF descriptor [6] is a bit string description of an
image patch constructed from a set of binary intensity tests.
Consider a smoothed image patch, p. A binary test τis
defined by:
τ(p;x,y) := 1 : p(x)<p(y)
0 : p(x)p(y),(4)
where p(x)is the intensity of pat a point x. The feature is
defined as a vector of nbinary tests:
fn(p) := X
Many different types of distributions of tests were consid-
ered in [6]; here we use one of the best performers, a Gaus-
sian distribution around the center of the patch. We also
choose a vector length n= 256.
It is important to smooth the image before performing
the tests. In our implementation, smoothing is achieved us-
ing an integral image, where each test point is a 5×5sub-
window of a 31 ×31 pixel patch. These were chosen from
our own experiments and the results in [6].
Steered BRIEF
We would like to allow BRIEF to be invariant to in-plane
rotation. Matching performance of BRIEF falls off sharply
for in-plane rotation of more than a few degrees (see Figure
7). Calonder [6] suggests computing a BRIEF descriptor
for a set of rotations and perspective warps of each patch,
Number of Bits
Bit Response Mean
Histogram of Descriptor Bit Mean Values
steered BRIEF
Figure 3. Distribution of means for feature vectors: BRIEF, steered
BRIEF (Section 4.1), and rBRIEF (Section 4.3). The X axis is the
distance to a mean of 0.5
but this solution is obviously expensive. A more efficient
method is to steer BRIEF according to the orientation of
keypoints. For any feature set of nbinary tests at location
(xi,yi), define the 2×nmatrix
Using the patch orientation θand the corresponding rotation
matrix Rθ, we construct a “steered” version Sθof S:
Now the steered BRIEF operator becomes
gn(p, θ) := fn(p)|(xi,yi)Sθ(6)
We discretize the angle to increments of 2π/30 (12 de-
grees), and construct a lookup table of precomputed BRIEF
patterns. As long at the keypoint orientation θis consistent
across views, the correct set of points Sθwill be used to
compute its descriptor.
4.2. Variance and Correlation
One of the pleasing properties of BRIEF is that each bit
feature has a large variance and a mean near 0.5. Figure 3
shows the spread of means for a typical Gaussian BRIEF
pattern of 256 bits over 100k sample keypoints. A mean
of 0.5 gives the maximum sample variance 0.25 for a bit
feature. On the other hand, once BRIEF is oriented along
the keypoint direction to give steered BRIEF, the means are
shifted to a more distributed pattern (again, Figure 3). One
way to understand this is that the oriented corner keypoints
present a more uniform appearance to binary tests.
High variance makes a feature more discriminative, since
it responds differentially to inputs. Another desirable prop-
erty is to have the tests uncorrelated, since then each test
will contribute to the result. To analyze the correlation and
0 5 10 15 20 25 30 35 40
Steered BRIEF
Figure 4. Distribution of eigenvalues in the PCA decomposition
over 100k keypoints of three feature vectors: BRIEF, steered
BRIEF (Section 4.1), and rBRIEF (Section 4.3).
0 64 128 192 256
Relative Frequency
Descriptor Distance
Distance Distribution
steered BRIEF
Figure 5. The dotted lines show the distances of a keypoint to out-
liers, while the solid lines denote the distances only between inlier
matches for three feature vectors: BRIEF, steered BRIEF (Section
4.1), and rBRIEF (Section 4.3).
variance of tests in the BRIEF vector, we looked at the re-
sponse to 100k keypoints for BRIEF and steered BRIEF.
The results are shown in Figure 4. Using PCA on the data,
we plot the highest 40 eigenvalues (after which the two de-
scriptors converge). Both BRIEF and steered BRIEF ex-
hibit high initial eigenvalues, indicating correlation among
the binary tests – essentially all the information is contained
in the first 10 or 15 components. Steered BRIEF has signif-
icantly lower variance, however, since the eigenvalues are
lower, and thus is not as discriminative. Apparently BRIEF
depends on random orientation of keypoints for good per-
formance. Another view of the effect of steered BRIEF is
shown in the distance distributions between inliers and out-
liers (Figure 5). Notice that for steered BRIEF, the mean for
outliers is pushed left, and there is more of an overlap with
the inliers.
4.3. Learning Good Binary Features
To recover from the loss of variance in steered BRIEF,
and to reduce correlation among the binary tests, we de-
velop a learning method for choosing a good subset of bi-
nary tests. One possible strategy is to use PCA or some
other dimensionality-reduction method, and starting from a
large set of binary tests, identify 256 new features that have
high variance and are uncorrelated over a large training set.
However, since the new features are composed from a larger
number of binary tests, they would be less efficient to com-
pute than steered BRIEF. Instead, we search among all pos-
sible binary tests to find ones that both have high variance
(and means close to 0.5), as well as being uncorrelated.
The method is as follows. We first setup a training set of
some 300k keypoints, drawn from images in the PASCAL
2006 set [8]. We also enumerate all possible binary tests
drawn from a 31 ×31 pixel patch. Each test is a pair of 5×5
sub-windows of the patch. If we note the width of our patch
as wp= 31 and the width of the test sub-window as wt= 5,
then we have N= (wpwt)2possible sub-windows. We
would like to select pairs of two from these, so we have N
binary tests. We eliminate tests that overlap, so we end up
with M= 205590 possible tests. The algorithm is:
1. Run each test against all training patches.
2. Order the tests by their distance from a mean of 0.5,
forming the vector T.
