=Paper=
{{Paper
|id=Vol-1175/CLEF2009wn-ImageCLEF-BinderEt2009
|storemode=property
|title=Fraunhofer FIRST's Submission to ImageCLEF2009 Photo Annotation Task: Non-sparse Multiple Kernel Learning
|pdfUrl=https://ceur-ws.org/Vol-1175/CLEF2009wn-ImageCLEF-BinderEt2009.pdf
|volume=Vol-1175
|dblpUrl=https://dblp.org/rec/conf/clef/BinderK09a
}}
==Fraunhofer FIRST's Submission to ImageCLEF2009 Photo Annotation Task: Non-sparse Multiple Kernel Learning==
https://ceur-ws.org/Vol-1175/CLEF2009wn-ImageCLEF-BinderEt2009.pdf
Fraunhofer FIRST’s Submission to
ImageCLEF2009 Photo Annotation Task:
Non-sparse Multiple Kernel Learning
Alexander Binder, Motoaki Kawanabe
Fraunhofer Institute FIRST, Kekuléstr. 7, 12489 Berlin, Germany
{alexander.binder, motoaki.kawanabe}@first.fraunhofer.de
23.08.2009
Abstract
In order to achieve good performance in image annotation tasks, it is necessary to com-
bine information from various image features. In our submission, we applied the non-
sparse multiple kernel learning for feature combination proposed by Kloft et al.(2009)
to the ImageCLEF2009 photo annotation data. Since some of the concepts of the Im-
ageCLEF task are rather abstract, we conjectured that color histograms are informative
for some categories such as sky and snow. Therefore we tried pyramid histograms of
pixel colors. Since the images are not aligned, we sorted histograms at different places,
when computing similarity of two images. Short description of our methods will be
presented and obtained results will be discussed in this manuscript.
Categories and Subject Descriptors
I.4 [Image Processing and Computer Vision]: I.4.8 Scene Analysis; I.4.9 Image Representa-
tion; I.5 [Pattern Recognition]: I.5.2 Design Methodology; I.5.4 Applications
General Terms
Measurement, Performance, Experimentation
Keywords
ImageCLEF2009, Photo Annotation, Support Vector Machine, Multiple Kernel Learning, Spatial
Pyramid Representation, Bag of Words
1 Introduction
Recent research results show that combining information from various image features is inevitable
to achieve good performance in image annotation tasks. With the support vector machine (SVM),
this is implemented by mixing kernels (similarities between images) constructed from different
image descriptors with appropriate weights. For instance, the average kernel with uniform weights
or the optimal kernel trained by multiple kernel learning (MKL) have been used so far. Recently,
Kloft et al.(2009) proposed the non-sparse MKL with Lp -regularizer, which bridges the above-
mentioned two extremes, i.e. the average kernel SVM and the standard MKL. The non-sparse
MKL is successfully applied to object classification tasks; it could outperform the two baseline
methods by optimizing the tuning parameter p ≥ 1 through cross validation. In our submission,
we applied the novel technique for feature combination to ImageCLEF2009 photo annotation data
based on bag-of-words [2] over SIFT features which is similar to a subset of the Pascal VOC 2008
winners [10, 3] submission. Furthermore, we tested sorted pyramid kernel with color histograms
which will be explained in Section 2.1.1.
2 Experiment Description
2.1 Features
We computed SIFT features [6, 7] on a dense grid of step size six over the color channels red,
green, blue, normalized red, normalized green, normalized blue, opponent color 1, opponent color2,
normalized opponent color 1,normalized opponent color 2, grey. We used radii 4,8,12,16 for the
SIFT feature supports.
For the bag of words models we combined the color channels into the following five sets: grey,
red-green-blue, normalized red-green-blue, opponent color 1 - opponent color2. The bag of words
prototypes have been obtained using kmeans clustering with 800000 SIFT features randomly
sampled from 3000 files. The number of prototypes was fixed to 4000 for each of the five channel
combinations. Finally we computed the bag of words for image tilings 1x1, 2x2 and 3x1 in the
sense of spatial pyrmaid levels [5, 1]. This results in 12 features of which we applied in each
learning setting at most eleven due to 32 bit memory restrictions. Furthermore we employed
pyramid histograms over color intensities features based on the color channels grey, opponent
color 1,opponent color 2, green and blue. We chose the number of bins per pyramid cell to be ten.
2.1.1 Modifications of pyramid histograms over color intensities
In order to improve the information content of the spatial pyramids for pyramid histograms over
color intensities (PHoCol), we sorted the cells within a pyramid level according to the slant of
the histogram. This reflects the notion that an image patch having high intensities in a certain
colour channel results in an intensity histogram slanted towards higher intensities. The slant was
determined via the entropy of the accumulated histogram of a histogram:
X a(h)i
e X
a(h)i =
e hk , a(h)i = P , sl(h) = −a(h)i ln(a(h)i )
kea(h)k i
k≤i
In addition to that we included the difference between sort costs in the kernel. Algebraically a
sort cost is just a mapping of the sort permutation of a pyramid level into the real numbers. In
our specific case we resorted to
X
sc(π) = C max(v(π(k)) − v(k), 0)
k
where v(i) is the height of the cell with index i with C being chosen so that the sort cost is
upperbounded by one and lies at a similar scale compared to the χ2 -distance. The intuition
behind such a sort cost is to punish large changes of the vertical position of an image patch
associated to a pyramid cell caused by sorting compared to the unsorted pyramid. This sort cost
is obviously invariant to changes in horizontal position of pyramid cells.
The sort cost modified kernel is a product of the χ2 -Kernel and a gaussian kernel over the
difference of the sortcosts between two features
k(x, y) = exp(−σ(dχ2 (hx , hy ) + (scx − scy )2 ))
. While it would be straightforward to replace this fixed weight kernel modification by an linearized
MKL-approach, we omitted it due to 32 bit memory restrictions. We also experimented with an
alternative formulation
k(x, y) = exp(−σ(dχ2 (hx , hy )(1 + (scx − scy )2 ))
for which the kernel property is not proven which however turned out to be nonnegative definite.
In the end we computed these features using tilings 4x4, no sorting, 3x1, sorted with sort cost,
16x16, sorted without sort cost over color channels noted above. While we do not regard such
features as good standalone features, we realized their speed and potential value for contributing
to some classifiers of the 53 classes in a proper MKL-procedure. We had low structured concepts
with high variance in mind like sky, water, portrait, daytime.
2.2 Kernels
We used the χ2 -Kernel with the exception of pyramid histograms over color intensities. All kernels
have been normalized.
2.3 Experiments
We used support vector machines with average kernels, sparse L1 -MKL, and the recently developed
general non-sparse Lp -MKL [4]. The regularization parameter was fixed to one.
We submitted one single kernel result, namely phows over red-green-blue with 1x1 tiling as
an early fallback, furthermore we submitted for each binary classifier the single best support
vector learning algorithm measured by AP/AUC and FRR rate score over ten-fold crossvalidation
(Lp -joint) using which generates three additional submissions
A) eleven bag of words kernels, namely red-green-blue, normalized red-green-blue and grey, all
tilings, and opponent color 1 - opponent color2, 1x1 and 3x1 tiling.
B) nine bag of words kernels, namely red-green-blue, normalized red-green-blue and grey, all
tilings,
C) 10 kernels, namely bag of words red-green-blue, all tilings, grey 1x1 tiling, PHoCol over
green and blue channels 4x4,3x1,16x16 tilings (various sort strategies)
D) eleven kernels, namely bag of words red-green-blue and grey, all tilings, opponent color 1 -
opponent color2 and normalized opponent color 1 - opponent color2, both with 1x1 tiling,
PHoCol over grey channels 4x4,3x1,16x16 tilings (various sort strategies)
E) 10 kernels, namely bag of words red-green-blue, all tilings, grey 1x1 tiling, PHoCol over
opponent color 1 and 2 channels 4x4,3x1,16x16 tilings (various sort strategies)
We used for the standard SVM and the Lp -MKL code from the shogun toolbox [9, 8].
The experiments were conducted with C++-code on an Opteron-Cluster running on a 32 bit
Linux which implied a hard limit for memory usage of 3GB. On average we used 20 CPUs.
2.4 Conclusion
On average combinations of only Bag of words features perform best. There exists classes for
which adding the low level color descriptors improves performance. Since each binary classifier
can be selected separately based on crossvalidation error adding the color features improves the
final submission. By comparing the results from A and B versus C, D, and E one can clearly
observe overfitting. In our point of view it comes from noisy feature extraction in connection with
small sample sizes for many of the smaller concept classes. We doubt that the applied features
permit a Bayes Error close to zero. In that sense there is still a need for better feature extraction
despite the success story of Bag of Words representations.
