- ...set
- An
alternative method is to compute the principal components of
22#22, the covariance of 21#21.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...be
- In practice, the dimensionality may even be
smaller than m - some of the diagonal values of 29#29 may be
zero, or small enough to be affected by limited machine precision. In
this case, the corresponding eigenvectors are removed.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...magnitude
- We normalise the input vectors
to have unit magnitude in order to counter the effects of
lighting variation.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...orientation
- While this constraint
can be readily relaxed, we will later demonstrate a method for
estimating orientation under the conditions that the database
orientation is fixed.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...image
- Note that this does not guarantee that the landmark
images will be unique, since the environment may contain
self-similarities.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...acceptable
- The scale of this ellipse can be controlled by a
user-defined threshold.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...noise
- All of our experiments
employ the Canny edge map without thresholding, since we are
primarily interested in edge density. For this reason, the edge map
will contain a large amount of highly unstable, low intensity
noise.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.