Intrinsic quality analysis of binary partition trees

The binary partition tree (BPT) is a well-known hierarchical data-structure, frequently involved image segmentation procedures. The efficiency of segmentation based on BPTs depends on the segmentation process ("how to use a BPT?"), but also on the quality of the data-structure ("how to build a BPT?"). In this article, we propose a scheme for BPT quality analysis, with the purpose of answering the latter question. It relies on the observation of the very structure of a BPT, with respect to a given ground-truth example. Our hypothesis is that such intrinsic scheme can bring relevant clues about the ability of a BPT to provide correct segmentation results. Experiments carried out on satellite images illustrate the relevance of this scheme.