A survey on the state-of-the-art superpixel segmentation

Abstract

In contrast to the rapid progress of superpixel segmentation, their methods are often compared only with classical approaches. Also, the previous superpixel methods categorizations are insufficient to cover the recent literature. In addition, although the inner color similarity usually underlies superpixel methods, both color homogeneity measures have important drawbacks. In this work, we fill these gaps by providing a new taxonomy for superpixel segmentation, a new color homogeneity measure, and an extensive comparison among 20 superpixel methods. Experiments show that the proposed measure, named Similarity between Image and Reconstruction from Superpixels (SIRS), is more robust to slight color variations than Explained Variation. Using SIRS and the commonly used superpixel metrics, we evaluated 20 superpixel segmentation methods and provided insights into the different approaches based on the clustering categories in our taxonomy.