Superpixel Generation by the Iterative Spanning Forest Using Object Information

Abstract

Superpixel segmentation methods aim to partition the image into homogeneous connected regions of pixels (i.e., superpixels) such that the union of its comprising superpixels precisely defines the objects of interest. However, the homogeneity criterion is often based solely on color, which, in certain conditions, might be insufficient for inferring the extension of the objects (e.g., low gradient regions). In this dissertation, we address such issue by incorporating prior object information — represented as monochromatic object saliency maps — into a state-of-the-art method, the Iterative Spanning Forest (ISF) framework, resulting in a novel framework named Object-based ISF (OISF). For a given saliency map, OISF-based methods are capable of increasing the superpixel resolution within the objects of interest, whilst permitting a higher adherence to the map’s borders, when color is insufficient for delineation. We compared our work with state-of-the-art methods, considering two classic superpixel segmentation metrics, in three datasets. Experimental results show that our approach presents effective object delineation with a significantly lower number of superpixels than the baselines, especially in terms of preventing superpixel leaking.