Graphics Reference
In-Depth Information
the quality measure
2
C
=
det
(
H
)
k trace
(
H
)
(4.4)
where k is a tunable parameter (frequently set to around 0.04; the lower the value
of k , the more sensitive the detector). When both eigenvalues are large, C will be a
large positive number, while C will be near zero if one eigenvalue is small. Figure 4.2
illustrates the candidate blocks from Figure 4.1 , along with the corresponding error
surfaces E
, eigenvalues of H , and quality measures C . We can see that both
eigenvalues are large for Candidates C and D, with correspondingly high quality
measures, while the quality measures for Candidates A and B are very low.
To detect features in an image, we simply evaluate the quality measure at each
block in the image, and select feature points where the quality measure is above
a minimum threshold. The resulting points are called Harris corners . Figure 4.3
illustrates Harris corners detected in an example image; we can see that most of
the features lie on actual image corners and other distinctive features, while few
features are found in flat regions or along edges. Since the test only depends on the
eigenvalues and not the direction of the eigenvectors, the detected feature locations
are approximately rotation-invariant (meaning that we would detect roughly the
same apparent features if the image were rotated). 4
We usually apply non-maximal suppression to the results of Harris corner detec-
tion, since theHarris qualitymeasurewill behigh formanypixels in theneighborhood
of a corner. That is, to avoid multiple detections for the same underlying corner, we
(
u , v
)
Figure 4.3. Harris corners detected in an image, using 15 × 15 windows and a threshold of one
percent of the maximum quality measure value. Non-maximal suppression is applied to avoid
generating many responses for the same feature.
4 The proper term is actually rotation- covariant . That is, the feature locations detected in a rotated
image will be approximately the same as the rotations of the locations in the original image.
However, the term “invariant” is often misused to mean “covariant” in the context of feature
detection.
 
Search WWH ::




Custom Search