In this article we propose a modification to the Iterated Conditional Modes algorithm applied to image segmentation. To do so, we introduce a new confidence criterion to determine whether or not the algorithm should keep updating the segmentation. This criterion relies on a transition matrix parameter and is based on the compactness of the segmented areas rather than the image global energy. By using this method, we ensure a more effective segmentation that gives more compact and meaningful areas than when using the traditional global energy criterion. Our initial experiments on Very High Resolution images have shown promising results.