3. Greedy search:
(a) Put the first test into the result vector R and re-
move it from T.
(b) Take the next test from T, and compare it against
all tests in R. If its absolute correlation is greater
than a threshold, discard it; else add it to R.
(c) Repeat the previous step until there are 256 tests
in R. If there are fewer than 256, raise the thresh-
old and try again.
This algorithm is a greedy search for a set of uncorrelated
tests with means near 0.5. The result is called rBRIEF.
rBRIEF has significant improvement in the variance and
correlation over steered BRIEF (see Figure 4). The eigen-
values of PCA are higher, and they fall off much less
quickly. It is interesting to see the high-variance binary tests
produced by the algorithm (Figure 6). There is a very pro-
nounced vertical trend in the unlearned tests (left image),
which are highly correlated; the learned tests show better
diversity and lower correlation.
4.4. Evaluation
We evaluate the combination of oFAST and rBRIEF,
which we call ORB, using two datasets: images with syn-
thetic in-plane rotation and added Gaussian noise, and a
real-world dataset of textured planar images captured from
different viewpoints. For each reference image, we compute
the oFAST keypoints and rBRIEF features, targeting 500
keypoints per image. For each test image (synthetic rotation
or real-world viewpoint change), we do the same, then per-
form brute-force matching to find the best correspondence.
Figure 6. A subset of the binary tests generated by considering
high-variance under orientation (left) and by running the learning
algorithm to reduce correlation (right). Note the distribution of the
tests around the axis of the keypoint orientation, which is pointing
up. The color coding shows the maximum pairwise correlation of
each test, with black and purple being the lowest. The learned tests
clearly have a better distribution and lower correlation.
0 45 90 135 180 225 270 315 360
Percentage of Inliers
Angle of Rotation (Degrees)
Percentage of Inliers considering In Plane Rotation
Figure 7. Matching performance of SIFT, SURF, BRIEF with
FAST, and ORB (oFAST +rBRIEF) under synthetic rotations
with Gaussian noise of 10.
The results are given in terms of the percentage of correct
matches, against the angle of rotation.
Figure 7shows the results for the synthetic test set with
added Gaussian noise of 10. Note that the standard BRIEF
operator falls off dramatically after about 10 degrees. SIFT
outperforms SURF, which shows quantization effects at 45-
degree angles due to its Haar-wavelet composition. ORB
has the best performance, with over 70% inliers.
ORB is relatively immune to Gaussian image noise, un-
like SIFT. If we plot the inlier performance vs. noise, SIFT
exhibits a steady drop of 10% with each additional noise
increment of 5. ORB also drops, but at a much lower rate
(Figure 8).
To test ORB on real-world images, we took two sets of
images, one our own indoor set of highly-textured mag-
azines on a table (Figure 9), the other an outdoor scene.
The datasets have scale, viewpoint, and lighting changes.
Running a simple inlier/outlier test on this set of images,
90 180 270 360
Percentage of Inliers
Angle of Rotation (Degrees)
Comparison of SIFT and rBRIEF considering Gaussian Intensity Noise
Figure 8. Matching behavior under noise for SIFT and rBRIEF.
The noise levels are 0, 5, 10, 15, 20, and 25. SIFT performance
degrades rapidly, while rBRIEF is relatively unaffected.
Figure 9. Real world data of a table full of magazines and an out-
door scene. The images in the first column are matched to those in
the second. The last column is the resulting warp of the first onto
the second.
we measure the performance of ORB relative to SIFT and
SURF. The test is performed in the following manner:
1. Pick a reference view V0.
2. For all Vi, find a homographic warp Hi0that maps
3. Now, use the Hi0as ground truth for descriptor
matches from SIFT, SURF, and ORB.
inlier % Npoints
ORB 36.180 548.50
SURF 38.305 513.55
SIFT 34.010 584.15
Boat ORB 45.8 789
SURF 28.6 795
SIFT 30.2 714
ORB outperforms SIFT and SURF on the outdoor dataset.
It is about the same on the indoor set; [6] noted that blob-
detection keypoints like SIFT tend to be better on graffiti-
type images.
Figure 10. Two different datasets (7818 images from the PASCAL
2009 dataset [9] and 9144 low resolution images from the Caltech
101 [29]) are used to train LSH on the BR IEF, steered BR IEF and
rBRIEF descriptors. The training takes less than 2 minutes and
is limited by the disk IO. rBRIEF gives the most homogeneous
buckets by far, thus improving the query speed and accuracy.
5. Scalable Matching of Binary Features
In this section we show that ORB outperforms
SIFT/SURF in nearest-neighbor matching over large
databases of images. A critical part of ORB is the recovery
of variance, which makes NN search more efficient.
5.1. Locality Sensitive Hashing for rBrief
As rBRIEF is a binary pattern, we choose Locality Sen-
sitive Hashing [10] as our nearest neighbor search. In LSH,
points are stored in several hash tables and hashed in differ-
ent buckets. Given a query descriptor, its matching buckets
are retrieved and its elements compared using a brute force
matching. The power of that technique lies in its ability
to retrieve nearest neighbors with a high probability given
enough hash tables.