Below are AP and AUC scores over 5-fold crossvalidation, the result of the best method is
marked bold. For multiclass subproblems we computed for each method the average score over
the subproblem and marked the subproblem in total bold.
AP / methods A+avg A+L1 A+L1.5 A+L2 B+L1.0625 B+avg
average 43.904 42.751 43.823 43.839 42.474 43.353
Partylife 26.32 21.73 26.05 26.15 22.47 26.01
Familiy Friends 40.58 38.35 39.06 39.13 38.63 39.98
Beach Holidays 19.49 14.99 18.80 19.39 16.15 17.83
Building Sights 35.92 34.12 36.33 36.27 34.62 35.47
Snow 16.89 13.19 13.61 14.57 10.50 15.25
Citylife 48.26 45.17 48.12 48.17 46.18 48.00
Landscape Nature 63.51 62.50 63.43 63.46 62.48 62.84
Sports 5.06 4.10 5.01 4.94 3.32 4.74
Desert 17.33 12.96 17.15 17.16 14.62 12.90
Spring 18.00 19.03 18.02 18.11 17.21 17.69
Summer 45.62 44.74 44.97 45.02 43.32 44.31
Autumn 15.33 16.40 15.97 15.97 16.20 13.95
Winter 26.67 23.88 26.67 26.60 21.76 25.54
No Visual Season 94.02 93.51 93.89 93.92 93.65 93.85
Indoor 68.27 67.27 68.57 68.55 68.67 68.24
Outdoor 89.70 89.65 90.15 90.17 89.25 89.53
No Visual Place 39.55 38.91 39.61 39.42 39.09 39.51
Plants 62.96 62.88 63.79 63.73 61.57 62.77
Flowers 56.94 56.02 57.46 57.56 54.54 56.85
Trees 52.21 52.15 52.07 51.81 51.86 52.53
Sky 72.74 72.68 72.51 72.39 71.50 72.03
Clouds 73.84 73.44 73.88 73.97 71.74 73.16
Water 55.13 53.76 55.34 55.27 52.33 54.47
Lake 28.01 25.83 27.59 27.75 30.05 27.47
River 26.11 20.63 26.37 25.77 20.01 24.70
Sea 45.44 44.21 45.35 45.37 44.90 45.06
Mountains 33.77 29.76 33.76 34.05 30.71 32.17
Day 87.64 87.56 87.74 87.69 86.52 87.39
Night 51.99 50.24 52.26 52.35 49.91 51.33
No Visual Time 72.42 72.06 73.01 73.04 70.95 72.35
Sunny 30.47 27.89 29.08 29.18 28.01 28.28
Sunset Sunrise 69.66 69.20 69.78 69.72 68.84 68.99
Canvas 13.26 12.79 13.16 13.24 12.34 12.39
Still Life 30.88 31.11 31.07 31.09 30.40 31.30
Macro 23.90 23.22 23.45 23.30 24.33 23.98
Portrait 51.96 51.98 52.52 52.52 50.23 52.08
Overexposed 3.54 3.36 3.41 3.49 3.79 3.76
Underexposed 37.09 36.62 36.92 36.58 35.37 36.62
Neutral Illumination 97.50 97.44 97.53 97.51 97.46 97.50
Motion Blur 9.28 9.66 9.84 9.67 9.79 9.69
Out of focus 5.79 4.77 5.72 5.72 4.26 5.08
Partly Blurred 68.30 68.12 69.01 68.84 67.53 68.14
No Blur 88.35 87.78 88.38 88.44 87.77 88.10
Single Person 57.16 57.69 57.34 57.40 56.57 56.51
Small Group 27.10 26.80 26.53 26.55 24.89 24.75
Big Group 28.33 27.21 28.80 28.76 26.35 29.73
No Persons 89.49 89.10 89.61 89.62 89.05 89.17
Animals 51.23 48.75 51.12 51.18 48.12 49.79
Food 41.00 37.06 40.71 40.93 36.61 38.75
Vehicle 32.69 32.32 32.93 32.87 32.40 32.13
Aesthetic Impression 25.11 25.39 25.15 25.05 25.00 25.36
Overall Quality 35.47 34.78 34.23 34.35 36.23 36.06
Fancy 19.64 21.07 19.79 19.68 21.07 21.60
AP / methods B+L1 B+L1.5 B+L2 C+L1 C+L1.0625 C+L2
average 42.47 43.33 43.36 41.00 41.23 40.66
Partylife 21.60 25.33 25.49 19.17 20.19 24.39
