IJCB 2014
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE.
IJCB 2014
A Software Framework For Tattoo Matching
Anonymous IJCB 2014 submission
Because challenges such as tattoo-matching potentially
benefit from many pixel-processing approaches, this work
presents the design, deployment, and results of from a
software framework that combines computer vision
algorithms to gain insights into the potential and challenges
involved with using tattoos for biometric purposes. The
system allows users to test out and analyze the performance
of different keypoint detectors, descriptor extractors,
descriptor matchers, match scorers, and other algorithms,
described in section 3.
Many of these algorithms have been implemented in
OpenCV, but several others do not; the framework's
interface enables easy incorporation of almost any
off-the-shelf resources. Since tattoo matching is a relatively
unexplored problem, our system's structure is designed to
thoroughly test all variations and combinations of existing
computer vision algorithms and can compare multiple
metrics for the system's performance. Such a framework
may serve researchers of other image-based biometric
tasks, as well.
Abstract
Tattoo matching is a compelling problem at the
intersection of biometrics and vision. Tattoo-matching
systems offer potential aid in identifying groups or
individuals who associate with or can be identified by their
tattoos. Because tattoos of the same class (e.g. tattoos of
hearts) can appear very different, many computer vision
algorithms provide helpful partial solutions to the problem,
but finding the optimal combination of those algorithmic
tools is not at all trivial. This paper describes a software
framework designed to compare a large set of computer
vision algorithms as efficiently and thoroughly as possible.
It provides one-to-one, one-to-many, and many-to-many
comparisons of performance, while being as simple as
possible to run. Although the original motivation served
tattoo-matching, Although designed with tattoo-matching
in mind, the framework and its architecture also provide a
foundation for efficiently exploring other vision-based
biometric challenges.
1. Introduction
More and more common, tattoos are of increasing
interest and importance to the biometrics community.
Depending on the specific application at hand, systems for
matching tattoos can leverage a wide variety of
pixel-processing approaches. Figure 1 (top) hints at the
variety of tattoos that might be considered "matching": the
right pair are the different images of the same tattoo and the
left pair are distinct renderings of the same figure.
Resources such as OpenCV [13] and others offer
collections of algorithms for exploring this challenging
matching problem, but they provide less structure for
composing and comparing those algorithms for specific
tasks. The bottom panel of Figure 1 shows two very
different combinations of feature detection, extraction,
matching, and scoring for the two tattoo pairs above.
Figure 1: (Top) Examples of similar-class tattoos, e.g.,
painted woman and lion. (Bottom) Different keypointmatching pipelines run on the pairs: at left, BRISK and
hamming distance; at right, SURF and RANSAC
1
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
IJCB 2014
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE.
IJCB 2014
create a single program that takes two images and gives
them a score because images taken under one set of
conditions would have different results than another set of
images taken under different conditions. Since the actual
use and types of images of tattoos the system will be used
on is unknown, we created a system that allows users to
swap out algorithms and parameters to find the best
combination for their particular set of images.
2. Background
Tattoos could be very important for the field of
biometrics. While there exist techniques that are able to
match two dimensional image patterns, tattoos have a
number of specific issues that make them difficult to match
using traditional computer vision algorithms. Issues
include lighting, skin imperfections, the changing surface
they exist on, artistic ability, etc. We would like images to
have a better score based on their similarity. For example,
two snake tattoos should be a very close match, a snake
tattoo should be a match with a different snake tattoo, and a
snake tattoo should not be a match with a dragon tattoo.
3. Algorithms
Many algorithms can serve tattoo-matching. Most of
them start by seeking corresponding keypoints between
images. Keypoint-matching generally has three stages:
feature detection, descriptor extraction, and descriptor
matching. These can overlap - some algorithms can both
detect and extract keypoints - but are distinct enough to
form the backbone of our system design, Figure 2: our
current system uses the OpenCV implementations of the
algorithms noted. The final stage, match-scoring, quantifies
the overall match results. The framework allows easy
combination and comparison of various detectors,
extractors, matchers, and scorers in order to determine the
combination with the best precision, accuracy, recall,
and/or performance.
