Summary: Multiresolution image alignment
Pedro M. Q. Aguiar
Institute for Systems and Robotics / IST
These notes address the problem of aligning, or registering, two
images (or image patches). This is done by estimating the motion of
the brightness pattern between the two images in a global way, i.e.,
without any pre-processing step such as selecting and matching salient
features. The resultant optimization problem is solved in an efficient
way by using a Gauss-Newton method in a multiresolution pyramid.
These notes address the estimation of the two-dimensional (2D) motion of
the brightness pattern in the image plane. This is a crucial step in solving
any motion analysis task. For example, the problems of building a mosaic,
or panorama, from uncalibrated images, segmenting out moving objects, and
inferring three-dimensional (3D) structure from video, require the step of
estimating 2D motion to accomplish their higher level goals.
The problem of estimating the 2D motion of the brightness pattern has