Database Reference
In-Depth Information
Another performance metric is Normalized Discounted Cumulative Gain
(
NDCG
). Given a query
q
,the
NDCG
at the depth
d
in the ranked list is defined by:
2
r
j
−
1
d
j
∑
NDCG
@
d
=
Z
d
(4.9)
=
1
log
(
1
+
j
)
where
r
j
is the rating of the
j
-th pair,
Z
d
is a normalization constant and is chosen
so that the
NDCG
@
d
of a perfect ranking is 1.
4.4.2
Objective Evaluations
4.4.2.1
Evaluation of Location-Based Recognition
In Fig.
4.11
, the CVT-based CBIR method with and without location-based GPS
filter is evaluated in both MAP and NDCG measurements for different database
sizes. In this case, original image query is used without any visual intent regulation.
The performance suffers a degradation with the increment of database size. For the
location-based recognition method, images with related geographical regions have
been firstly isolated from irrelevant images, and then, recognition by search algo-
rithm is implemented solely on the filtered dataset. Performance is maintained and
demonstrates that the system is applicable for dealing with large-scale databases.
For the location-based filter
ˆ
(
)
, the GPS effective region Q utilizes the Quadkey
level 5, which is equivalent to the resolution of 4,891 m in ground. Since landmarks
groundtruth includes various object types: from statuaries and buildings, to city
skylines and famous mountains, the aforementioned contextual filter will guarantee
the inclusion of enough potential image candidates. In summary, such an analysis
and investigation demonstrate the usage of location-based filter as an important tool
in mobile visual search and recognition.
q
0.9
1
50K
500K
1M
1.5M
2M
2M(GPS)
50K
500K
1M
1.5M
2M
2M (GPS)
0.8
0.7
0.6
0.5
0.9
0.8
0.7
0.4
0.3
0.2
0.1
0
0.6
0.5
0.4
3
5
6
9
10
12
3
5
6
9
10
12
top
N
top
N
Fig. 4.11
Top
N
returns for both MAP and NDCG evaluations with GPS context, on the whole
image itself as query
Search WWH ::
Custom Search