2.1. Tattoos and Biometrics
Tattoos have a high potential for use in identification
since they are larger and more accessible than other
physiological characteristics such as fingerprints or irises.
They could be used to distinguish a gang member or
identify a person after a natural disaster, along with many
other security and safety applications.
2.2. Problem space for tattoo-matching
Because tattoos are on bodies, they can appear stretched
or warped in different images of the same tattoo.
Perspective changes can also change a tattoo’s shape in a
two dimensional image. In order for a computer vision
algorithm to perform well, it has to account for these visual
differences by using algorithms that are invariant to many
geometric transformations, including shearing, translation,
rotations, and scaling; and fortunately, many feature
comparison algorithms are invariant to these transforms.
Tattoos are also affected by lighting and context changes
that are common problems in image processing. Images
taken with surveillance cameras may not be high quality
and are unlikely to be taken under good lighting conditions.
Finally, tattoos are often partially obscured by clothing or
may be cut off when they are not the main focus of the
image.
Beyond image processing challenges, tattoos carry their
own set of problems. Varied skin tones can make it difficult
to detect where skin ends and the tattoo begins, especially
when skin is close to the color of a tattoo that is only an
outline as opposed to a fully filled in design. Also, a tattoo
artist’s skill or different interpretations of the same image
can greatly impact the kind of tattoo produced. Thus, due to
differences in the design, interpretation, and production of a
tattoo, tattoo images that are of the same class (e.g. a tattoo
of a heart) can appear very different.
We wanted our system to be accurate no matter what
database of images it is used on. It would not be possible to
Figure 2. (Top) The four stages common to many
image-matching applications. (Bottom) The
command-line interface to the framework allows a
user to mix and match those algorithms arbitrarily.
3.1. Keypoint Detectors
The most common type of algorithms implemented,
detectors locate all areas of interest in an image. These
areas of interest can include keypoints, curves, or regions.
Keypoints designate regions where the image gradient is
changing in a locally distinct way, e.g., at a corner. Often
this means that such a point would be re-identifiable
regardless of modest rotation, warping or other simple
transforms. The detector algorithms return keypoint
locations for use in the following extraction stage. The
current system uses SimpleBlob, BRISK[1], Dense,
FAST[2][3], GFTT[4], HARRIS, MSER[5], ORB[6],
SIFT[7], STAR[8], and SURF[9] detectors; others would
be straightforward to incorporate, as well.
2
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
IJCB 2014
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE.
IJCB 2014
descriptors from other images. In this vein the current
system includes contour-matching by ShapeContext [16],
GIST [14] and local binary patterns, LBP [15]. In each of
these cases the upstream processing is idiosyncratic, but all
of them take in two images and output a match score
between them. From a high-level perspective, then, these
algorithms add three more scorers to the two noted above.
3.2. Descriptor Extractors
After finding the locations of an image’s keypoints, the
system passes those locations along to the designated
extractors. An extractor algorithm creates a vector
characterizing the image region around a keypoint. For
example, if the descriptor has found a point where the
gradient changes in a useful way, the extractor might
include a histogram of the orientations along which the
gradients change. Implemented extractors include
BRIEF[10], BRISK, FREAK[11], ORB, SIFT, and SURF.
In addition to returning a descriptor vector, some extractors
also provide a scale and orientation reference: this
information can increase the set of conditions for which
matching, the next pipeline stage, will succeed.
3.6. Ensemble Learning
The framework's primary strength is in its organization
of these component algorithms and its facility for creating
combinations of them. Ensemble learning naturally builds
from this foundation, and our system includes an
implementation of AdaBoost [17], an adaptive boosting
algorithm. To use AdaBoost, the user simply appends a flag
to the algorithms they want to run (see Section 5). The
AdaBoost algorithm will then use each pair of images as an
instance and each specified computer vision algorithm as a
weak learner. The result is a composite, learned algorithm
whose prediction is a weighted sum of all of the weak
learners’ predictions.
