Information Technology Reference
In-Depth Information
The object of interest
S
is matched with the object
O
i
with the highest value of
the aggregated result
R
i
. If there are more than one object with the same maximum
values of
R
i
, the object with the highest ratio
D
i
of individual positive decisions is
chosen:
NV
ij
NI
i
j
=
0
j
=
0
1
NV
ij
d
ijk
1
for positive decision
D
i
=
,
d
ijk
=
(12.33)
0
for negative decision
NI
i
12.7 Experiments and Results
This section presents experiments carried out and their outcomes. First, datasets used
in experiments, involving human and vehicle images, are presented. Then, descriptor
evaluation measures are employed to assess particular descriptors in the task of object
re-identification. Based on the results, combined feature vectors for both object types
are proposed. The vectors (as well as single descriptors) are evaluated with object
re-identification experiments involving four classifiers.
12.7.1 Datasets
For the purpose of visual descriptor evaluation for multi-camera object identification,
two datasets were created. The first one, HUMAN, contains images of persons
walking indoors. The second set, VEHICLE, contains images of cars driving in
the parking lot near an office building. This location has already been exploited in
our previous experiments regarding parking event detection [
15
]. Both sets contain
images of objects acquired from various cameras with different horizontal and ver-
tical angles (Figs.
12.12
and
12.13
). Not all objects appear in all cameras, therefore
the number of object/camera pairs in the set is lower than the product of numbers of
objects and cameras. Datasets also differ significantly by the number of objects and
the duration of data acquisition. Furthermore, in case of the VEHICLE set, for the
re-identification task a subset of objects (based on the time criterion) has been used
comparing with the descriptor evaluation task. Object amount reduction has been
made due to the enormous computational cost of the object re-identification exper-
iments (see the Sect.
12.7.3
). All these differences (summarized in Table
12.4
)are
meant to evaluate visual descriptors in the large inconstancy of conditions, including
small and large datasets, different object types and recording conditions.
Each object image in the dataset is accompanied with a binary mask that denotes
image pixels belonging to the object. The masks have been obtained automatically
with moving object detection and tracking algorithms that were developed with
the Indect project and deployed successfully in the area of video event detection
[
16
,
46
]. Only masked pixels of object are used for visual descriptor calculating.
Search WWH ::
Custom Search