Oversegmentation methods: a new evaluation

Abstract

Using superpixels instead of pixels has become a popular pre-processing step in computer vision. Currently, about fifteen oversegmentation methods have been proposed. The last evaluation, realized by Stutz et al. in 2015, concludes that the five more competitive algorithms achieve similar results. By introducing HSID, a new dataset, we point out unexpected difficulties encountered by state-of-the-art oversegmentation methods.