3.3. Descriptor Matchers
The extracted vectors are matched by the algorithms in
this stage of the pipeline. Matches are made by pairing
keypoints with the most similar visual characteristics. For
instance,
typical
matchers
will
compare
the
gradient-orientation histograms, and some methods will
also account for characteristics such as the difference in
pixel location. They can be augmented by the geometric
consistency of a set of correspondences, e.g., through
RANSAC. Matchers currently implemented are variations
on brute force (Hamming, L1), radius matching, KNN,
K-Nearest Neighbors, and a fast KNN approximation
known as FLANN[12].
3.7. Tattoo Segmentation
Thus far, we have assumed that the images are primarily
of tattoos and nothing else, but that is generally not the case
when images are pulled from surveillance cameras or other
realistic scenarios. Most images that contain tattoos are not
solely of tattoos. For example, there could be an image of a
man who has a tattoo on his arm. Certainly, there might be a
Frisbee, a dog, a slide, birds, or any other various things in
the image, but we only care about the tattoo on the man’s
arm.
3.4. Scorers
In the final stage of the system, the list of keypoint matches
is passed along to a scorer algorithm that quantifies the
matching of the two images. Our current tattoo-matching
task uses two scoring algorithms, known as Simple and
Distance scoring.The Simple scorer consists of the ratio of
the number of matches found to the number of possible
matches that could have been made, i.e., the number of
keypoints in the sparser of the two images. The Distance
scorer yields the average vector distance between the
matched descriptors, i.e., the visual similarity found.
Figure 3. Stages of the optional image segmentation.
Thus, we have incorporated tattoo-segmenting software
into our system that is run before any of the other steps. The
segmentation software generates a mask, i.e., a grayscale
image in which intensity represents the probability of that
pixel being a part of a tattoo. When specified, the
framework uses a threshold to binarize the image and
extracts the the convex hull of the tattoo region. Finally, a
bounding box of the convex hull is created and the
corresponding area of that bounding box in the original
tattoo image becomes the image that the computer vision
algorithms will operate on. (Figure 3.) As with ensemble
3.5. Flexibility in handling other algorithms
While most algorithms fit into this set of stages, some do
not fall into precisely one of these categories. Our
framework can also incorporate approaches that work
outside this basic pipeline. Rather than finding keypoints
and passing information along from stage to stage, many
algorithms provide a global descriptors of an image,
already prepared for matching against the corresponding
3
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
IJCB 2014
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE.
learning, this is an option that can be included – or not – at
run time in order to determine the performance of the
matching and/or segmentation portions of the system.
IJCB 2014
designed to be extensible so that it is easy to add algorithms
to the system.
4.3. Parameter Specification
4. Framework
When finer tuning of the algorithms is desired, there is a
system in place to tweak the values of the parameters for
each algorithm that can be specified as previously
described via the command line flags. This system consists
of text files containing key value pairs for each of the
algorithms. The user of the system can edit these files to
tweak the parameters.
The framework seeks to balance the conflicting demands
of being both user-friendly and extensible. Although
OpenCV has a wide array of vision algorithms, it is not as
user-friendly in providing support for testing combinations
of those algorithms. Our system, built alongside OpenCV,
simplifies the task of running "contests" among component
algorithms in order to determine which pieces work best for
a particular vision-based biometric task. Thus, though
initially tattoo matching, the framework is not limited to
that domain – it will support a wide variety of situations in
which a user needs to quickly compare the performance of
various combinations of vision algorithms.
4.4. Serialization
In
a
one-versus-many
contest
run
(or
many-versus-many) the naïve way of matching the images
is to run them against each other N many times. However,
this system speeds things up by outputting the intermediate
keypoint and descriptor results for the “one” image to a
JSON file. Then, instead of re-computing the keypoints—a
slow operation—the keypoints are instead read from file.
This is done separately for each different algorithm
combination. We have found that serialization dramatically
speeds the process of matching and scoring similarity
between images.
