Selection of Visual Descriptors for the Purpose of Multi-camera Object Re-identification - Feature Selection for Data and Pattern Recognition - page 289

Table 12.10 Number of objects (mean values and standard deviations) in training and validation
sets for each identification task
Dataset

Number of Training set Validation set

identification Positive Negative Positive Negative

tasks Mean Std Mean Std Mean Std Mean Std

PERSON 282 1.00 0.00 3.00 0.00 1.00 0.00 2.94 0.24

VEHICLE 452 1.00 0.00 5.50 1.58 1.00 0.00 5.07 1.71

negative than positive samples, and its task is to find one valid object in a set of
4 other persons or 6 other vehicles. This is a very challenging task that is unlikely
to be carried out in the real scenarios with the application of spatial and temporal
constraints regarding cameras' geographical localizations.
During re-identification experiments, 18 different feature vectors have been used;
15 of them are formed with single descriptors, presented in Sect. 12.4 . Additionally,
3 combined vectors have been used that contain descriptors selected as the best
ones for re-identification of persons, vehicles and both groups, based on experiments
employing descriptor evaluation measures (Sect. 12.7.2 ). Due to different nature of
local image features and other descriptors, both parameter groups cannot be mixed
up within the same feature vector. Therefore, comparing to Table 12.9 , combined
feature vectors VectorPerson , VectorVehicle and VectorBoth contain
only standard (non-local) descriptors. They are presented in Table 12.11 .
Taking into account the number of feature vectors examined (18) and the number
of classifiers used (4), the total quantity of single identification tasks performed is
equal to 20,304 for PERSON dataset and 32,544 for VEHICLE dataset. In order to
make the results robust against the selection of training and validation samples, all
identification tasks are repeated 5 times, with different, random selection of both sets.
Furthermore, classifications with ANN and RTree (for the given training and valida-
tion sets) are further repeated 5 times due to the stochastic nature of the classifiers'
learning. Therefore, the re-identification tasks are computationally very complex and
full experiments take a few days to complete on a modern PC. The averaged results
from all iterations for PERSON and VEHICLE datasets are reported in Figs. 12.21
and 12.22 , separately for each classifier used. Table 12.12 presents results averaged
over all classifiers.
Absolute accuracy values of object re-identification (i.e. the results of finding the
one object in the group of a few) are not impressive at the first sight, but taking into

Table 12.11 Combined visual feature vectors
Feature vector name Descriptors included in the vector

VectorPerson CLDTrans , CLD , EHD , MomentInvGPSO

HistFull , MomentInvGPSO , VertTrace VectorVehicle

CLDTrans , MomentInvGPSO , HistFull , EHD VectorBoth