For binary features, the hash function is simply a subset
of the signature bits: the buckets in the hash tables contain
descriptors with a common sub-signature. The distance is
the Hamming distance.
We use multi-probe LSH [18] which improves on the
traditional LSH by looking at neighboring buckets in which
a query descriptor falls. While this could result in more
matches to check, it actually allows for a lower number of
tables (and thus less RAM usage) and a longer sub-signature
and therefore smaller buckets.
5.2. Correlation and Leveling
rBRIEF improves the speed of LSH by making the
buckets of the hash tables more even: as the bits are less
correlated, the hash function does a better job at partitioning
Figure 11. Speed vs. accuracy. The descriptors are tested on
warped versions of the images they were trained on. We used 1,
2 and 3 kd-trees for SIFT (the autotuned FLANN kd-tree gave
worse performance), 4 to 20 hash tables for rBRIEF and 16 to 40
tables for steered BRIEF (both with a sub-signature of 16 bits).
Nearest neighbors were searched over 1.6M entries for SIFT and
1.8M entries for rBRIEF.
the data. As shown in Figure 10, buckets are much smaller
in average compared to steered BRIEF or normal BRIEF.
5.3. Evaluation
We compare the performance of rBRIEF LSH with kd-
trees of SIFT features using FLANN [20]. We train the dif-
ferent descriptors on the Pascal 2009 dataset and test them
on sampled warped versions of those images using the same
affine transforms as in [1].
Our multi-probe LSH uses bitsets to speedup the pres-
ence of keys in the hash maps. It also computes the Ham-
ming distance between two descriptors using an SSE 4.2
optimized popcount.
Figure 11 establishes a correlation between the speed
and the accuracy of kd-trees with SIFT (SURF is equiv-
alent) and LSH with rBRIEF. A successful match of the
test image occurs when more than 50 descriptors are found
in the correct database image. We notice that LSH is faster
than the kd-trees, most likely thanks to its simplicity and the
speed of the distance computation. LSH also gives more
flexibility with regard to accuracy, which can be interesting
in bag-of-feature approaches [21,27]. We can also notice
that the steered BRIEF is much slower due to its uneven
6. Applications
6.1. Benchmarks
One emphasis for ORB is the efficiency of detection and
description on standard CPUs. Our canonical ORB detec-
tor uses the oFAST detector and rBRIEF descriptor, each
computed separately on five scales of the image, with a scal-
ing factor of 2. We used an area-based interpolation for
efficient decimation.
The ORB system breaks down into the following times
per typical frame of size 640x480. The code was executed
in a single thread running on an Intel i7 2.8 GHz processor:
Time (ms) 4.43 8.68 2.12
When computing ORB on a set of 2686 images at 5
scales, it was able to detect and compute over 2×106fea-
tures in 42 seconds. Comparing to SIFT and SURF on the
same data, for the same number of features (roughly 1000),
and the same number of scales, we get the following times:
Time per frame (ms) 15.3 217.3 5228.7
These times were averaged over 24 640x480 images from
the Pascal dataset [9]. ORB is an order of magnitude faster
than SURF, and over two orders faster than SIFT.
6.2. Textured object detection
We apply rBRIEF to object recognition by implement-
ing a conventional object recognition pipeline similar to
[19]: we first detect oFAST features and rBRIEF de-
scriptors, match them to our database, and then perform
PROSAC [7] and EPnP [16] to have a pose estimate.
Our database contains 49 household objects, each taken
under 24 views with a 2D camera and a Kinect device from
Microsoft. The testing data consists of 2D images of sub-
sets of those same objects under different view points and
occlusions. To have a match, we require that descriptors are
matched but also that a pose can be computed. In the end,
our pipeline retrieves 61% of the objects as shown in Figure
12.The algorithm handles a database of 1.2M descriptors
in 200MB and has timings comparable to what we showed
earlier (14 ms for detection and 17ms for LSH matching in
average). The pipeline could be sped up considerably by not
matching all the query descriptors to the training data but
our goal was only to show the feasibility of object detection
with ORB.
6.3. Embedded real-time feature tracking
Tracking on the phone involves matching the live frames
to a previously captured keyframe. Descriptors are stored
with the keyframe, which is assumed to contain a planar
surface that is well textured. We run ORB on each incom-
ing frame, and proced with a brute force descriptor match-
ing against the keyframe. The putative matches from the
descriptor distance are used in a PROSAC best fit homog-
raphy H.
Figure 12. Two images of our textured obejct recognition with
pose estimation. The blue features are the training features super-
imposed on the query image to indicate that the pose of the object
was found properly. Axes are also displayed for each object as
well as a pink label. Top image misses two objects; all are found
in the bottom one.
While there are real-time feature trackers that can run on
a cellphone [15], they usually operate on very small images
(e.g., 120x160) and with very few features. Systems com-
parable to ours [30] typically take over 1 second per image.
We were able to run ORB with 640 ×480 resolution at 7
Hz on a cellphone with a 1GHz ARM chip and 512 MB of
RAM. The OpenCV port for Android was used for the im-
plementation. These are benchmarks for about 400 points
per image:
ORB Matching HFit
Time (ms) 66.6 72.8 20.9
7. Conclusion
In this paper, we have defined a new oriented descrip-
tor, ORB, and demonstrated its performance and efficiency
relative to other popular features. The investigation of vari-
ance under orientation was critical in constructing ORB
and de-correlating its components, in order to get good per-
formance in nearest-neighbor applications. We have also
contributed a BSD licensed implementation of ORB to the
community, via OpenCV 2.3.
