Information Technology Reference
In-Depth Information
Let's denote
w
2
⎡
⎣
⎤
⎦
Σ
−
1
/
2
1
0
.
.
.
P
=
(21)
w
2
0
Σ
−
1
/
2
K
(22)
H
t
H
1
;
;
H
t
K
]
=[
···
(23)
VCV
T
=
P
(
I
−
)
φ
(
Z
t
)
,
(24)
where
H
t
summarizes information from the
t
th
training image.
With Equations 17, 20 and 21, it can be shown that the per feature vector contri-
bution function can be written as in Equation 25.
K
k
=
1
∑
t
α
t
H
k
•
1
n
k
Pr
W
j
=
(
k
|
z
j
)
z
j
.
(25)
5
Experiments
In this paper, we carry out object localization experiments using the proposed effi-
cient object localization approach based on the Gaussinized vector representation.
We compare the detection performance with a similar object localization system
based on the generic histogram of keywords. In addition, we demonstrate that the
proposed normalizing approach can be effectively incorporated in object localiza-
tion based on Gaussianized vector representation.
5.1
Dataset
We use a multi-scale car dataset[1] for the localization experiment. There are 1050
training images of fixed size 100
40 pixels, half of which exactly showing a car
and the other half showing other scenes or objects. Since the proposed localization
approach has the benefit of requiring no heuristics about the possible locations and
sizes of the bounding boxes, we use a test set consisting of 107 images with varying
resolution containing 139 cars in sizes between 89
×
85. This dataset
also includes ground truth annotation for the test images in the form of bounding
rectangles for all the cars. The training set and the multi-scale test set are consistent
with the setup used in [5].
A few sample test images of the dataset is shown in figure 1. Note that some test
images contain multiple cars and partial occlusion may exist between different cars
as well as between a car and a “noise” object, such as a bicyclist, a pedestrian or a
tree.
×
36 and 212
×
Search WWH ::
Custom Search