Familiy Friends 38.49 39.59 39.67 35.21 35.58 35.55
Beach Holidays 15.91 17.70 17.75 14.78 15.91 15.10
Building Sights 33.91 35.31 35.35 33.06 32.73 33.96
Snow 13.21 14.89 15.07 8.27 8.88 11.58
Citylife 45.12 47.64 47.70 44.19 44.47 43.84
Landscape Nature 62.34 63.11 63.00 62.23 61.65 60.48
Sports 4.12 4.69 4.71 2.95 2.96 6.08
Desert 12.52 12.84 12.90 9.71 8.30 5.93
Spring 19.04 17.69 17.70 18.09 16.77 14.75
Summer 44.02 43.85 43.89 38.12 38.22 35.55
Autumn 16.19 14.52 14.47 11.30 10.03 10.23
Winter 22.72 25.44 25.46 20.28 24.18 25.27
No Visual Season 93.39 93.73 93.74 93.02 93.06 92.70
Indoor 67.33 68.23 68.23 67.48 67.62 67.28
Outdoor 89.41 89.86 89.85 88.74 88.53 88.58
No Visual Place 38.50 39.17 39.21 39.29 39.63 37.78
Plants 61.99 63.66 63.61 59.77 59.66 59.50
Flowers 55.44 57.28 57.27 50.70 51.05 51.07
Trees 52.19 52.59 52.63 50.24 50.71 47.61
Sky 72.42 71.86 71.93 71.10 70.95 70.05
Clouds 72.51 72.98 73.09 72.32 71.57 70.10
Water 53.06 54.76 54.88 50.91 51.47 49.64
Lake 25.70 27.83 27.91 25.26 23.65 21.38
River 20.60 24.62 24.78 16.44 19.06 15.78
Sea 43.38 44.57 44.71 44.29 43.31 36.80
Mountains 29.76 32.14 32.24 30.09 31.19 29.71
Day 87.23 87.41 87.40 85.96 86.12 85.43
Night 49.72 51.52 51.68 51.14 51.17 49.71
No Visual Time 71.78 72.82 72.81 70.28 70.30 69.92
Sunny 27.01 27.76 27.79 26.40 26.87 27.07
Sunset Sunrise 68.84 69.18 69.18 70.25 69.99 66.70
Canvas 12.06 12.30 12.34 11.57 13.38 16.78
Still Life 31.18 31.18 31.14 28.67 29.92 29.73
Macro 23.08 23.89 23.83 21.18 22.30 23.66
Portrait 51.81 52.24 52.30 48.72 49.06 47.23
Overexposed 3.36 3.75 3.75 3.85 4.47 5.45
Underexposed 36.46 36.38 36.36 37.31 37.50 35.14
Neutral Illumination 97.44 97.46 97.46 97.69 97.76 97.63
Motion Blur 9.68 9.53 9.68 9.06 8.83 7.56
Out of focus 4.50 5.00 4.99 4.37 4.52 4.47
Partly Blurred 67.86 68.70 68.64 66.20 66.02 65.50
No Blur 87.64 88.08 88.06 87.30 87.23 86.82
Single Person 57.36 56.73 56.69 55.60 55.22 55.38
Small Group 25.73 24.54 24.49 22.75 23.30 24.92
Big Group 27.29 29.49 29.69 21.07 21.65 22.14
No Persons 88.77 89.24 89.27 88.60 88.23 88.76
Animals 47.99 49.83 49.84 42.64 43.19 42.12
Food 35.67 38.73 38.86 31.27 32.81 31.34
Vehicle 32.25 32.05 32.09 32.53 32.84 29.48
Aesthetic Impression 25.31 25.22 25.23 24.84 24.79 25.07
Overall Quality 34.80 36.04 35.99 34.85 34.79 35.60
Fancy 21.04 21.33 21.45 21.64 21.37 20.57
AP / methods D+L1 D+L1.0625 D+L2 E+L1 E+L1.0625 E+L2
average 41.21 41.66 41.37 41.06 41.35 41.47
Partylife 17.50 19.70 22.45 18.34 19.98 21.11
Familiy Friends 36.01 36.83 36.26 35.30 36.03 36.36
Beach Holidays 12.21 14.56 15.28 13.02 13.76 12.45
Building Sights 32.85 34.30 34.50 32.66 32.16 33.74
Snow 9.68 10.72 13.61 7.06 7.71 8.35
Citylife 43.98 45.20 45.44 43.87 43.99 46.98
Landscape Nature 62.14 62.59 62.75 61.87 61.31 62.51
Sports 3.53 3.01 3.36 3.00 2.80 2.66
Desert 13.64 7.36 8.71 17.30 18.14 14.98
Spring 18.55 18.24 15.82 17.74 17.47 16.69
