Information Technology Reference
In-Depth Information
0.5 h. This gives a total number of 17 subsets; each contains from 10 to 39 objects
(25 on average).
For each image in the datasets, all visual descriptors have been calculated. Then,
within each subset, visual descriptor evaluation measures presented in Sects.
12.5.1
-
12.5.3
have been calculated. Unfortunately, it is not possible to calculate
SD
and
RS
indexes for local image features. Therefore, analysis for these descriptors is limited
to
DSIM
measure only.
Next, the results were aggregated for each visual descriptor according to the steps
described in Sect.
12.5.4
. For the reference, the same procedure has been performed
for full HUMAN and VEHICLE datasets, assuming there is only one subset in each
set containing all objects. Naturally, in this case it is not possible to calculate internal
stabilities
SI
.
Table
12.5
contains mean and standard deviation values of descriptor evaluation
measures (DEMs) over subsets of each dataset, for every visual descriptor. Because
for the local image features only dissimilarity measure
DSIM
can be calculated,
results for other measures are not available. The data is used to calculate individual
aggregation measures presented in Tables
12.6
and
12.7
.
Summarized aggregation results (for all datasets and DEMs) are presented in
Figs.
12.14
and
12.15
, and for each dataset separately—in Figs.
12.16
and
12.17
.
Based on the data, it is possible to evaluate and compare descriptors with each
other based on the data type and different criteria. Among non-local image fea-
tures, the best results (based on
RV
and
RP
) were achieved for colour-based
descriptors (
Hist
,
HistFull
), texture-based ones (
CMSP
) and statistical-based
ones (
MomentCentNorm
,
MomentInvGPSO
) in case of VEHICLE dataset, and
for colour-based feature (
CLDTrans
), edge-based one (
EHD
) and statistical ones
(
MomentCentNorm
,
MomentInvGPSO
) in case of PERSON dataset. It means
that the statistical features are universal enough to be suitable for both human and
vehicle identification task, however they are not the best in each category, sepa-
rately. The reason for high effectiveness of histograms for VEHICLE set results
from the vehicle nature: they are usually monochromatic, therefore histogram-based
vectors are more or less orthogonal. This is not happening in case of objects with
more complex colours, like people images. On the other hand, edge orientation-
based descriptor
EHD
provides good results (comparing to other descriptors) for
PERSON dataset because human pose changes less with different camera viewing
angles (the pose remains vertical) while vehicle edges orientation depends heavily
on the object rotation angle and camera viewing angle. It seems strange that the
VertTrace
extraction technique provides low descriptor evaluation measures for
PERSON. This might be due to characteristics of each image row colour that can
vary depending on people outfit. Hence, the mean value utilized to depict this variety
might to be inadequate for this task. However, a practical evaluation of this descriptor
is performed during re-identification experiments.
Ranking aggregation measures are not enough to select the best descriptors. It is
also important for the descriptor to guarantee stable results in case of different objects
in the dataset (it should not be dependent on particular objects in a dataset). Analysis
of the stability aggregation measures
SI
and
SE
reveals that the
CMSP
feature has
Search WWH ::
Custom Search