Region saliency as a measure for colour segmentation stability

This paper proposes a goodness function for colour segmentation, which allows to predict whether the segmented regions will be stable against noise, variation of lighting, and change of viewpoint. As such a measure, colour saliency defined from average border contrast is proposed. While the idea to maximise border contrast in segmentation is not novel per se, it is shown here empirically that maximisation of border contrast indeed leads to improved region stability. Experiments for three different algorithms show that the effect is independent of the particular functional principle of segmentation. Thus, the measure can be applied for the automatic and context-free optimisation of segmentation parameters.