Summer 37.54 37.20 34.98 40.10 40.13 38.87
Autumn 9.86 9.83 9.50 11.57 10.47 10.46
Winter 19.03 21.60 23.65 19.58 20.84 21.23
No Visual Season 92.93 93.08 92.87 93.15 93.19 93.40
Indoor 67.01 67.86 67.85 68.23 68.15 68.70
Outdoor 88.75 88.88 88.86 89.01 88.74 89.07
No Visual Place 38.88 39.54 37.74 39.22 39.51 38.39
Plants 60.09 60.75 61.28 60.13 59.57 61.91
Flowers 51.07 51.90 52.34 52.61 53.51 53.94
Trees 50.89 51.10 50.56 49.59 49.50 50.39
Sky 71.19 71.22 71.12 71.62 71.33 71.02
Clouds 71.67 71.82 71.59 72.10 71.53 71.19
Water 50.45 51.20 49.63 50.04 50.60 48.63
Lake 31.18 32.39 27.44 25.86 26.82 25.78
River 16.73 17.97 19.58 17.88 20.34 23.36
Sea 43.97 43.29 40.31 44.22 44.45 41.00
Mountains 28.07 29.22 26.56 30.74 31.29 30.85
Day 85.93 86.09 85.62 85.99 86.09 85.97
Night 49.62 50.39 50.40 49.45 49.50 48.95
No Visual Time 69.97 70.35 70.72 70.16 70.20 71.11
Sunny 25.45 25.91 27.49 26.91 27.65 27.07
Sunset Sunrise 70.41 70.70 69.84 70.42 70.59 70.42
Canvas 13.17 15.84 16.30 11.56 12.82 14.65
Still Life 29.01 30.41 28.88 28.61 29.19 25.57
Macro 23.46 23.93 21.93 20.91 21.58 19.69
Portrait 50.27 50.28 48.92 48.97 49.18 49.61
Overexposed 3.75 4.39 5.01 3.71 4.36 7.66
Underexposed 36.58 36.64 35.03 35.60 36.03 34.67
Neutral Illumination 97.68 97.73 97.65 97.45 97.49 97.44
Motion Blur 10.92 10.46 10.39 9.08 8.78 7.48
Out of focus 4.56 4.62 4.38 4.38 4.30 4.07
Partly Blurred 66.29 66.82 66.50 66.27 66.13 67.29
No Blur 87.12 87.38 87.11 87.30 87.19 86.80
Single Person 57.08 57.46 57.12 55.24 55.06 56.90
Small Group 23.80 24.21 24.74 22.76 22.69 25.08
Big Group 21.29 20.92 22.56 20.24 20.49 24.04
No Persons 88.41 88.66 88.61 88.66 88.19 88.87
Animals 43.91 44.49 43.31 42.87 44.08 46.33
Food 31.78 34.27 33.60 31.10 32.82 30.10
Vehicle 31.97 32.35 31.29 31.48 32.57 32.43
Aesthetic Impression 25.01 24.91 24.80 24.82 24.83 25.66
Overall Quality 36.03 36.23 35.95 34.73 34.93 35.55
Fancy 21.05 20.99 20.40 21.76 21.64 20.61
AUC / methods A+avg A+L1 A+L1.5 A+L2 B+L1.0625 B+avg
average 82.22 81.69 82.19 82.17 81.56 82.03
Partylife 81.69 80.53 81.93 81.83 80.32 81.53
Familiy Friends 81.91 80.25 81.57 81.61 80.35 81.53
Beach Holidays 90.47 88.12 90.37 90.41 87.80 88.92
Building Sights 84.04 83.16 84.01 84.03 83.46 84.00
Snow 84.57 85.18 85.08 84.88 84.85 85.77
Citylife 82.93 82.00 82.67 82.68 82.35 82.63
Landscape Nature 89.23 88.67 89.12 89.13 88.94 89.01
Sports 66.34 67.72 66.44 66.37 64.14 68.03
Desert 88.84 87.31 88.79 88.80 88.17 89.08
Spring 83.08 82.20 82.88 82.92 81.15 82.35
Summer 81.18 79.60 80.83 80.88 79.49 80.30
Autumn 83.96 84.13 84.00 83.98 83.79 84.24
Winter 83.78 83.47 84.00 83.94 83.30 83.58
No Visual Season 81.43 80.05 81.07 81.09 80.30 80.89
Indoor 85.19 84.57 84.97 84.97 84.85 84.90
Outdoor 87.93 87.88 88.50 88.51 87.51 87.73
No Visual Place 75.93 75.49 75.57 75.52 75.60 75.87
Plants 86.34 86.16 86.47 86.43 85.97 86.11
