The Importance of Object-Based Seed Sampling for Superpixel Segmentation

Abstract

Superpixel segmentation can be defined as an image partition into connected regions, such that image objects may be represented by the union of their superpixels. In this context, multiple iterations of superpixel segmentation from improved seed sets is a strategy exploited by several algorithms. The Iterative Spanning Forest (ISF) framework divides this strategy into three independent components: a seed sampling method, a superpixel delineation algorithm based on strength of connectedness between seeds and pixels, and a seed recomputation procedure. A recent work shows that object information can be added to each component of ISF such that the user can control the number of seeds inside the objects and so improve superpixel segmentation. However, it is uncertain how the added information impacts each component of the pipeline. Therefore, in this work, a study is conducted to evaluate such inclusion in the seed sampling procedure, partially elucidating its benefits. Additionally, we introduce a novel object-based sampling approach, named Object Saliency Map sampling by Ordered Extraction (OSMOX), and demonstrate the results for supervised and unsupervised object information. The experiments show considerable improvements in under-segmentation error, specially with a low number of superpixels.