4.1. "Contest" support
The framework efficiently runs many different image
comparison "contests," where the contest winner is the
most accurate combination of algorithms for a particular set
of images. Such contests to be one-versus-one,
one-versus-many, and many-versus-many. In terms of
tattoo-matching, this translates to one image versus one
image, one image versus a folder of images, and a folder of
images versus a folder of images. We foresee several
possible use cases for each of these contests. For example,
one versus many may be used to determine the closest
matches for one image against a database. The one versus
one, on the other hand, may be used for finding the match
score between two particular images. A winner for one
contest may not be the same for a different contest with
different images.
4.5. Other components
In addition to the four general steps in the matching
process, the system also features other components.
Ensemble learning, e.g., -ensemble, and masking are
two such components that can be added in addition to the
typical image matching steps. The framework makes it easy
to add additional components to handle specific use cases.
5. Output
4.2. Algorithm Specification
After running any algorithm or pair of algorithms, the
output from the run is stored in a csv file whose name
contains the information about the run. In addition, the csv
file is stored in a directory structure organized by folders or
images given as input to the testing framework. Finally, the
parameters used for the run are stored in a text file whose
name is based on the timestamp. Thus, all information
about the input to a run of the framework is encoded in the
names of the output, so that users can easily identify and
compare the output from various runs.
The output, in turn, is stored in columns of a
comma-separated values (csv) file, includes: (1) the name
of first image being compared, (2) the name of second
image being compared, (3) index, (4) score, and (5) truth.
The index is an increasing unique identifier for each trial,
The system does not take on the constraints (and
dependencies) of a graphical interface; Figure 2 shows an
example invocation of the Python-based framework. For
flexibility, command line flags are used to specify the
algorithms to be compared. For example, a user might enter
-d SURF SIFT as an option. The -d specifies the stage of
matching of concern – keypoint detection in this case, and
SURF SIFT indicates that the matching should use both
SURF and SIFT algorithms for detecting keypoints. This
will not combine the algorithms (yet), but will calculate
each algorithm's match scores separately. Similar flags for
extraction and scoring steps. Algorithm specification is
4
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
IJCB 2014
IJCB 2014
the score is the final output of the testing framework. The
ground-truth column holds 1 if the images are of the same
class (we use filenames to indicate class, e.g.,
butterfly1.png and butterfly2.jpg) and 0 otherwise.
scores that are considered positive). This metric shows the
user how well their algorithms perform at recognizing
when a pair of images is a match, as shown in Figure 4.
5.1. Analysis
5.1.2
The analysis of the data is primarily dependent upon the
score, which serves as our prediction and the truth. The
metric tools provided in the framework include Receiver
Operator Characteristic curves and Cumulative Match
Characteristic curves. We use R to generate these
visualizations from the framework's output, noted above.
Cumulative Match Characteristic (CMC) curves display
the percent likelihood that an image of the same class is in
the top X images, where X is shown on the horizontal axis.
Thus, the program looks at every image (let’s say imgA)
and the score between it and all other images are examined
in order of the images that have the highest score when
matched against imgA. The program keeps track of how
many images it has to look at before it finds an image of the
same class, and this number is then kept track of for every
imgA and used to derive the probabilities Figure 4 shows a
large run comparing CMC curves of many algorithm
combinations, along with the AdaBoost ensemble of all of
them, further to the upper left, a clear improvement in
performance over any single component.
