Summary: Studies in GradientBased Registration
Assume we have two images I 1 and I 2 of the same scene taken from two different points
of view. Image registration means warping I 1 towards I 2 in some fashion, so that the two
images will be superimposed. Gradientbased registration relates to those techniques that
find the warping parameters from the grayscale similarities between the two images .
Alternative approaches find the warping parameters using local features (such as, points,
lines, etc.) extracted from the images (e.g. ).
The main advantage of the gradientbased techniques is that a correspondence between local
features in the images is not required (finding the correspondence can be exponential in the
number of features). However, the performance of gradientbased algorithms is adequate only
in the cases where difference in the viewing position is relatively small and the illumination
conditions in the two images are similar .
2 The Motion Model
Let I 1 (x; y) and I 2 (x; y) be two images of the same scene taken from different points of view.
If the illumination condition is about the same in the two images and the viewing positions
are not distant from each other we can assume the following:
I 1 (x + u; y + v) = I 2 (x; y) : (1)