Flowers 91.21 90.95 91.21 91.21 90.27 91.26
Trees 88.34 88.24 88.41 88.35 88.54 88.25
Sky 92.11 92.07 92.07 92.03 91.92 92.04
Clouds 94.60 94.48 94.62 94.62 94.05 94.47
Water 89.04 88.65 89.06 89.05 88.08 88.85
Lake 80.33 82.74 80.50 80.45 82.94 80.97
River 93.27 93.45 93.54 93.45 93.07 93.46
Sea 92.60 92.17 92.54 92.56 92.39 92.35
Mountains 89.21 87.31 89.11 89.16 88.38 88.47
Day 85.03 84.95 85.26 85.24 83.88 84.77
Night 88.37 88.01 88.38 88.36 87.98 88.46
No Visual Time 83.54 83.27 83.70 83.68 82.56 83.41
Sunny 72.16 70.01 71.59 71.65 70.66 70.54
Sunset Sunrise 95.02 95.03 95.07 95.05 94.79 94.85
Canvas 82.83 82.63 82.74 82.77 82.53 82.31
Still Life 78.53 78.58 78.52 78.49 78.70 78.64
Macro 79.79 79.30 79.73 79.70 77.81 79.21
Portrait 83.40 82.68 83.44 83.42 82.23 83.47
Overexposed 73.46 70.40 73.39 73.41 71.90 73.63
Underexposed 83.67 83.65 83.79 83.72 84.06 83.62
Neutral Illumination 79.59 79.34 79.66 79.58 79.46 79.49
Motion Blur 72.85 72.44 73.06 72.97 74.26 73.71
Out of focus 74.88 70.81 74.71 74.79 70.90 73.75
Partly Blurred 81.38 81.15 81.43 81.34 80.73 81.35
No Blur 82.38 81.86 82.54 82.50 81.77 82.11
Single Person 79.84 79.47 79.54 79.52 79.30 79.83
Small Group 72.53 71.16 72.14 72.18 71.61 71.44
Big Group 85.71 86.28 85.72 85.73 84.30 84.72
No Persons 83.33 82.88 83.57 83.60 82.49 82.89
Animals 84.03 82.98 83.89 83.91 82.57 83.18
Food 88.26 87.37 88.18 88.21 87.28 88.09
Vehicle 83.51 84.18 83.61 83.58 83.24 83.24
Aesthetic Impression 58.80 58.85 58.54 58.48 59.29 58.84
Overall Quality 60.84 60.26 59.94 59.88 60.70 60.91
Fancy 54.22 55.20 54.40 54.29 55.55 54.78
AUC / methods B+L1 B+L1.5 B+L2 C+L1 C+L1.0625 C+L2
average 81.49 81.96 81.97 80.50 80.70 80.04
Partylife 80.34 81.60 81.58 76.27 77.11 77.80
Familiy Friends 80.09 81.25 81.28 77.97 78.09 77.05
Beach Holidays 87.86 88.85 88.87 84.98 83.97 80.45
Building Sights 83.03 83.86 83.87 82.59 81.99 82.06
Snow 85.19 85.96 85.89 84.53 85.83 85.30
Citylife 81.85 82.36 82.36 81.03 81.08 80.22
Landscape Nature 88.58 88.90 88.89 88.70 88.43 88.26
Sports 67.75 67.88 67.96 65.52 65.92 68.97
Desert 87.33 89.03 89.08 85.94 84.73 78.61
Spring 81.50 82.17 82.23 79.86 79.77 78.35
Summer 79.02 79.94 79.97 76.77 77.31 75.93
Autumn 84.10 84.18 84.20 79.36 78.75 77.31
Winter 83.22 83.75 83.71 84.12 84.67 82.98
No Visual Season 79.73 80.58 80.60 78.56 78.93 78.20
Indoor 84.49 84.66 84.64 84.12 83.94 83.42
Outdoor 87.59 88.20 88.22 87.03 86.72 87.00
No Visual Place 75.04 75.52 75.50 75.10 75.35 74.12
Plants 85.85 86.16 86.15 84.41 84.42 84.22
Flowers 90.77 91.20 91.21 89.90 89.76 89.00
Trees 88.20 88.33 88.28 87.97 87.92 86.78
Sky 91.91 91.94 91.93 91.91 91.81 91.39
Clouds 94.28 94.44 94.46 94.07 93.96 93.98
Water 88.47 88.87 88.89 87.28 87.27 85.92
Lake 82.73 81.06 81.03 83.03 83.75 83.52
River 93.49 93.61 93.55 91.43 91.77 89.30
Sea 92.03 92.27 92.29 92.87 92.79 90.59
Mountains 87.28 88.42 88.44 87.35 88.19 87.64
