=Paper=
{{Paper
|id=Vol-1391/33-CR
|storemode=property
|title=CNRS TELECOM ParisTech at ImageCLEF 2015 Scalable Concept Image Annotation Task: Concept Detection with Blind Localization Proposals
|pdfUrl=https://ceur-ws.org/Vol-1391/33-CR.pdf
|volume=Vol-1391
|dblpUrl=https://dblp.org/rec/conf/clef/Sahbi15
}}
==CNRS TELECOM ParisTech at ImageCLEF 2015 Scalable Concept Image Annotation Task: Concept Detection with Blind Localization Proposals==
https://ceur-ws.org/Vol-1391/33-CR.pdf
CNRS TELECOM ParisTech at ImageCLEF
2015 Scalable Concept Image Annotation Task:
Concept Detection with Blind Localization
Proposals
Hichem SAHBI
CNRS TELECOM ParisTech
hichem.sahbi@telecom-paristech.fr
Abstract. We introduce our participation at the ImageCLEF 2015 scal-
able concept detection and localization task. This edition focuses on
generating not only annotations (concept detections) but also localiz-
ing concepts into a large image collection.
Concept detection part of our runs is based on standard nonlinear sup-
port vector machines (SVMs). The localization part is blind and based
on a priori learned statistics that generate multiple localization propos-
als. In spite of its blindness, the performance of this concept localization
framework is promising.
Keywords: Support vector machines, histogram intersection kernels,
concept detection, blind concept localization proposals
1 Introduction
The general problem of visual category recognition generally includes three dif-
ferent tasks: concept detection (also known as image annotation) [1–3], concept
localization [4, 5] and object category segmentation [6, 7]. Concept detection con-
sists in inferring a list of keywords that best describes the visual and the semantic
content of a given image while localization seeks to find a list of bounding boxes
that defines the span of detected concepts. As a variant of concept localization,
object category segmentation consists in delimiting the extent of detected con-
cepts with a high precision.
We are interested in this paper in concept detection and localization; we
present our solutions submitted to the ImageCLEF 2015 scalable concept im-
age annotation task [8, 9]. This edition focuses on concept localization, which
consists in finding all the occurrences of a list of concepts into a given test im-
age. This task has been widely studied in different related challenges including
Pascal VOC [10], ImageNET [11] and more recently MS-COCO [12]. Existing
solutions usually parse images using sliding windows [13], image segmentation
and superpixels [14] as well as multiple segmentation proposals [15]. In these
methods, detection and segmentation results are scored using machine learning
techniques (such as SVM [16], deep networks [17], and decision forests [18]) and
Fig. 1. Sample of pictures taken from the ImageCLEF2015 database.
consolidated using spatial layout and geometric relationships usually described
with graphical models (such as conditional and Markov random fields [19, 20]).
For more detailed discussions of related work in concept detection and localiza-
tion, see [21] and references therein.
Among existing object localization (and segmentation) methods those based
on region proposals are currently receiving a particular attention. Their general
principle consists in defining multiple partitions of test images into sets of blobs
that potentially correspond to actual objects. Only few of these partitions are
scored and used to annotate and localize concepts in test images. Even though
relatively successful, these approaches are highly dependent on the quality of im-
age segmentation, which is known to be challenging especially when no a priori
information is used about the statistics of these concept-localization proposals.
Our proposed solution, discussed in this paper, avoids image segmentation and
it is based on two steps: first, we train SVM classifiers that detect concepts be-
longing to different test images. Afterwards, we use an priori (trained) statistical
model in order to infer their most likely locations, without observing the content
of these test images. We will show that in spite of the simplicity of this ap-
proach, the results are reasonably decent, and very promising, and this opens a
new direction towards refining these models and obtaining better performances
by combining annotation and concept localization results.
The rest of this paper is organized as follows; first, we describe our concept
detection algorithm, based on SVMs and an efficient evaluation of the histogram
intersection kernel. Then, we describe our a priori statistical model for blind
concept localization, and we present and discuss our ImageCLEF 2015 results.
Finally, we conclude the paper, with possible extensions for a future work.
