Fine-structured object segmentation via neighborhood propagation

Abstract

This paper presents a novel method for the challenging task of fine-structured (FS) object segmentation. The task is formulated as a label propagation problem on an affinity graph. The proposed method has mainly three advantages. First, to enhance the completeness and connectivity of FS objects, we introduce a novel neighborhood system combining both local and nonlocal connections, with a robust scheme for edge weight calculation. Second, appearance models are explicitly incorporated into the energy function as a term of region cost. This helps to further preserve the connectivity of the fine parts for which the label information is hard to propagate correctly via neighboring pixels alone. Third, the resulting energy minimization problem has a closed-form solution with global optimum guaranteed, showing an advantage over the FS object segmentation methods that suffer from NP-hardness. To enrich the evaluation of FS object segmentation methods, we created a new challenging data set. It consists of 100 natural images involving diverse FS objects, with accurately hand-labeled ground truth. Extensive experimental results demonstrate that our method is effective in handling FS objects and achieves the state-of-the-art performance.