Spatial Repulsion Between Markers Improves Watershed Performance

The Watershed Transformation is a powerful segmentation tool from Mathematical Morphology. Here we focus on the markers selection step. We hypothesize that introducing some kind of repulsion between them leads to improved segmentation results when dealing with natural images. To do so, we compare the usual watershed transformation to waterpixels, i.e. regular superpixels based on the watershed transformation which include a regularity constraint on the spatial distribution of their markers. Both methods are evaluated on the Berkeley segmentation database.