Database Reference
In-Depth Information
Table 5.1
Average recognition accuracy (%). Performance comparison
for the Singapore Landmark Dataset
Method I
Method II
No.
Train:test
Difference
SA + SVT + Re-ranking
SA + SVT
1
90:10
81.87
78.93
2.94
2
80:20
72.67
70.68
1.99
3
70:30
65.43
63.94
1.49
4
60:40
59.67
59.11
0.56
5
50:50
56.8
55
1.8
Table 5.2
Average recognition accuracy (%). Performance comparison
for the World Landmark Dataset
Method I
Method II
No.
Train:test
Difference
SA + SVT + Re-ranking
SA + SVT
1
90:10
77.25
76.01
1.24
2
80:20
67.89
66.36
1.53
3
70:30
60.58
59.41
1.17
4
60:40
56.82
55.67
1.15
5
50:50
52.24
50.83
1.41
The experimental results were analyzed in two parts. The first part presented the
performance of the re-ranking method with the procedural parameter adjustment. In
the second part, the recognition accuracy of the re-ranking method was measured
at different ratios of training images and testing images, and compared to the
previously proposed method discussed [
346
].
The procedural parameters for the re-ranking approach are:
R
[the number of top
retrievals for training in Eq. (
5.20
)], and the weight parameters
ʵ
,
ʵ
2
in Eq. (
5.28
).
1
The system was tested at
(
ʵ
,
ʵ
)=
{
(
0
.
2
,
0
.
8
)
,
(
0
.
5
,
0
.
5
)
,
(
0
.
8
,
0
.
2
)
}
. For each
1
2
setting, the value of
R
was varied from
R
20. It was observed that the
recognition accuracy for the Singapore landmark database at 90:10 ratio of training
and testing was at 81.6 % at
R
=
3
,...,
=
4 and
(
ʵ
,
ʵ
)=(
0
.
5
,
0
.
5
)
.Next,
R
was fixed at
1
2
4 and the values of weights
(
ʵ
,
ʵ
)
were varied from 0.1 to 0.9. The best result
1
2
was obtained at 81.9 % at
3. According to Eq. (
5.28
), this
result indicates that the new query should be modified by adding the most important
features and subtracting the less important features. The parameter
ʵ
≥
0
.
6 and
ʵ
≤
0
.
1
2
ʵ
1
should be
greater than
2
.
In the second part of the experiment, both of the datasets were used. The dataset
in each category was divided into two groups; training group and testing group, in
the ratios of 90:10, 80:20, 70:30, 60:40 and 50:50. This experiment compared the
two methods, which were the re-ranking method implementing SA, SVT, and re-
ranking, and the SA and SVT methods demonstrated in [
346
]. The experimental
result is shown in Tables
5.1
and
5.2
.
ʵ
Search WWH ::
Custom Search