5.1.1
Receiver Operator Characteristic
Cumulative Match Characteristic
1397628765_BRISK−BRISK−Br uteForce−dist−CMC
1397628765_BRISK−BRISK−Br uteForce−L1
1397628765_BRISK−FREAK−Br uteForce−dist−CMC
1397628765_BRISK−FREAK−Br uteForce−L1
1397628765_BRISK−ORB−Br uteForce−dist−CMC
1397628765_BRISK−ORB−Br uteForce−L1
1397628765_BRISK−SIFT−Br uteForce−dist−CMC
1397628765_BRISK−SIFT−Br uteForce−L1
1397628765_CONTOUR−CMC
1397628765_FAST−BRISK−Br uteForce−dist−CMC
1397628765_FAST−BRISK−Br uteForce−L1
1397628765_FAST−FREAK−Br uteForce−dist−CMC
1397628765_FAST−FREAK−Br uteForce−L1
1397628765_FAST−ORB−BruteForce−dist−CMC
1397628765_FAST−ORB−BruteForce−L1
1397628765_FAST−SIFT−BruteForce−dist−CMC
1397628765_FAST−SIFT−BruteForce−L1
1397628765_GFTT−BRISK−Br uteForce−dist−CMC
1397628765_GFTT−BRISK−Br uteForce−L1
1397628765_GFTT−FREAK−Br uteForce−dist−CMC
1397628765_GFTT−FREAK−Br uteForce−L1
1397628765_GFTT−ORB−Br uteForce−dist−CMC
1397628765_GFTT−ORB−Br uteForce−L1
1397628765_GFTT−SIFT−Br uteForce−dist−CMC
1397628765_GFTT−SIFT−Br uteForce−L1
1397628765_MSER−BRISK−Br uteForce−dist−CMC
1397628765_MSER−BRISK−Br uteForce−L1
1397628765_MSER−FREAK−Br uteForce−dist−CMC
1397628765_MSER−FREAK−Br uteForce−L1
1397628765_MSER−ORB−Br uteForce−dist−CMC
1397628765_MSER−ORB−Br uteForce−L1
1397628765_MSER−SIFT−Br uteForce−dist−CMC
1397628765_MSER−SIFT−Br uteForce−L1
1397628765_ORB−BRISK−Br uteForce−dist−CMC
1397628765_ORB−BRISK−Br uteForce−L1
1397628765_ORB−FREAK−Br uteForce−dist−CMC
1397628765_ORB−FREAK−Br uteForce−L1
1397628765_ORB−ORB−Br uteForce−dist−CMC
1397628765_ORB−ORB−Br uteForce−L1
1397628765_ORB−SIFT−Br uteForce−dist−CMC
1397628765_ORB−SIFT−Br uteForce−L1
1397628765_SURF−BRISK−Br uteForce−dist−CMC
1397628765_SURF−BRISK−Br uteForce−L1
1397628765_SURF−FREAK−Br uteForce−dist−CMC
1397628765_SURF−FREAK−Br uteForce−L1
1397628765_SURF−ORB−Br uteForce−dist−CMC
1397628765_SURF−ORB−Br uteForce−L1
1397628765_SURF−SIFT−Br uteForce−dist−CMC
1397628765_SURF−SIFT−Br uteForce−L1
1397635766−ensemble−CMC
0.2
0.4
0.6
0.8
1.0
Receiver Operator Characteristic (ROC) curves display
the true positive rate (sensitivity) on the y-axis and the false
positive rate (fall-out) on the x-axis. In order to convert our
scores into binary predictions, we apply a threshold so that
scores above the threshold are converted to 1 and scores
below are converted to 0. Then, to vary the false positive
rate, the threshold is changed (decreasing the threshold
increases the false positive rate by increasing the number of
Percent of Matches
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE.
0
50
100
150
Rank
Figure 4: (Left) Example of a multi-algorithm ROC curve. Each curve on the plot represents a different combination
of the detector/extractor/matcher/scorer algorithms; an AdaBoost-learned ensemble algorithm is also included, though
for the task of maximizing the ratio of true-to-false positives, it is not the winner. (Right) In terms of CMC
performance, the ensemble algorithm is a clear improvement over the components. At right the curves summarize
"how far one must look through the search results" in order to find a category match. Both approaches have broad use
in biometric applications. The contribution of this paper's framework is increased human efficiency in exploring
many different combinations/ensembles of off-the-shelf routines. Both use a dataset of 200 tattoos distributed across
40 different categories. The file-based, caching framework eases the transition to larger datasets, as available.
5
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
IJCB 2014
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE.
6. Conclusion
[9]
This work originally sought to find a "single" algorithm
that would best suit the important biometric application of
matching images of tattoos. Yet after studying the
challenges involved in tattoo matching, it became evident
that progress depended on a framework that would allow
investigators to quickly try out any existing popular
computer vision algorithm and, singly or in combination,
generate relative performance statistics.