Day 84.62 84.93 84.92 83.27 83.33 82.78
Night 87.98 88.47 88.47 88.02 88.24 87.33
No Visual Time 82.98 83.43 83.43 82.01 81.95 81.69
Sunny 69.09 70.08 70.08 68.85 69.60 68.20
Sunset Sunrise 94.78 94.90 94.89 95.06 94.97 94.35
Canvas 82.70 82.36 82.33 81.48 82.15 82.79
Still Life 78.60 78.62 78.59 76.33 76.60 75.78
Macro 78.80 79.07 79.05 76.65 77.69 78.66
Portrait 82.55 83.41 83.46 79.65 79.94 79.95
Overexposed 70.42 73.43 73.53 71.22 74.01 76.61
Underexposed 83.54 83.58 83.56 85.34 85.82 84.76
Neutral Illumination 79.30 79.41 79.42 80.91 81.27 80.21
Motion Blur 72.45 73.75 73.73 72.21 72.78 71.16
Out of focus 70.52 73.48 73.59 69.84 70.20 70.40
Partly Blurred 81.03 81.27 81.20 80.04 79.94 79.45
No Blur 81.68 82.17 82.14 81.19 80.95 80.50
Single Person 79.39 79.47 79.46 78.19 77.87 78.01
Small Group 70.56 71.17 71.20 69.19 69.62 69.50
Big Group 85.93 84.86 84.80 80.78 81.79 81.40
No Persons 82.37 83.09 83.12 81.28 80.83 81.21
Animals 82.53 83.08 83.10 79.62 79.76 79.46
Food 87.16 88.04 88.06 85.27 85.71 85.50
Vehicle 83.99 83.36 83.31 82.76 82.76 81.66
Aesthetic Impression 58.85 58.72 58.69 59.16 59.43 59.05
Overall Quality 60.26 60.17 60.16 60.73 60.86 59.67
Fancy 55.20 54.83 54.77 54.73 54.68 53.65
AUC / methods D+L1 D+L1.0625 D+L2 E+L1 E+L1.0625 E+L2
average 80.75 81.11 80.78 80.50 80.64 80.66
Partylife 78.10 78.76 78.56 76.60 77.75 79.03
Familiy Friends 78.36 78.88 78.13 78.21 78.49 78.31
Beach Holidays 85.76 85.48 83.74 85.66 85.61 88.17
Building Sights 82.47 83.22 82.80 82.41 81.89 82.32
Snow 82.96 84.19 86.24 83.35 83.55 82.85
Citylife 81.26 81.66 81.30 81.18 81.24 81.45
Landscape Nature 88.50 88.77 88.97 88.55 88.22 88.60
Sports 65.47 64.94 66.17 65.50 65.22 64.08
Desert 86.14 87.32 87.62 89.50 88.11 87.25
Spring 79.13 78.45 75.68 79.43 79.45 78.60
Summer 76.48 76.63 75.55 77.67 77.99 77.94
Autumn 80.63 80.08 79.74 79.59 79.34 78.75
Winter 81.91 82.97 83.15 83.59 83.95 83.25
No Visual Season 78.20 78.63 78.30 78.93 79.05 79.40
Indoor 84.10 84.34 83.87 84.49 84.37 84.32
Outdoor 86.89 87.04 87.10 87.22 86.91 87.29
No Visual Place 74.59 75.28 74.93 75.19 75.49 74.98
Plants 84.62 84.94 84.88 84.60 84.55 85.64
Flowers 89.98 90.08 89.20 89.75 89.84 89.68
Trees 88.36 88.66 88.35 87.79 87.72 87.74
Sky 91.81 91.86 91.75 91.77 91.61 91.45
Clouds 93.79 93.92 93.76 93.96 93.81 93.59
Water 87.21 87.40 86.47 87.18 87.29 86.11
Lake 84.61 85.14 86.23 82.28 82.50 83.34
River 92.65 92.67 90.84 91.69 92.24 91.99
Sea 92.86 92.81 91.48 92.96 93.04 91.68
Mountains 86.72 87.60 87.38 88.42 89.13 88.85
Day 83.08 83.31 82.91 83.09 83.18 83.06
Night 87.62 88.00 87.82 86.90 87.02 86.18
No Visual Time 81.80 82.22 81.92 81.91 81.94 82.08
Sunny 68.86 69.54 69.57 69.18 69.89 69.80
Sunset Sunrise 95.20 95.29 95.09 95.59 95.55 95.55
Canvas 84.35 84.69 84.36 81.23 80.76 80.34
Still Life 78.05 78.28 77.26 76.35 76.78 77.21
Macro 76.89 77.08 76.62 76.12 76.28 75.30