TEST IMAGE
2 Concept detection criteria
Criteria (C1) Criteria (C2)
concept1 conceptN concept1 conceptN
SVM1 SVMN SVM1 SVMN
5 Localization heuristics
Localization Heuristic 2 Localization Heuristic 5
concept1 conceptN concept1 conceptN
Loc1 LocN Loc1 LocN
TEST IMAGE TEST IMAGE TEST IMAGE TEST IMAGE
conceptN
5 x 2 Runs
concept1 concept1
conceptN conceptN
concept1 concept1 concept1 concept1
conceptN
concept1 concept1
RUN 3 (statistical + C1) RUN 4 (statistical + C2) RUN 9 (adapted dimension+shift + C1) RUN 10 (adapted dimension+shift + C2)
Fig. 2. This figure shows the two-step process used for concept detection and localiza-
tion
2 Concept Detection with Blind Localization Proposals
Our concept detection and localization results are obtained according to the two
following steps (see Fig. 2):
i) Holistic concept detection: this step is achieved using global (holistic)
visual and textual features. For that purpose, we train “one versus all” SVMs
for each concept, in order to detect whether that concept exists in a given test
image (see extra details in Section 2.1).
ii) Blind concept localization proposals: in contrast to concept detection,
concept localization is achieved blindly, i.e., without observing the content of a
given test image. As will be shown subsequently, localization is achieved using a
priori knowledge about possible locations of these bounding boxes. These knowl-
edges correspond to learned localization statistics, of bounding boxes, taken from
a training/dev set of concepts and their associated bounding boxes (i.e., from
the file “imageclef2015.dev.bbox.v20150226”; see extra details in Section 2.2).
2.1 Holistic concept detection: training and classification
We used only the holistic features provided in this ImageCLEF task including
GIST, Color Histograms, SIFT, C-SIFT, RGB-SIFT, OPPONENT-SIFT, etc.
We build 10 gram matrices (9 visual and 1 textual), based on efficient histogram
intersection kernel, associated to these features. Then, we linearly combine those
matrices into a single one. Notice that this combination does not result from
multiple kernel learning but just a convex combination of kernels with uniform
weights. We plug the resulting kernel into SVMs for training and testing.
For each concept, we train “one-versus-all” SVM classifiers; we use many random
folds (taken from training/dev data in “imageclef2015.dev.bbox.v20150226”) for
multiple SVM training and we use these SVMs in order to predict the concepts
on the test set1 . We repeat this training process, for each concept, through
different random folds from the training set and we take the average scores of the
underlying SVM classifiers. This makes classification results less sensitive to the
sampling of the training set and also allows us to re-balance classification results
mainly for concepts with unbalanced distributions of positive and negative data.
2.2 Blind concept localization proposals
Several heuristics are tried in order to suggest multiple concept localization pro-
posals. Given a test image and the list of concepts attached to it (see Sec-
tion 2.1), concept localization is achieved without consulting the content of
the test image (but only its detected concepts). Indeed, concept localization
is blind and bounding boxes (BBs) are either fixed (using test image dimen-
sions) or based on statistics estimated offline on the training/dev set (in “im-
ageclef2015.dev.bbox.v20150226”) as described subsequently.
In what follows, pc = (x, y, w, h) denotes the bounding box coordinates of a
given detected concept c in a given test image; here (x, y) (resp. (w, h)) corre-
sponds to the center (resp. dimensions; width and height) of the bounding box pc .
Heuristic 1 (fixed BBs): for a detected concept c in a given test image, its
bounding box pc is set to (W/2, H/2, W, H); here W and H respectively denote
the width and the height of the test image. In what follows, we consider that
all the test images are re-sized and have the same dimensions (i.e., W , H are
constant for all the test images)2 .
In the subsequent heuristics (2–5), we introduce the following notation: given a
concept c, we consider Tc = {pic }i as the union of all the bounding boxes (in the
training/dev set “imageclef2015.dev.bbox.v20150226”) that belong to c. We also
1
A given test image is assigned to a given concept, iff the underlying SVM score is
positive.
2
Of course, the actual dimensions of the test images are taken into account in order
to re-scale concept localization results.
Fig. 3. This figure shows: (left) BB clustering process and the union of bounding boxes
in the training set that belong to a given concept, (middle) BB shift using the first
principal direction and (right) BB re-scale using the first principal direction.
consider Nc as the average number of bounding boxes (par image) associated
to c; Nc is evaluated from the training set. Prior to use the following heuristics
(2–5), we consider an offline step that clusters the coordinates in Tc (using k-
means) with a number of clusters fixed to Nc (see Fig. 3, left).
Heuristic 2 (concept-dependent BBs): for a detected concept c in a given
test image, we generate Nc bounding boxes whose coordinates correspond to the
cluster centers obtained after applying k-means on Tc .
