Information Technology Reference
In-Depth Information
with low importance may be omitted from the classification. The Gradient Boosted
Trees (GBTree) algorithm [
19
] extends the idea of decision trees by employing a set
(an ensemble) of
weak
decision trees. A weighted sum of decisions from each tree in
the ensemble provides the final re-identification result. In the experiments described
here, the GBTree classifier is employed to solve the regression problem, returning
a value in the range
; the grater the output value is, the grater similarity of the
object feature vectors. A squared loss function is used for the ensemble training as
the most suitable one for this setup. 200 iterations of boosting algorithms during
GBTree training has been performed (resulting in 200 trees in the ensemble) and for
each iteration, the randomly chosen subset of 80 % training vectors has been used.
Random Forest (RTree)
. This algorithm also uses an ensemble of binary trees
(called
a forest
here) for enhanced classification [
6
]. In contrast to the GBTree algo-
rithm, each tree is trained with only a subset of object features, different for each tree.
The data not used for training of a given tree is then used for its validation. During
the re-identification stage, each tree processes the same vector of object features and
makes a decision. The final classification result is a class that was chosen by most of
the decision trees. The depth of each tree is set to 5, which seems to be an optimal
value based on initial experiments using the cross-validation. The minimum samples
required at a leaf node for it to be split is equal to 10, which accounts for the small
ratio of the total data. The size of the randomly selected subset of features at each
tree node that are used to find the best split is equal to the square root of the number
of elements in the feature vector. During the re-identification experiments, the RTree
classifier was configured to solve a binary classification problem. However, the direct
output (a discrete class label) is ignored. Instead, the probability (confidence) of the
sample belonging to the object of interest is used. It returns the number between
0 and 1 that is calculated as the proportion of the decision trees that assigned the
sample to the class representing an object of interest.
In order to employ diversified classifiers and visual features to object re-identifi-
cation task, a method of training the classifier and aggregation of the results must be
proposed. The solution presented in the chapter resembles the real-life application
scenario as close as possible: a classifier is trained with object image features obtained
from one camera and then it is used to recognize the same object in video frames
acquired from another camera (Fig.
12.10
). Therefore, one classifier is trained per
each object and each transition between fields of view of two particular cameras.
Positive training samples are formed by the image features of an object of interest
in the source camera. Negative samples for training are formed by features of other
objects from the dataset that passed the field of view of the source camera. Therefore
the classifier is learned to distinguish one particular object from other, possible similar
objects that could be found in the same area. Positive samples for validation are
created from the images of the object of interest in another, destination camera;
negative samples are formed with features of other object images. It is assured that
the same objects are not found in negative training and validation sets.
In order to identify the object
S
observed in camera
C
1
(the source one) in video
frames acquired from camera
C
2
(the destination one), it is necessary to find the
most similar object out of all candidates for matching. Let
O
i
,
[
0
,
1
]
=
...
i
1
NO
denote
Search WWH ::
Custom Search