High Entropy Ensembles for Holistic Figure-ground Segmentation

In this paper we approach the task of figure-ground segmentation of natural images using a novel framework to generate highly collaborative tree-based structures, called High Entropy Ensembles (HEE).While many model combination frameworks adopt rejection rules to improve the classification time of the ensembles at the cost of restricting the interactions between the different elements in the structures, throughout our work we prove that, similarly to the Cascade Classification Model [3], when execution time is not critical, better results can be obtained when encouraging that kind of interaction by combining heterogeneous suboptimal classifiers into highly connected tree-based ensembles in which the different algorithms communicate with each other to let the strengths of one overcome the weaknesses of the others and vice versa. Inspired by randombased model combination approaches [2], we do not focus on looking for the optimal classifiers to be added to the HEE, instead we pick them from a pool of randomly configured segmentation algorithms. This randomness injection increases the effectiveness of HEE while also decreasing both the computational complexity of the model creation procedure and the risk of overfitting the training data, which is a common issue for most model combination frameworks.