One of the issues that we have not adequately addressed
here is scale invariance. Although we use a pyramid scheme
for scale, we have not explored per keypoint scale from
depth cues, tuning the number of octaves, etc.. Future work
also includes GPU/SSE optimization, which could improve
LSH by another order of magnitude.
[1] M. Aly, P. Welinder, M. Munich, and P. Perona. Scaling
object recognition: Benchmark of current state of the art
techniques. In First IEEE Workshop on Emergent Issues in
Large Amounts of Visual Data (WS-LAVD), IEEE Interna-
tional Conference on Computer Vision (ICCV), September
2009. 6
[2] H. Bay, T. Tuytelaars, and L. Van Gool. Surf: Speeded up ro-
bust features. In European Conference on Computer Vision,
May 2006. 1,2
[3] M. Brown, S. Winder, and R. Szeliski. Multi-image match-
ing using multi-scale oriented patches. In Computer Vision
and Pattern Recognition, pages 510–517, 2005. 2
[4] M. Calonder, V. Lepetit, and P. Fua. Keypoint signatures for
fast learning and recognition. In European Conference on
Computer Vision, 2008. 2
[5] M. Calonder, V. Lepetit, K. Konolige, P. Mihelich, and
P. Fua. High-speed keypoint description and matching us-
ing dense signatures. In Under review, 2009. 2
[6] M. Calonder, V. Lepetit, C. Strecha, and P. Fua. Brief: Bi-
nary robust independent elementary features. In In European
Conference on Computer Vision, 2010. 1,2,3,5
[7] O. Chum and J. Matas. Matching with PROSAC - pro-
gressive sample consensus. In C. Schmid, S. Soatto, and
C. Tomasi, editors, Proc. of Conference on Computer Vision
and Pattern Recognition (CVPR), volume 1, pages 220–226,
Los Alamitos, USA, June 2005. IEEE Computer Society. 7
[8] M. Everingham. The PASCAL Visual Ob-
ject Classes Challenge 2006 (VOC2006) Results.
[9] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn,
and A. Zisserman. The PASCAL Visual Object Classes
Challenge 2009 (VOC2009) Results. http://www.pascal-
[10] A. Gionis, P. Indyk, and R. Motwani. Similarity search in
high dimensions via hashing. In M. P. Atkinson, M. E. Or-
lowska, P. Valduriez, S. B. Zdonik, and M. L. Brodie, editors,
VLDB’99, Proceedings of 25th International Conference on
Very Large Data Bases, September 7-10, 1999, Edinburgh,
Scotland, UK, pages 518–529. Morgan Kaufmann, 1999. 6
[11] C. Harris and M. Stephens. A combined corner and edge
detector. In Alvey Vision Conference, pages 147–151, 1988.
[12] Y. Ke and R. Sukthankar. Pca-sift: A more distinctive rep-
resentation for local image descriptors. In Computer Vision
and Pattern Recognition, pages 506–513, 2004. 2
[13] G. Klein and D. Murray. Parallel tracking and mapping for
small AR workspaces. In Proc. Sixth IEEE and ACM Inter-
national Symposium on Mixed and Augmented Reality (IS-
MAR’07), Nara, Japan, November 2007. 1
[14] G. Klein and D. Murray. Improving the agility of keyframe-
based SLAM. In European Conference on Computer Vision,
2008. 2
[15] G. Klein and D. Murray. Parallel tracking and mapping on a
camera phone. In Proc. Eigth IEEE and ACM International
Symposium on Mixed and Augmented Reality (ISMAR’09),
Orlando, October 2009. 7
[16] V. Lepetit, F. Moreno-Noguer, and P. Fua. EPnP: An accurate
O(n) solution to the pnp problem. Int. J. Comput. Vision,
81:155–166, February 2009. 7
[17] D. G. Lowe. Distinctive image features from scale-invariant
keypoints. International Journal of Computer Vision,
60(2):91–110, 2004. 1,2
[18] Q. Lv, W. Josephson, Z. Wang, M. Charikar, and K. Li.
Multi-probe LSH: efficient indexing for high-dimensional
similarity search. In Proceedings of the 33rd international
conference on Very large data bases, VLDB ’07, pages 950–
961. VLDB Endowment, 2007. 6
[19] M. Martinez, A. Collet, and S. S. Srinivasa. MOPED:
A Scalable and low Latency Object Recognition and Pose
Estimation System. In IEEE International Conference on
Robotics and Automation. IEEE, 2010. 7
[20] M. Muja and D. G. Lowe. Fast approximate nearest neigh-
bors with automatic algorithm configuration. VISAPP, 2009.