In this way, our framework enables users to explore any
matching-based computer vision problem, not just the
problem of matching tattoo images. If a user has a database
of images representative of a phenomenon they hope to
study by matching, then this framework can efficiently and
exhaustively explore a wide variety of existing computer
vision algorithms, their variations, and their combinations.
Whether trying to match images of kittens or images of
tattoos, this framework empowers users to find the best
algorithm for their specific problem.
[10]
[11]
[12]
[13]
[14]
[15]
7. Acknowledgments
The authors thank Nick Orlans and Mikel Rodriguez of
the MITRE corporation for their support, suggestions, and
guidance in this work. This work is made possible by the
MITRE corporation and the Harvey Mudd College
computer science clinic program.
[16]
[17]
References
[1] S. Leutenegger, M. Chli, and Y. Siegwart. BRISK: Binary
Robust Invariant Scalable Keypoints. Porc. IEEE Int. Conf.
Computer Vision (ICCV), 2548–2555, 2011.
[2] E. Rosten and T. Drummond. Fusing Points and Lines for
High Performance Tracking. IEEE Int. Conf. Computer
Vision (ICCV), 1508-1511, 2005.
[3] E. Rosten and T. Drummond. Machine Learning for
High-Speed Corner Detection. ECCV, 430 – 443, 2006.
[4] J. Shi and C. Tomasi. Good Features to Trak. IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR), 593-600, 1994.
[5] J. Matas, O. Chum, M. Urban, and T. Pajdla. Robust Wide
Baseline Stereo from Maximally Stable Extremal Regions.
Proc. Of British Machine Vision Conference, 384 – 396,
2002.
[6] E. Rublee, V. Rabaud, K. Konolige, and G. Bradski. ORB: an
efficient alternative to SIFT or SURF. Proceedings of the
ICCV, 13, 2011.
[7] D. Lowe. Object recognition from local scale-invariant
features. Proceedings of the ICCV, 2,1150–1157, 1999
[8] M. Agrawal, K. Konolige, and M. R. Blas. CenSurE: Center
Surround Extremas for Realtime Feature Detection and
6
IJCB 2014
Matching. Computer Vision – ECCV 2008, 4,103-115,
2008.
H. Bay, A Ess, T. Tuytelaars, L. Van Gool. SURF: Speeded
Up Robust Features.
Computer Vision and Image
Understanding (CVIU), 110(3), 346-359, 2008.
M. Calonder, V. Lepetit, C. Strecha, and P. Fua. BRIEF:
Binary Robust Independent Elementary Features. In
Proceedings of the European Conference on Computer
Vision (ECCV), 2010.
A. Alahi , R. Ortiz and P. Vandergheynst. Freak: Fast retina
keypoint. Proc. IEEE CVPR, 510 -517, 2012.
M. Muja and D. G. Lowe. Fast Approximate Nearest
Neighbors with Automatic Algorithm Configuration.
International Conference on Computer Vision Theory and
Applications (VISAPP'09), 2009.
OpenCV Dev Team. (Dec 31, 2013). Common Interfaces of
Descriptor
Matchers.
OpenCV.
Available:
http://docs.opencv.org/modules/features2d/doc/common_int
erfaces_of_descriptor_matchers.html. [2014, April 7].
C. Siagian, L. Itti. Rapid Biologically-Inspired Scene
Classification Using Features Shared with Visual Attention.
IEEE Transactions on Pattern Analysis and Machine
Intelligence, 29(2), 300-312, 2007.
T. Ojala, M. Pietikäinen, and D. Harwood. Performance
evaluation of texture measures with classification based on
Kullback discrimination of distributions. Proceedings of the
12th IAPR International Conference on Pattern Recognition
(ICPR 1994), 1, 582 - 585, 1994
S. Belongie, J. Malik, and J. Puzicha. Shape Context: A new
descriptor for shape matching and object recognition,
Proceedings, NIPS pp. 831-837, 2000.
J. Friedman, T. Hastie, and R. Tibshirani. Additive Logistic
Regression: a Statistical View of Boosting, Annals of
Statistics v. 28 1998.
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699