Image Segmentation Using Hierarchical Merge Tree

This paper investigates one of the most fundamental computer vision problems: image segmentation. We propose a supervised hierarchical approach to object-independent full image segmentation. Starting with over-segmenting superpixels, our approach uses a tree-like constrained conditional model to incorporate the region merging hierarchy with their likelihoods, which are predicted using an ensemble boundary classifier. Final segmentations can be inferred by finding the globally optimal solution to the model efficiently. We also present an iterative method for increasing the diversity of region merging hierarchies and emphasizing accurate boundaries by segmentation accumulation. Extensive experimental validation and comparison with other recent methods demonstrate that our approach achieves the state-of-the-art segmentation performance and is very competitive in image segmentation without semantic priors.