Graphics Reference
In-Depth Information
Figure 4.19.
Log-polar bins used to construct the GLOH
descriptor.
Figure 4.20.
Support region centers used to construct
a DAISY descriptor. The circles' radii correspond to
the standard deviation of the Gaussian function at
each position used to weight contributions to the
orientation histogram.
function at each support region, as illustrated in Figure
4.20
. Like GLOH, the support
regions are arranged radially around a center point. Like the dominant orientation
estimation algorithm in SIFT, the Gaussian function at each center point specifies a
weighting function for the gradients in the neighborhood, so that points further from
the center contribute less to the orientation histogram.
4.2.3.3 Shape Contexts
Belongie et al. [
38
] proposed
shape contexts
as amethod for matching shapes, which
were modified by Mikolajczyk and Schmid [
328
] for feature point description. The
approach is similar to GLOH in that a log-polar location grid is constructed at the fea-
ture point location. However, instead of using the gradients of all points to construct
the histograms in each subregion, only edge points detected with the Canny detector
are allowed to contribute their gradient orientations. Each edge point's contribution
is further weighted by its gradient magnitude. Shape contexts are also used in 3D
feature detection, as discussed in Section
8.4.1
.
4.2.3.4 Spin Images
Spin images were originally proposed by Johnson andHebert [
224
] for describing fea-
tures in range data (see Section
8.4.1
). Lazebnik et al. [
264
] proposedmodifying them
to create feature descriptors for grayscale images. We simply compute a histogramof
quantized intensities for each of several rings around the feature location, after the
