Hybrid superpixel segmentation

Superpixel over-segment image into meaningful clusters so that pixels in each cluster belong to one object. Many state-of-art superpixel algorithms have to make trade-offs between different concerns. As a result, algorithms that can produce good result in some situations fail in another. In order to take advantage of different algorithms and at the same time avoid their limitation, we propose a new fusion approach based on an efficient lazy greedy optimization. It incorporates two different superpixel algorithms as its ancestors and produces a hybrid result. The result is then refined based on a novel energy function that consists of two terms. The region term uses histogram diffusion distance and enforces intra-region similarity from an overall perspective; the boundary term models interregion dissimilarity from a local perspective. In experiments, the result of proposed algorithm matches the best superpixel algorithm and shows outstanding performance over its ancestor algorithms in all the standard evaluation metrics.