we are trying to use NCC to do template matching. Once a feature is
extracted using Harris Corner Detector, we try to track this feature
in the next images. To do the matching, we use the normalized
correlation. It considers changes in the position of the feature (x,y
coordinates) but not rotation.
The first idea is to rotate the template to several different angles
and do the correlation with those different rotations of the template.
We cannot rotate the template, because when rotating (for example
using imrotate) the template is padded with zero-value pixels, that
make NCC work badly.
Currently, we rotate the whole image instead of just the template, but
using imrotate adds noise to the image that makes the matching
Some other ideas here on how to do template matching with different
rotations? It is not valid using rotation invariant approaches, as we
need to know not only the displacement of the features but also their
||10/15/2008 11:02:15 AM