Selection of Visual Descriptors for the Purpose of Multi-camera Object Re-identification - Feature Selection for Data and Pattern Recognition - page 291

Table 12.12 Object identification results averaged over all classifiers
PERSON dataset

VEHICLE dataset

Feature vector Mean (%) Standard Feature vector Mean (%) Standard

deviation (%) deviation (%)

VectorPerson 57.54 2.31 Surf64 60.56 1.31

56.40 1.65 60.44 1.41

OpponentSift Surf128

55.77 2.31 57.88 1.80

CLD Sift

CLDTrans 54.05 2.40 OpponentSift 54.68 1.16

53.54 1.95 53.85 1.99

VertTrace HistFull

52.54 1.93

OpponentSurf64 52.07

1.37

Sift

CMSP 51.72 3.22 MomentInvGPSO 51.38 1.56

MomentCentNorm 51.51

1.55 50.44 1.63

Hist

OpponentSurf64 51.39

1.75 49.17 1.27

LBPHist

EHD 50.19 2.10 VectorVehicle 49.16 1.50

49.19 2.25 48.99 1.88

Hist VectorBoth

MomentInvGPSO 49.16

2.50 48.32 2.21

CMSP

VectorBoth 48.97 2.57 VertTrace 47.99 1.75

48.76 2.09

MomentCentNorm 46.26

1.28

HistFull

48.58 1.69 45.34 1.53

Surf128 EHD

Surf64 48.12 1.55 CLD 43.26 1.62

45.97 2.25 42.59 1.69

LBPHist VectorPerson

VectorVehicle 45.04 2.08 CLDTrans 40.80 1.84

results achieved by each classifier (less than 2 % points). Therefore, for object re-
identification tasks, kNN and RTree classifiers seem to be the most suited.
Classification results obtained for different descriptors are coherent with the analy-
sis performed with descriptor evaluation measures (Sect. 12.7.2 ); the best candidate
descriptors obtained during this analysis turned out to perform better than majority
of other descriptors. Especially, CLD , CLDTrans and OpponentSIFT descrip-
tors provide better results for PERSON dataset and SURF64 , OpponentSIFT ,
HistFull and MomentInvGPSO —for VEHICLE dataset. The only major
difference between descriptor evaluation and identification results is caused by
VertTrace descriptor that performs very well for PERSON dataset, although
descriptor evaluation measures state otherwise. However, such a performance is con-
sistent with the descriptor extraction method that should provide meaningful clues
for people images.
Furthermore, it is clear that in case of vehicles, local image features outperform
other descriptors, while for people, descriptors based on whole image representations
are leading. This is due to the fact that vehicle images are basically texture-less and
very similar to each other, thus a proper vehicle identification may be performed by
looking at local, distinguishable features only.