Information Technology Reference
In-Depth Information
75%
kNN
ANN
RTree
GBTree
70%
65%
60%
55%
50%
45%
40%
35%
30%
Vector
Person
Opponent
Sift
CLD
CLD
Trans
Vert
Trace
Sift
CMSP
MomentOpponent
CentNorm Surf64
EHD
Hist
Moment
InvGPSO
Vector
Both
HistFull
Surf128
Surf64
LBPHist
Vector
Vehicle
Fig. 12.21 Average object identification results for different feature vectors and classifiers, in
descended order of the accuracy averaged over all classifiers—PERSON set
75%
kNN
ANN
RTree
GBTree
70%
65%
60%
55%
50%
45%
40%
35%
30%
Surf64
Surf128
Sift
Opponent
Sift
HistFull OpponentMoment
Surf64 InvGPSO
Hist
LBPHist
Vector
Vehicle
Vector
Both
CMSP
Vert
Trace
Moment
CentNorm
EHD
CLD
Vector
Person
CLD
Trans
Fig. 12.22 Average object identification results for different feature vectors and classifiers, in
descended order of the accuracy averaged over all classifiers—VEHICLE set
account the large diversity of object poses and camera orientations in the datasets and
hard conditions of experiments that are unlikely to happen in real-life scenarios, they
could be interpreted as promising. However, the main purpose of the experiments
was to find the best descriptors for object re-identification.
Standard deviation values of results obtained for the tested classifiers and feature
vectors are similar to each other and low (less than 4 % points). This proves that
results obtained are independent of the selection of the specific objects into classifier
training and validation sets.
Analysing results for used classifiers it may be noticed that kNN classifier achieved
better results using local image features, comparing to other classifiers. In the same
time, it performed worse with combined feature vectors. Averaging over all feature
vectors, RTree classifier performed best for PERSON dataset (54.02 % accuracy)
and kNN—for VEHILCE dataset (54.49 %). On the other hand, ANN classifier per-
formed the worst for both datasets (49.72 % and 48.97 %, respectively). This may
be caused by the training set that contained a few object only, the amount too small
for proper neural network training. Although the accuracy span among classifiers
is small (approx. 5 % point), it is larger than the averaged standard deviation of
Search WWH ::




Custom Search