Wrapping up: Week 6

In this week we have seen how we can fit a model to a target surface. To this end we have introduced a variant of the classical Iterative Closest Points (ICP) algorithm.

The ICP algorithm alternates between two steps. It finds corresponding points by selecting the closest point on the target, and then fits the model to the surface using these points. As we fixed correspondences (and hence known deformations) we could apply Gaussian Process regression to compute the best fit of the model to the points. By iterating this procedure, we should get closer and closer to the true correspondences. We have seen how this algorithm can be implemented and visualized in Scalismo Lab, using only a few lines of code.

The ICP method is an approach for establishing correspondence, and hence closes the big conceptual gap, which we had left open in our exposition of building shape models in Week 3.

However, we would like to emphasize, that while the approach works and has been successfully used in practical applications, it is definitely not the final solution to the registration problem. Many other, much more sophisticated approaches for establishing correspondence have been proposed and also within Scalismo we have implemented several other registration approaches.

Yet, for this course here we continue along the line of the ICP algorithm. It turns out that by applying a minor modification to the algorithm, we can derive one of the most popular methods for model based image segmentation – namely the Active Shape Model fitting algorithm. This will be the topic of next week.