From min tree to watershed lake tree: Evaluation

Recently, several tree based hierarchical image descriptions have been proposed for image segmentation and analysis. This paper considers the problem of evaluating such algorithms. Recently we proposed a new algorithm for constructing the watershed lake tree by transforming the min tree structure as these two image trees share some similarities. We use this algorithm to illustrate the evaluation approach. The algorithm is evaluated by considering its computational complexity, memory usage and the cost of manipulating the resulting tree structure. Our results show that considerable care is needed when evaluating algorithms of this kind. In particular, comparisons cannot be made simply on the basis of computational complexity alone and other parameters such as image/tree ‘complexity’ also need to be considered.