In the remaining three heuristics (3–5), we update the coordinates of the bound-
ing boxes, by manipulating i) their dimensions in heuristic 3, ii) their centers in
heuristic 4, and iii) both their centers and dimensions in heuristic 5.
Heuristic 3 (re-scaled concept-dependent BBs): each bounding box pc =
(x, y, w, h) generated in heuristic 2, is replaced by re-scaled BB. First, principal
component analysis (PCA) is applied offline to the BB dimensions {(wi , hi )}i
in the training set that also belong to concept c, afterwards, the dimensions
(w, h) of pc are moved towards the first principal component of PCA3 , with an
amplitude proportional to its eigenvalue (and this corresponds a re-scale of the
dimensions of pc ). In this heuristic (x, y) remains unchanged (see Fig. 3, right).
Heuristic 4 (shifted concept-dependent BBs): for each bounding box
pc = (x, y, w, h) generated in heuristic 2, we generate two extra BBs, with shifted
coordinates. Again, PCA is applied offline to the BB coordinates {(xi , yi )}i in
the training set that also belong to concept c, afterwards, the (x, y) coordinates
of pc are shifted towards two opposite directions corresponding to the first prin-
cipal component of PCA. In this heuristic (w, h) remains unchanged (see Fig. 3,
middle).
Heuristic 5 (shifted and re-scaled concept-dependent BBs): this heuris-
tic corresponds to the combination of the two heuristics 3 and 4.
3
i.e., the eigenvector with the largest eigenvalue.
3 ImageCLEF 2015 Evaluation
The targeted task is, again, concept detection and localization: given a picture,
the goal is to predict which concepts (classes) are present into that picture and
a proposal of bounding boxes surrounding these concepts.
3.1 ImageCLEF 2015 Collection
A very large amount of images was gathered by the organizers, and using asso-
ciated web pages, tags and meta-data were also provided. This set includes 500k
images with only 2k images with known ground truth (i.e., labels and bounding
boxes are given). These images belong to 251 concepts (see example in Fig. 1).
Each image is again described with nine holistic visual features provided by the
organizers, and we compute one extra textual feature using a normalized vector
space model; first, a vocabulary of keywords V is defined4 in order to query the
associated meta-data that include 500k textual descriptions. For each keyword
ω ∈ V, only images whose textual descriptions include ω have their ω vector
entry set to non-zeros.
3.2 Submitted Runs
All our submitted runs are based on SVM training and classification with the
same kernel function (i.e., histogram intersection kernel), and the differences re-
side in the used decision criteria for concept detection and localization. Our ten
submitted runs correspond to the combination of the five concept localization
heuristics described earlier (see Section 2.2) and the two following concept de-
tection criteria
i) Criterion 1 (C1): the first concept detection results are obtained by follow-
ing the setting in Section 2.1.
ii) Criterion 2 (C2): the second set of concept detection results is obtained
using a slightly different criterion; more precisely, if an image has no detected
concepts, i.e., all the SVM scores are negatives for all concepts, then we select
the top 3 concepts (i.e., with the highest negative SVM scores) as annotations
for that image. This makes it possible to increase the recall, with a possible
impact on the precision.
Our runs are summarized in table 1. For all the submitted runs, performances
are evaluated, by the organizers, using a variant of the Jaccard measure; the lat-
ter is defined as the intersection over union of bounding boxes provided in the
submitted runs and those in the ground truth. Mean average precision (MAP)
measures based on different percentages of bounding box overlaps are given for
each concept and also averaged through different concepts (see our results in
4
Including relevant keywords that are used in concept definitions.
Table 1: This table shows the definition of the ten runs submitted to the Image-
CLEF 2015 challenge.
Heuristic 1 Heuristic 2 Heuristic 3 Heuristic 4 Heuristic 5
Criterion 1 (C1) Run 1 Run 3 Run 5 Run 7 Run 9
Criterion 2 (C2) Run 2 Run 4 Run 6 Run 8 Run 10
Table 2: Performances (in %) of our different concept detection and localization
proposal heuristics sorted from the highest to the lowest (taken from ImageCLEF
2015
X
results).
