Groupwise Registration Algorithms

Groupwise image registration aims to find sets of corresponding points on each of a group of images.
This is a challenging problem.
We have advocated various approaches, built on the Minimum Description Length approach. This postulates that the best correspondences will be those that lead to an appearance model which most compactly describes the training set of images.
For more details see our publications on correspondence methods

Example: Faces

Here's the mean of a model automatically constructed from over 300 images of different people from the XM2VTS databases:

This used an advanced algorithm based on our BMVC2005 paper.

Example: 3D MR Images of the Brain

Here are some cross sections through the mean obtained by applying the registration technique to 270 3D MR images of the brain: