Content-Adaptive Superpixel Segmentation Via Image Transformation

Abstract

We propose simple and efficient method that produces content-adaptive superpixels, i.e. smaller segments in content-dense areas and larger segments in content-sparse areas. Previous adaptive methods distribute superpixels over the image according to image content. In contrast, we transform the image itself to redistribute the content density uniformly across the image area. This transformation is guided by a significance map, which characterizes the ‘importance’ of each pixel. Arbitrary superpixel algorithm can be utilized to segment the transformed image into regular superpixels, providing a suitable representation for subsequent tasks. Regular superpixels in the transformed image induce content-adaptive superpixels in the original image facilitating the improved segmentation accuracy.