XX Overlap
XXX 0% 10% 20% 30% 40% 50% 60% 70% 80% 90%
XXX
Runs #
5 (heuristic 3+C1) 30.73 27.64 25.11 22.47 19.80 16.92 14.68 12.46 10.05 07.85
9 (heuristic 5+C1) 30.73 27.21 24.87 21.69 19.01 16.58 14.08 11.65 09.32 07.62
3 (heuristic 2+C1) 30.73 26.13 24.48 21.52 18.72 15.82 13.01 10.30 08.03 06.32
7 (heuristic 4+C1) 30.73 25.73 23.50 20.58 17.77 14.61 11.80 09.38 07.38 05.55
1 (heuristic 1+C1) 30.73 26.11 20.79 16.81 13.29 10.49 08.53 06.81 04.99 03.17
6 (heuristic 3+C2) 19.40 17.39 15.83 14.08 12.44 10.63 09.25 08.00 06.56 05.20
10(heuristic 5+C2) 19.40 17.21 15.71 13.83 12.10 10.38 08.90 07.52 06.10 05.01
4 (heuristic 2+C2) 19.40 16.35 15.31 13.56 11.98 10.11 08.33 06.87 05.40 04.15
8 (heuristic 4+C2) 19.40 16.23 14.91 13.17 11.41 09.48 07.83 06.16 04.92 03.68
2 (heuristic 1+C2) 19.40 16.30 13.05 10.53 08.44 06.73 05.53 04.51 03.34 02.21
Tables 2, 3). Details about these measures can be found in the ImageCLEF 2015
website5 .
From tables 2 and 3, we observe the following issues
– Different methods for “concept localization proposals” provide much better
results when concept detection is relatively successful (see runs 1, 3, 5, 7, 9
vs runs 2, 4, 6, 8, 10 in table 2 for different overlap ratios). Following the
spirit of our two-step method, these results clearly corroborate the fact that
concept detection could be decoupled from localization as long as concept
detection is achieved with a relative success. This clearly opens a direction
towards enhancing the performances of localization proposals by further im-
proving concept detection results.
– From table 2, heuristic 3 (BB re-scaling) provides the best overall perfor-
mances; indeed, even though shifting is important, it has less impact on
performances compared to re-scaling. This is mainly due to the variability
and non-rigidity of many concepts (such as animals), that require an adap-
tation of the dimensions of BBs, while shifting is already well captured by
the statistical model (in heuristics 2, 4, 5); see again k-mean clustering in
Section 2.2.
5
http://www.imageclef.org/2015/annotation.
Table 3: Some “concept-by concept” performances (in %) of our different con-
cept detection and localization proposal heuristics (taken from ImageCLEF 2015
results).
shift
ift
ale+sh
escale+
ale
ale
t
t
al+resc
al+resc
al+resc
al+shif
al+shif
tical+r
al
al
statistic
statistic
statistic
statistic
statistic
statistic
statistic
) statis
fixed
fixed
(run 10
(run 5)
(run 9)
(run 3)
(run 7)
(run 1)
(run 6)
(run 4)
(run 8)
(run 2)
concepts description
n01639765 frog 18.18 36.36 36.36 27.27 18.18 18.18 27.27 45.45 27.27 18.18
n01896031 feather 20.00 20.00 40.00 40.00 20.00 20.00 20.00 40.00 40.00 20.00
n02084071 dog 50.00 50.00 50.00 50.00 50.00 33.33 33.33 33.33 33.33 33.33
n02114100 wolf 22.86 20.00 28.57 25.71 20.00 22.86 20.00 28.57 25.71 20.00
n02129165 lion 0 0 0 0 0 02.22 02.22 02.22 02.22 02.22
n02131653 bear 0 0 0 0 50.00 0 0 0 0 50.00
n02206856 bee 66.67 66.67 66.67 66.67 66.67 50.00 50.00 50.00 50.00 50.00
n02330245 mouse 100.0 100.0 100.0 100.0 100.0 50.00 50.00 50.00 50.00 50.00
n02395406 hog 40.00 40.00 52.00 52.00 36.00 40.00 40.00 52.00 52.00 36.00
n02411705 sheep 52.94 52.94 35.29 41.18 47.06 52.94 52.94 23.53 29.41 47.06
n02416519 goat 50.00 50.00 0 0 50.00 33.33 33.33 0 0 33.33
n02430045 deer 25.00 25.00 25.00 25.00 0 25.00 25.00 25.00 25.00 0
n02484322 monkey 57.14 57.14 64.29 64.29 50.00 52.94 52.94 58.82 58.82 47.06
n02503517 elephant 100.0 100.0 100.0 100.0 100.0 28.57 28.57 28.57 28.57 14.29
n02512053 fish 60.00 60.00 70.00 60.00 60.00 46.67 46.67 53.33 46.67 40.00
n02691156 airplane 0 100.0 0 100.0 0 0 0 33.33 33.33 0
n02709367 anchor 63.16 73.68 42.11 47.37 52.63 63.64 63.64 45.45 59.09 54.55
n02774152 bag 10.00 10.00 0 0 10.00 07.69 07.69 0 0 07.69