[21] D. Nist´er and H. Stew´enius. Scalable recognition with a vo-
cabulary tree. In CVPR, 2006. 2,6
[22] P. L. Rosin. Measuring corner properties. Computer Vision
and Image Understanding, 73(2):291 – 307, 1999. 2
[23] E. Rosten and T. Drummond. Machine learning for high-
speed corner detection. In European Conference on Com-
puter Vision, volume 1, 2006. 1
[24] E. Rosten, R. Porter, and T. Drummond. Faster and bet-
ter: A machine learning approach to corner detection. IEEE
Trans. Pattern Analysis and Machine Intelligence, 32:105–
119, 2010. 1
[25] S. Se, D. Lowe, and J. Little. Mobile robot localization and
mapping with uncertainty using scale-invariant visual land-
marks. International Journal of Robotic Research, 21:735–
758, August 2002. 1
[26] S. N. Sinha, J. michael Frahm, M. Pollefeys, and Y. Genc.
Gpu-based video feature tracking and matching. Technical
report, In Workshop on Edge Computing Using New Com-
modity Architectures, 2006. 1
[27] J. Sivic and A. Zisserman. Video google: A text retrieval
approach to object matching in videos. International Con-
ference on Computer Vision, page 1470, 2003. 2,6
[28] N. Snavely, S. M. Seitz, and R. Szeliski. Skeletal sets for
efficient structure from motion. In Proc. Computer Vision
and Pattern Recognition, 2008. 1
[29] G. Wang, Y. Zhang, and L. Fei-Fei. Using dependent regions
for object categorization in a generative framework, 2006. 6
[30] A. Weimert, X. Tan, and X. Yang. Natural feature detection
on mobile phones with 3D FAST. Int. J. of Virtual Reality,
9:29–34, 2010. 7
... Estimating the relative pose between two images is a fundamental vision problem [17], with applications including 3D understanding [22,34,57] and extended reality [30,35,38,73]. Early work focused on robust [14] fitting of models [17,19,31,42] on detected correspondences [2,32,54] between the images. This strategy can fail catastrophically with poor correspondence, which is especially frequent in the wide baseline setting, when the images have a substantial pose difference. ...
... Based on these observations, there has been much work applying learning to the problem. One line of attack [8,10,55,60,63] has been to follow the classic pipeline and replace classic correspondence methods [2,32,54] with learned ones. This approach is appealing since the learning method finds correspondence, an especially thorny challenge in the wide-baseline setting, and the conversion of correspondences to pose is done by a provably correct method [31,42]. ...
... Relative pose estimation from an image pair is a sufficiently broad problem to preclude a full account. We refer readers to [17], and focus on the closest works, which all follow a strategy of solving for pose given correspondences from local descriptors [2,32,54]. We revisit the 8-point algorithm [19,31] that maps correspondences to an Essential Matrix, which was invented by Longuet-Higgins and extended to Fundamental Matrices by [13,18]. ...
We present a simple baseline for directly estimating the relative pose (rotation and translation, including scale) between two images. Deep methods have recently shown strong progress but often require complex or multi-stage architectures. We show that a handful of modifications can be applied to a Vision Transformer (ViT) to bring its computations close to the Eight-Point Algorithm. This inductive bias enables a simple method to be competitive in multiple settings, often substantially improving over the state of the art with strong performance gains in limited data regimes.
... The MSCKF pipelines proposed in this work have been developed using the C++ programming language in Linux Ubuntu 16.04 environment on an NVIDIA Jetson TX-2 embedded board that is empowered by a Dual-Core Denver2 and a Quad-Core ARM Cortex-A57 CPUs both running at 2GHz, an NVIDIA Pascal GPU with 256 CUDA cores running at 1300MHz, and an 8GB of LPDDR4 RAM. OpenCV library (CPU libraries only) [24] has been used for ORB features [25] processing. All matrix operations have been performed using the Eigen library [26]. ...
... In these experiments, the ORB features [25] were employed. In the Feature Detection module, the FAST detector's threshold [16] was adjusted to 30, and a multi-scale pyramid scheme with 6 levels and a scale factor of 1.2 was considered. ...
... In the Feature Extraction module, features were assigned a BRIEF descriptor [32]. In the Feature Matching module, the Nearest Neighbor method with Bruteforce matcher and Hamming distance [25] was employed. ...
Full-text available
In autonomous navigation technologies, the Multi-State Constraint Kalman Filter (MSCKF) is one of the most accurate and robust tightly-coupled fusion frameworks for Visual-Inertial Navigation (VIN). However, the adoption of the MSCKF VIN system in real-time embedded applications depends heavily on an efficient implementation of its tangled pipeline. This work initially proposes a novel parallel multi-thread implementation of the MSCKF VIN pipeline on an embedded CPU-enabled hardware that has speeded up the per-epoch processing time of the pipeline by 41% compared to the conventional sequential implementation. The heart of the MSCKF pipeline’s visual backend is an inertially-aided 3D localization of visual feature points that are reduced to a set of nonlinear optimization problems which were conventionally solved in a serial fashion using the single-objective Gauss-Newton optimization algorithm. This work leveraged the parallel architecture of an embedded GPU and further proposes an efficient parallel implementation of a multi-objective Gauss-Newton algorithm. Integration of the proposed GPU-accelerated feature localization technique in the MSCKF parallel pipeline has resulted in 33% faster per-epoch processing time and consequently, the satisfaction of strict real-time constraints. The proposed parallel MSCKF VIN pipelines have been developed using C++ and CUDA on the NVIDIA Jetson TX2 embedded board. Experimental evaluations on a real visual-inertial odometry dataset have been provided to validate the efficacy and real-time performance enhancement of the proposed parallel implementation.
