Waterfall Segmentation of Complex Scenes

We present an image segmentation technique using the morphological Waterfall algorithm. Improvements in the segmentation are brought about by using improved gradients. These are based on the detection of object boundaries learnt from human segmentations introduced by Martin et al. (2004). We avoid the usual pitfall found when applying Watershed algorithms to these boundaries, namely that the boundary lines usually contain gaps, by making use of distance functions on the boundary image. Two types of distance function are used: the classic distance function and a distance function for numerical images recently introduced by Beucher (2005). Resulting segmentations are compared to human segmentations using the Berkeley segmentation benchmark. The benchmark results show that the proposed segmentation algorithm produces segmentations comparable to those produced by the Normalised Cuts algorithm.