n02778669 ball 16.67 16.67 16.67 16.67 0 12.50 12.50 12.50 12.50 0
n02782093 balloon 0 05.26 0 05.26 0 0 05.26 0 05.26 0
n02800213 baseball 0 0 0 0 0 01.49 01.49 01.49 01.49 01.49
n02828884 bench 20.00 20.00 25.00 25.00 20.00 20.00 20.00 25.00 25.00 20.00
n02834778 bicycle 10.00 05.00 10.00 05.00 05.00 13.16 05.26 10.53 05.26 07.89
n02839910 bin 30.43 30.43 30.43 30.43 30.43 30.43 30.43 30.43 30.43 30.43
n02883344 box 0 0 0 0 0 04.35 04.35 0 04.35 0
n02909870 bucket 46.67 53.33 53.33 53.33 53.33 41.18 47.06 47.06 47.06 47.06
n02933112 cabinet 60.00 60.00 80.00 80.00 60.00 24.24 21.21 27.27 24.24 21.21
n02942699 camera 0 05.26 0 05.26 0 05.41 05.41 02.70 08.11 05.41
n02984061 cathedral 36.76 37.50 13.97 08.82 34.56 36.76 37.50 15.44 16.18 34.56
n02990373 ceiling 36.36 36.36 36.36 36.36 36.36 23.26 23.26 23.26 25.58 18.60
n03001627 chair 0 0 0 0 0 10.26 10.26 10.26 15.38 0
n03046257 clock 11.32 09.43 07.55 07.55 05.66 09.43 09.43 09.43 07.55 05.66
n03135532 cross 25.00 0 25.00 0 0 20.00 0 20.00 0 0
– From table 3, for almost all the concepts, statistical bounding box estimation
(i.e., heuristics 2, 3, 4, 5) is very helpful in order to improve the quality
of localization; for some concepts such as “frog”, re-scaling and shifting are
important, as this category is highly non-rigid while for other categories such
as “bear”, adaptation does not improve performances as “bear” localization
is less predictable. Note also that for rigid (and man-made) objects, such
as “cathedral” and “bicycle”, re-scaling is more important than shifting as
the proportions of the w-h dimensions, in these concepts, are very changing
while for others (including natural objects and also some other man-made
objects such as “airplane”, “balloon”, “bucket”, “camera”), the adaptation
of shift is more important than scale; as the variability of w-h proportions
is small in these concepts. In sum, bounding box re-scaling and shifting is
important for some concepts and less for others. This suggests, as a future
extension, to mix different heuristics for different concepts (and we already
observe this improvement in the “concept-by-concept” results).
4 Conclusion
We discussed in this paper, our participation at the ImageCLEF 2015 Scalable
Concept Image Annotation Task. Our runs are based on a two-step process
that decouples concept detection from localization. The former is achieved us-
ing SVMs trained with linear combination of elementary histogram intersection
kernels, while the latter is accomplished blindly using a simple statistical model
that allows us to generate multiple localization proposals (without image seg-
mentation). Observed results show that i) the accuracy of concept detection has
an impact on the performance of localization, and ii) the adaptation of scale
and shift of concept localization is essential to improve performances mainly for
concepts with a large variability in their extents.
A future possible extension, of this work, is to make concept localization non-
blind and also coupled with concept detection. Another possible extension is to
mix and select different localization heuristics for different concepts.
Acknowledgments. This work is supported in part by a grant from the French
Research Agency ANR (Agence Nationale de la Recherche) under the MLVIS
project.
References
1. V. Lavrenko, R. Manmatha, and J. Jeon, “A model for learning the semantics of
pictures,” In: Proc. of NIPS, 2004.
2. Jia Li and James Ze Wang, “Real-time computerized annotation of pictures,”
Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 30, no. 6,
pp. 985–1002, 2008.
3. H. Sahbi and X. Li, “Context based support vector machines for interconnected
image annotation (the saburo tsuji best regular paper award),” In the Asian
Conference on Computer Vision (ACCV), 2010.
4. Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik, “Rich feature
hierarchies for accurate object detection and semantic segmentation,” in Computer
Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on. IEEE, 2014,
pp. 580–587.
5. Olga Barinova, Victor Lempitsky, and Pushmeet Kholi, “On detection of mul-
tiple object instances using hough transforms,” Pattern Analysis and Machine
Intelligence, IEEE Transactions on, vol. 34, no. 9, pp. 1773–1784, 2012.
6. X. Li and H. Sahbi, “Superpixel based object class segmentation using conditional
random fields,” In the International Conference on Acoustics, Speech, and Signal
Processing (ICASSP), 2011.
7. Jian Yao, Sanja Fidler, and Raquel Urtasun, “Describing the scene as a whole: Joint
object detection, scene classification and semantic segmentation,” in Computer
Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on. IEEE, 2012,
pp. 702–709.
8. Andrew Gilbert, Luca Piras, Josiah Wang, Fei Yan, Emmanuel Dellandrea, Robert
Gaizauskas, Mauricio Villegas, and Krystian Mikolajczyk, “Overview of the Im-
ageCLEF 2015 Scalable Image Annotation, Localization and Sentence Genera-
tion task,” in CLEF2015 Working Notes, Toulouse, France, September 8-11 2015,
CEUR Workshop Proceedings, CEUR-WS.org.
9. Mauricio Villegas, Henning Müller, Andrew Gilbert, Luca Piras, Josiah Wang,
Krystian Mikolajczyk, Alba Garcı́a Seco de Herrera, Stefano Bromuri, M. Ashra-
ful Amin, Mahmood Kazi Mohammed, Burak Acar, Suzan Uskudarli, Neda B.
Marvasti, José F. Aldana, and Marı́a del Mar Roldán Garcı́a, “General Overview
of ImageCLEF at the CLEF 2015 Labs,” Lecture Notes in Computer Science.
Springer International Publishing, 2015.
10. Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and An-
drew Zisserman, “The pascal visual object classes (voc) challenge,” International
journal of computer vision, vol. 88, no. 2, pp. 303–338, 2010.
11. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei, “Imagenet:
A large-scale hierarchical image database,” in Computer Vision and Pattern Recog-
nition, 2009. CVPR 2009. IEEE Conference on. IEEE, 2009, pp. 248–255.
12. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva
Ramanan, Piotr Dollár, and C Lawrence Zitnick, “Microsoft coco: Common objects
in context,” in Computer Vision–ECCV 2014, pp. 740–755. Springer, 2014.
13. Paul Viola and Michael Jones, “Rapid object detection using a boosted cascade
of simple features,” in Computer Vision and Pattern Recognition, 2001. CVPR
2001. Proceedings of the 2001 IEEE Computer Society Conference on. IEEE, 2001,
vol. 1, pp. I–511.
14. Xuming He, Richard S Zemel, and MA Carreira-Perpindn, “Multiscale conditional
random fields for image labeling,” in Computer vision and pattern recognition,
2004. CVPR 2004. Proceedings of the 2004 IEEE computer society conference on.
IEEE, 2004, vol. 2, pp. II–695.
15. Pekka Rantalankila, Juho Kannala, and Esa Rahtu, “Generating object segmen-
tation proposals using global and local search,” in Computer Vision and Pattern
Recognition (CVPR), 2014 IEEE Conference on. IEEE, 2014, pp. 2417–2424.
16. Tomasz Malisiewicz, Abhinav Gupta, and Alexei A Efros, “Ensemble of exemplar-
svms for object detection and beyond,” in Computer Vision (ICCV), 2011 IEEE
International Conference on. IEEE, 2011, pp. 89–96.
17. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton, “Imagenet classifica-
tion with deep convolutional neural networks,” in Advances in neural information
processing systems, 2012, pp. 1097–1105.
18. Juergen Gall and Victor Lempitsky, “Class-specific hough forests for object detec-
tion,” in Decision Forests for Computer Vision and Medical Image Analysis, pp.
143–157. Springer, 2013.
19. Xuming He and Stephen Gould, “An exemplar-based crf for multi-instance object
segmentation,” in Computer Vision and Pattern Recognition (CVPR), 2014 IEEE
Conference on. IEEE, 2014, pp. 296–303.
20. Stan Z Li, Markov random field modeling in Image Analysis (was: Markov random
field modeling in computer vision)., 2011.
21. ImageNET webpage. http://image-net.org/about-publication.