Abstract

This paper addresses the search for a fast and meaningful image segmentationin the context of $k$-means clustering. The proposed method builds on awidely-used local version of Lloyd's algorithm, called Simple Linear IterativeClustering (SLIC). We propose an algorithm which extends SLIC to dynamicallyadjust the local search, adopting superpixel resolution dynamically tostructure existent in the image, and thus provides for more meaningfulsuperpixels in the same linear runtime as standard SLIC. The proposed method isevaluated against state-of-the-art techniques and improved boundary adherenceand undersegmentation error are observed, whilst still remaining among thefastest algorithms which are tested.

 

Quick Read (beta)