Information Technology Reference
In-Depth Information
5.2
Whole Image Classification
As we said before, the local feature based feature classification is composed of two
steps (see Section 5). In previous section we have dealt with the issue of classifying the
local feature of an image. Now, in this section, we discuss the second phase of the local
feature based classification of images. In particular we consider the classification of the
whole image given the label
Φ
(
p
x
)
and the confidence
ν
(
Φ,p
x
)
assigned to its local
features
p
x
∈
d
x
during the first phase.
To this aim, we use a confidence-rated majority vote approach. We first compute a
score
z
(
p
x
,c
i
)
for each label
c
i
∈
C
. The score is the sum of the confidence obtained
for the local features predicted as
c
i
. Formally,
z
(
d
x
,c
i
)=
p
x
∈d
x
, Φ
(
p
x
)=
c
i
ν
(
Φ,p
x
)
.
Then, the label that obtains the maximum score is chosen:
Φ
(
d
x
) = arg max
c
j
∈C
z
(
d
x
,c
j
)
.
As measure of confidence for the classification of the whole image we use ratio between
the predicted and the second best class:
arg
max
c
j
∈C−
z
(
d
x
,c
j
)
Φ
(
p
x
)
ν
img
(
Φ, d
x
)=1
−
.
arg max
c
i
∈
C
z
(
d
x
,c
i
)
This whole image classification confidence can be used to decide whether or not the pre-
dicted label has an high probability to be correct. In the experimental results Section 7
we will show that the proposed confidence is reasonable.
6
Evaluation Settings
For evaluating the various classifiers we need at least: a data set, an interest points
detector, a local feature extractor, some performance measures. In the following, we
present all the evaluation setting we used for the experimentation.
6.1
The Dataset
The dataset that we used for our tests is composed of 1,227 photos of landmarks lo-
cated in Pisa and was used also in [2]. The photos have been crawled from Flickr, the
well known on-line photo service. The dataset we built is publicly available. The IDs
of the photos used for these experiments together with the assigned label and extracted
features can be downloaded from [10]. In the following we list the classes that we used
and the number of photos belonging to each class. In Figure 1 we reported an example
for each class that are:
Leaning Tower
(119 photos);
Duomo
(130 photos);
Battistero