... Feature Descriptor Extraction and Matching In recent years, many local feature detectors and descriptors, such as SIFT [13], SURF [2] and ORB [27], have been developed and used for object recognition, image registration, classification, or 3D reconstruction. To enable real-time navigation assistance, we chose the ORB feature descriptor [27], which are oriented multi-scale FAST [1] corners with a 256-bits descriptor associated. ...
... Feature Descriptor Extraction and Matching In recent years, many local feature detectors and descriptors, such as SIFT [13], SURF [2] and ORB [27], have been developed and used for object recognition, image registration, classification, or 3D reconstruction. To enable real-time navigation assistance, we chose the ORB feature descriptor [27], which are oriented multi-scale FAST [1] corners with a 256-bits descriptor associated. There are two main advantages of ORB: 1) ORB uses an orientation compensation mechanism, making it rotation invariant; 2) ORB learns the optimal sampling pairs, whereas other descriptors like BRIEF [3] uses randomly chosen sampling pairs. ...
Full-text available
People with blindness and low vision (pBLV) experience significant challenges when locating final destinations or targeting specific objects in unfamiliar environments. Furthermore, besides initially locating and orienting oneself to a target object, approaching the final target from one's present position is often frustrating and challenging, especially when one drifts away from the initial planned path to avoid obstacles. In this paper, we develop a novel wearable navigation solution to provide real-time guidance for a user to approach a target object of interest efficiently and effectively in unfamiliar environments. Our system contains two key visual computing functions: initial target object localization in 3D and continuous estimation of the user's trajectory, both based on the 2D video captured by a low-cost monocular camera mounted on in front of the chest of the user. These functions enable the system to suggest an initial navigation path, continuously update the path as the user moves, and offer timely recommendation about the correction of the user's path. Our experiments demonstrate that our system is able to operate with an error of less than 0.5 meter both outdoor and indoor. The system is entirely vision-based and does not need other sensors for navigation, and the computation can be run with the Jetson processor in the wearable system to facilitate real-time navigation assistance.
... The hand-crafted methods typically estimate the scale and orientation with the analyses of gradients in the local region [5]. Some predict only one scale/orientation for a patch [32], termed as hard estimators in this paper. ...
... SURF [3] obtains the orientation according to the distribution of Haar-wavelet responses. ORB [32] defines the orientation as the direction from the keypoint location to the intensity centroid in the local region. ...
Conference Paper
Full-text available
Estimating keypoint scale and orientation is crucial to extracting invariant features under significant geometric changes. Recently, the estimators based on self-supervised learning have been designed to adapt to complex imaging conditions. Such learning-based estimators generally predict a single scalar for the keypoint scale or orientation, called hard estimators. However, hard estimators are difficult to handle the local patches containing structures of different objects or multiple edges. In this paper, a Soft Self-Supervised Estimator (S3Esti) is proposed to overcome this problem by learning to predict multiple scales and orien-tations. S3Esti involves three core factors. First, the esti-mator is constructed to predict the discrete distributions of scales and orientations. The elements with high confidence will be kept as the final scales and orientations. Second, a probabilistic covariant loss is proposed to improve the consistency of the scale and orientation distributions under different transformations. Third, an optimization algorithm is designed to minimize the loss function, whose convergence is proved in theory. When combined with different keypoint extraction models, S3Esti generally improves over 50% accuracy in image matching tasks under significant viewpoint changes. In the 3D reconstruction task, S3Esti decreases more than 10% reprojection error and improves the number of registered images. [code release]
... The aim is to find the image coordinates u l and v (in the left image) and u r (in the right image plane) of the world point m, as see in Figures 5 and 6. For feature extraction, ORB features [31,32] were used, where the extracted features are matched within the stereo image pair and between subsequent captured image pairs. It is assumed that the remaining image coordinates represent the same point on the robot platform, then, these points' 3D coordinates are reconstructed. ...
Full-text available
A slip-aware localization framework is proposed for mobile robots experiencing wheel slip in dynamic environments. The framework fuses infrastructure-aided visual tracking data (via fisheye lenses) and proprioceptive sensory data from a skid-steer mobile robot to enhance accuracy and reduce variance of the estimated states. The slip-aware localization framework includes: the visual thread to detect and track the robot in the stereo image through computationally efficient 3D point cloud generation using a region of interest; and the ego motion thread which uses a slip-aware odometry mechanism to estimate the robot pose utilizing a motion model considering wheel slip. Covariance intersection is used to fuse the pose prediction (using proprioceptive data) and the visual thread, such that the updated estimate remains consistent. As confirmed by experiments on a skid-steer mobile robot, the designed localization framework addresses state estimation challenges for indoor/outdoor autonomous mobile robots which experience high-slip, uneven torque distribution at each wheel (by the motion planner), or occlusion when observed by an infrastructure-mounted camera. The proposed system is real-time capable and scalable to multiple robots and multiple environmental cameras.
... The feature points in the image can be regarded as the more significant points in the image, some common feature point extraction algorithms include SIFT [41], SURF [42], FAST [43], ORB [44], and other methods. The ORB feature points use FAST feature points and combine the BRIEF descriptors to represent the attributes of these feature points. ...
