Selection of Visual Descriptors for the Purpose of Multi-camera Object Re-identification - Feature Selection for Data and Pattern Recognition - page 270

Table 12.2 Local image features listing
Feature symbol

Base local feature Colour information Feature vector

dimensionality
for a key point

Sift SIFT Grey-level 128

SURF Grey-level 64

Surf64

SURF Grey-level 128

Surf128

OpponentSift SIFT Opponent colour space 364

SURF Opponent colour space 192

OpponentSurf64

⊡ ⊤

R G

√ 2

−

⊡ ⊤

O 1
O 2
O 3

⊣ ⊦

R + G − 2 B

⊣ ⊦ =

√ 6

(12.12)

R + G + B

√ 3

where R , G , B denote appropriate channel colour values. This space is tuned to
mimic human perception of colour. The third channel is proportional to the intensity
channel of the HSV colour model while other two channels contain differences of
opponent pairs red - green and yellow - blue .
Table 12.2 lists all local image features used during object re-identification exper-
iments presented further in the chapter.

12.5 Descriptor Evaluation Methods

In order to evaluate the set of visual object descriptors objectively, valid measures
able to extract and highlight properties of the descriptors significant from the object
re-identification point of view, have to be employed. Therefore, two measures are
used that analyse inter- and intra-class scatter and clustering properties. They are
supplemented with a new, third method that is based on direct image pairs dissimi-
larity measurements.

12.5.1 RS Index

RS index is a parameter that estimates the degree of clusters dissimilarity [ 33 , 43 ].
This method is based on the Euclidean distances between instances of the dataset. The
measure is one of the approaches utilized for verifying the presented object descrip-
tion techniques. The R-squared parameter RS is calculated utilizing the following
relation: