Modified grabcut for unsupervised object segmentation

propose a fully automated variation of the GrabCut technique for segmenting comparatively simple images with little variation in background colour and relatively high contrast between foreground and background. The interactive trimap generation central to the original formulation of GrabCut is replaced by a tentative approximation of the background using active contours. Instead of waiting until convergence of the iterated graph cut, we terminate as soon as the Gaussian models of foreground and background are (locally) maximally separated. We demonstrate that this results in equivalent segmentation quality at significantly lower cost. A comparison with three alternative segmentation techniques, including normalised cut, indicates that the method is eminently suitable for the chosen image domain.