Full-text available
In the traditional visual simultaneous localization and mapping (SLAM), the strong static assumption leads to a large degradation in the accuracy of visual SLAM in dynamic environments. For this reason, many scholars incorporate semantic segmentation networks into the visual SLAM framework to extract dynamic information in images. However, most semantic segmentation networks consume a lot of computing time due to the large model size, which leads to the algorithm’s inability to meet real-time requirements in practical applications. In this paper, a real-time visual SLAM algorithm based on deep learning is proposed. This novel algorithm is based on ORB-SLAM2, and a parallel semantic thread based on the lightweight object detection network YOLOv5s is designed, which enables us to get semantic information in the scene more quickly. In the tracking thread, an optimized homography matrix module is proposed, which utilizes semantic information to optimize and solve the homography matrix so that we can calculate a more accurate optical flow vector. In the optical flow module, the semantic information is used to narrow down the calculation range of the optical flow value to improve the real-time performance of the system, and the dynamic feature points in the image are removed by the optical flow mask to improve the accuracy of the system. Experimental results show that compared with ORB-SLAM2, DynaSLAM, and other excellent visual SLAM algorithms, this algorithm can effectively reduce the absolute trajectory error of visual SLAM in dynamic environments. Compared with other deep learning-based visual SLAM algorithms, the real-time performance of this algorithm is also significantly improved.
... Feature descriptors are extracted from local regions surrounding each point, in order to match points with the highest feature similarity. Feature extraction can be performed by hand-crafting features [23,5,32] or with a learned convolutional neural network (CNN) [9,29]. SuperGlue [33] and LoFTR [37] introduce a GNN and Transformer module respectively, as a means to incorporate both local and global image information during feature extraction. ...
Full-text available
3D multi-object tracking (MOT) is a key problem for autonomous vehicles, required to perform well-informed motion planning in dynamic environments. Particularly for densely occupied scenes, associating existing tracks to new detections remains challenging as existing systems tend to omit critical contextual information. Our proposed solution, InterTrack, introduces the Interaction Transformer for 3D MOT to generate discriminative object representations for data association. We extract state and shape features for each track and detection, and efficiently aggregate global information via attention. We then perform a learned regression on each track/detection feature pair to estimate affinities, and use a robust two-stage data association and track management approach to produce the final tracks. We validate our approach on the nuScenes 3D MOT benchmark, where we observe significant improvements, particularly on classes with small physical sizes and clustered objects. As of submission, InterTrack ranks 1st in overall AMOTA among methods using CenterPoint detections.
Visual SLAM can be mainly divided into direct method and feature-based method, and these two methods develop relatively independently. In recent years, feature-based SLAM systems have been significantly improved by introducing more robust features, effective matching and optimization frameworks, or other sensors. The introduction of semantic information also promotes the development of direct methods. However, visual SLAM usually assumes that the system satisfies the constant velocity assumption. This assumption may lead to a poor initial pose so that the subsequent optimization falls into a local minimal. Meanwhile, large field of view changes often lead to an increase in the error of feature matching for feature-based method and large illumination changes often lead to an increase of photometric error for direct method, thus negatively influencing SLAM systems. In this paper, we mainly target at feature-based method. In detail, we focus on the number and quality of feature matches, as well as the accuracy of the initial pose, thus an interpolation mechanism for SLAM is proposed. Specifically, we introduce the interpolation network originally used to increase the number of video frames into visual SLAM. First, we point out that the traditional interpolation network evaluation metrics are not suitable for the SLAM systems, and we provide the corresponding evaluation metric. Secondly, we verify that it works both for hand-crafted and deep learning features. Thirdly, in order to verify the effectiveness and transferability of our method, we also apply our method to SLAM systems based on direct method, which proves that our method is also applicable to the direct method. Fourthly, we point out that the interpolation network effectively slows down the pose transformation of the SLAM system by inserting an intermediate frame between the previous frame and the current frame, so that the system can obtain a better initial pose based on the constant velocity assumption. This can also explain why visual-inertial systems can effectively improve the performance of visual SLAM. Finally, to ensure the efficiency of the SLAM system, we provide a turning detection module and propose a method to interpolate only at turnings. Extensive experiments and analyses verify the effectiveness and transferability of the proposed system.
In this study, we propose a novel end-to-end system called Human–Machine Collaborative Inspection (HMCI) to enable collaboration between inspectors with Mixed Reality (MR) headsets and a robotic data collection platform (robot) for structural inspections. We utilize the MR headset’s holographic display and precise head tracking to allow inspectors to visualize and localize information (e.g., structural defect) on the real scenes, which are gathered by the robot and processed by an offsite computational server. The primary use case of HMCI is to enable the inspector to visualize, supervise, and improve results produced by automated defect detection algorithms in near real-time. The workflow in HMCI starts with collecting images and depth data to generate 3D maps of the site from the robot. A technique called single-shot localization is developed to create visual anchors for real-time spatial alignment between the robot and the MR headset. The 3D map and images are then sent to the computational server for analysis to detect defects and their locations. Then, the information is received by the MR headset and overlaid on the actual scenes to visualize it with spatial context. An experimental study is conducted in a lab environment to demonstrate HMCI using Microsoft HoloLens 2 (HL2) as the MR headset and Turtlebot2 as the robot. We start with the reconstruction of a 3D environment using a 3D depth sensor (Azure Kinect) on Turtlebot2 and visually detect fiducial markers as regions-of-interest (replicating structural damage) along a predefined inspection path. Then, regions-of-interest are successfully anchored to the real scene and visualized through HL2. To our knowledge, HMCI is one of the first human–machine collaborative systems that can integrate robots and inspectors with the MR headset, which has been developed, tested, and presented for structural inspection.
