The images are organized into a tree with agglomerative clustering, using a measure of overlap as the distance. The reconstruction then follows this tree from the leaves to the root. As a result, the problems is broken into smaller instances, which a