Superpixel coherency and uncertainty models for semantic segmentation

Abstract

We present an efficient semantic segmentation algorithm based on contextual information which is constructed using super pixel-level cues. Although several semantic segmentation algorithms employing super pixel-level cues have been proposed and significant technical advances have been achieved recently, these algorithms still suffer from inaccurate super pixel estimation, recognition failure, time complexity and so on. To address problems, we propose novel super pixel coherency and uncertainty models which measure coherency of super pixel regions and uncertainty of the super pixel-wise preference, respectively. Also, we incorporate two super pixel models in an efficient inference method for the conditional random field (CRF) model. We evaluate the proposed algorithm based on MSRC and PASCAL datasets, and compare it with state-of-the-art algorithms quantitatively and qualitatively. We conclude that the proposed algorithm outperforms previous algorithms in terms of accuracy with reasonable time complexity.