Full-text available
This report presents the results of the 2006 PASCAL Visual Object Classes Challenge (VOC2006). Details of the challenge, data, and evaluation are presented. Participants in the challenge submitted descriptions of their methods, and these have been included verbatim. This document should be considered preliminary, and subject to change.
Full-text available
This paper describes novel implementations of the KLT feature track- ing and SIFT feature extraction algorithms that run on the graphics processing unit (GPU) and is suitable for video analysis in real-time vision systems. While significant acceleration over standard CPU implementations is obtained by ex- ploiting parallelism provided by modern programmable graphics hardware, the CPU is freed up to run other computations in parallel. Our GPU-based KLT im- plementation tracks about a thousand features in real-time at 30 Hz on 1024 £ 768 resolution video which is a 20 times improvement over the CPU. It works on both ATI and NVIDIA graphics cards. The GPU-based SIFT implementation works on NVIDIA cards and extracts about 800 features from 640 £ 480 video at 10Hz which is approximately 10 times faster than an optimized CPU implementation.
Full-text available
This report presents the results of the 2006 PASCAL Visual Object Classes Challenge (VOC2006). Details of the challenge, data, and evalu-ation are presented. Participants in the challenge submitted descriptions of their methods, and these have been included verbatim. This document should be considered preliminary, and subject to change.
Conference Paper
Full-text available
In this paper, we present a novel scale- and rotation-invariant interest point detector and descriptor, coined SURF (Speeded Up Robust Features). It approximates or even outperforms previously proposed schemes with respect to repeatability, distinctiveness, and robustness, yet can be computed and compared much faster. This is achieved by relying on integral images for image convolutions; by building on the strengths of the leading existing detectors and descriptors (in casu, using a Hessian matrix-based measure for the detector, and a distribution-based descriptor); and by simplifying these methods to the essential. This leads to a combination of novel detection, description, and matching steps. The paper presents experimental results on a standard evaluation set, as well as on imagery obtained in the context of a real-life object recognition application. Both show SURF’s strong performance.
In this paper, we present a novel feature detection approach designed for mobile devices, showing optimized solutions for both detection and description. It is based on FAST (Features from Accelerated Segment Test) and named 3D FAST. Being robust, scale-invariant and easy to compute, it is a candidate for augmented reality (AR) applications running on low performance platforms. Using simple calculations and machine learning, FAST is a feature detection algorithm known to be efficient but not very robust in addition to its lack of scale information. Our approach relies on gradient images calculated for different scale levels on which a modified9 FAST algorithm operates to obtain the values of the corner response function. We combine the detection with an adapted version of SURF (Speed Up Robust Features) descriptors, providing a system with all means to implement feature matching and object detection. Experimental evaluation on a Symbian OS device using a standard image set and comparison with SURF using Hessian matrix-based detector is included in this paper, showing improvements in speed (compared to SURF) and robustness (compared to FAST)
The Pascal Visual Object Classes (VOC) challenge consists of two components: (i) a publicly available dataset of images together with ground truth annotation and standardised evaluation software; and (ii) an annual competition and workshop. There are five challenges: classification, detection, segmentation, action classification, and person layout. In this paper we provide a review of the challenge from 2008–2012. The paper is intended for two audiences: algorithm designers, researchers who want to see what the state of the art is, as measured by performance on the VOC datasets, along with the limitations and weak points of the current generation of algorithms; and, challenge designers, who want to see what we as organisers have learnt from the process and our recommendations for the organisation of future challenges. To analyse the performance of submitted algorithms on the VOC datasets we introduce a number of novel evaluation methods: a bootstrapping method for determining whether differences in the performance of two algorithms are significant or not; a normalised average precision so that performance can be compared across classes with different proportions of positive instances; a clustering method for visualising the performance across multiple algorithms so that the hard and easy images can be identified; and the use of a joint classifier over the submitted algorithms in order to measure their complementarity and combined performance. We also analyse the community’s progress through time using the methods of Hoiem et al. (Proceedings of European Conference on Computer Vision, 2012) to identify the types of occurring errors. We conclude the paper with an appraisal of the aspects of the challenge that worked well, and those that could be improved in future challenges.
1 Challenge The goal of this challenge is to recognize objects from a number of visual object classes in realistic scenes (i.e. not pre-segmented objects). There are ten object classes: For each of the ten object classes predict the presence/absence of at least one object of that class in a test image. The output from your system should be a real-valued confidence of the object's presence so that an ROC curve can be drawn.
High-dimensional indexing plays a critical role in multidimensional data retrieval. In this work, we propose a new indexing method, named HVA-Index, for similarity search in high-dimensional vector space. This index is based on Vector Approximation and Hash Table. The outstanding advantage is that it stores all vectors in a hash table using approximation as key, and the vectors fall into same cell are organized in a linked list. Contrast to VA-File, the HVA-Index doesn't require scan the entire approximation file, and efficiently improves the speed of similarity search. Our experiments prove that HVA-Index outperforms both of the VA-File and the sequential scan in total elapsed time and the number of disk access, and it's still effective at high dimensionality.