RISF: Recursive Iterative Spanning Forest for Superpixel Segmentation

Abstract

Methods for superpixel segmentation have become very popular in computer vision. Recently, a graph-based framework named ISF (Iterative Spanning Forest) was proposed to obtain connected superpixels (supervoxels in 3D) based on multiple executions of the Image Foresting Transform (IFT) algorithm from a given choice of four components: a seed sampling strategy, an adjacency relation, a connectivity function, and a seed recomputation procedure. In this paper, we extend ISF to introduce a unique characteristic among superpixel segmentation methods. Using the new framework, termed as Recursive Iterative Spanning Forest (RISF), one can recursively generate multiple segmentation scales on region adjacency graphs (i.e., a hierarchy of superpixels) without sacrificing the efficiency and effectiveness of ISF. In addition to a hierarchical segmentation, RISF allows a more effective geodesic seed sampling strategy, with no negative impact in the efficiency of the method. For a fixed number of scales using 2D and 3D image datasets, we show that RISF can consistently outperform the most competitive ISF-based methods.