How do superpixels affect image segmentation?

Abstract

Computationally intensive segmentation algorithms often operate on an image pre-segmented into small regions referred to as "superpixels". We investigate the effect of the choice of the pre-segmentation algorithm and its parameters on the outcome of the final segmentation. Three pre-segmentation algorithms are compared. To avoid the particularities of sophisticated segmentation algorithms, the final segmentations are built using agglomerative hierarchical clustering. These segmentations are evaluated using 300 images from the Berkeley Segmentation Dataset. This leads to useful insights about the variations in the final segmentation caused by the choice of the pre-segmentation algorithm. © 2008 Springer-Verlag Berlin Heidelberg.