Selection of Visual Descriptors for the Purpose of Multi-camera Object Re-identification - Feature Selection for Data and Pattern Recognition - page 269

Fig. 12.9 Local image feature matching

12.4.8 Local Image Features

Local image features differ significantly from all other image descriptors presented
in the chapter. They are calculated in the nearest vicinity of key points that are found
automatically in the grey-level image. While dimensionality of the vector derived
from each key point is constant for each local image feature, the number of key
point varies significantly depending on the image. Therefore, the final local image
descriptor is a matrix with vectors for each key point stored in its rows. This means
that the width of the matrix remains constant while its height depends on the analysed
image (Fig. 12.9 ).
Key points in the image are found automatically using the Fast-Hessian detector
[ 5 ]. An adaptive, iterative procedure is used in order to find 25

10 % key points
for each image by changing Hessian threshold properly with each iteration.
Two types of local image features are used: SIFT (Scale Invariant Feature
Transform) and SURF (Speeded Up Robust Features). They are scale-, rotation- and
translation invariant, show robustness against illumination changes, and their values
remain very similar across a substantial range of affine distortions. SIFT features
[ 34 ] are derived from the image gradient magnitudes and orientations, sampled for
each pixel in the key point neighbourhood, while SURF descriptors [ 5 ] are based
on the sums of Haar wavelet responses in horizontal and vertical directions in the
square region centered around the key point. Depending on the number of sums, the
SURF vector may contain 64 or 128 elements per one key point.
Typically, local image features are calculated for grey-level images, therefore
they lack colour information. In order to overcome this inconvenience, features are
additionally calculated for each colour channel independently, thus increasing feature
vector dimensionality three times comparing to grey-level images. The opponent
colour space was used for this purpose [ 49 ]:

±