Portrait 81.02 81.09 80.50 79.83 80.01 81.46
Overexposed 70.06 73.56 77.29 71.25 74.35 80.14
Underexposed 85.32 85.57 84.99 83.80 84.23 83.22
Neutral Illumination 80.82 81.16 80.26 79.28 79.75 79.16
Motion Blur 71.74 72.23 71.49 72.16 72.68 70.19
Out of focus 71.58 72.28 71.77 69.67 69.75 71.36
Partly Blurred 79.98 80.32 80.01 80.02 79.85 80.25
No Blur 81.08 81.46 81.32 81.16 80.85 81.03
Single Person 78.75 78.97 78.81 78.17 77.93 78.44
Small Group 70.83 70.96 70.30 69.40 69.81 70.31
Big Group 83.51 83.33 82.76 79.97 81.32 81.48
No Persons 81.11 81.56 81.48 81.56 81.02 82.26
Animals 80.73 80.77 80.48 79.70 79.96 80.37
Food 86.12 86.82 86.88 85.18 85.27 84.32
Vehicle 83.38 83.35 82.60 82.72 82.87 82.58
Aesthetic Impression 58.96 59.28 58.72 59.10 59.25 58.60
Overall Quality 60.42 60.61 59.99 60.65 60.62 60.29
Fancy 55.11 55.19 53.92 54.80 54.80 53.60
Acknowledgements
This work was supported in part by Federal Ministry of Economics and Technology of Germany
under the project THESEUS (01MQ07018).
References
[1] Anna Bosch, Andrew Zisserman, and Xavier Muñoz. Representing shape with a spatial
pyramid kernel. In Proceedings of the 6th ACM international conference on Image and video
retrieval (CIVR ’07), pages 401–408, 2007.
[2] G. Csurka, C. Bray, C. Dance, and L. Fan. Visual categorization with bags of keypoints. In
Workshop on Statistical Learning in Computer Vision, ECCV, pages 1–22, Prague, Czech
Republic, May 2004.
[3] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and
A. Zisserman. The PASCAL Visual Object Classes Challenge 2008
(VOC2008) Results. http://www.pascal-network.org/challenges/VOC/
voc2008/workshop/index.html, 2008.
[4] Marius Kloft, Ulf Brefeld, Pavel Laskov, and Sören Sonnenburg. Non-sparse multiple kernel
learning, 2008. Abstract of the NIPS workshop on Kernel Learning: Automatic Selection of
Optimal Kernels.
[5] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. Beyond bags of features: Spatial
pyramid matching for recognizing natural scene categories. In Proceedings of the 2006 IEEE
Computer Society Conference on Computer Vision and Pattern Recognition (CVPR ’06),
volume 2, pages 2169–2178, New York, USA, 2006.
[6] David G. Lowe. Distinctive image features from scale-invariant keypoints. International
Journal of Computer Vision, 60(2):91–110, 2004.
[7] Krystian Mikolajczyk and Cordelia Schmid. A performance evaluation of local descriptors.
IEEE Transactions on Pattern Analysis & Machine Intelligence, 27(10):1615–1630, 2005.
[8] S. Sonnenburg. The shogun toolbox, http://www.shogun-toolbox.org/.
[9] S. Sonnenburg, G. Rätsch, C. Schäfer, and B. Schölkopf. Large scale multiple kernel learning.
Journal of Machine Learning Research, 7:1531–1565, 07 2006.
[10] K. E. A. van de Sande, T. Gevers, and C. G. M. Snoek. Evaluating color descriptors for object
and scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, (in
press), 2010.