In reconstructing a 3D image from a so-called 2-1/2 representation, it is necessary to assign a correspondence between points in a set of images that represent the same point in the 3D image. This is often done by using an iterative closest point algorithm to do a least squares approximation.
One drawback of this method is that it requires a pre-alignment step. The alternative approach, presented here, is to view the two images as being obtained by the combination of a rotation and translation (a pose, to use the language of the paper). A genetic algorithm was constructed in which the “genes” are the translation rotation pairs. The algorithm was implemented using a master/slave pattern, where the slaves are divided among several processors.
The resulting algorithm avoids being trapped by a local minimum, and requires no pre-alignment. Results for both pathological examples and more realistic ones are given.