Figure 8. Example segmentations on natural images.

Top row: Despite having a very noisy boundary map, using additional cues allows us to segment the objects successfully. Middle row: Although there are many weak edges, region-based texture information helps give a correct segmentation. Bottom row: A failure case, where the similar texture of elephants causes them to be merged even though a faint boundary exists between them. For all rows, the VI ODS threshold was used. The rows correspond top to bottom to the points identified in Figure 7.