Secrets of GrabCut and Kernel K-means

The log-likelihood energy term in popular model-fitting segmentation methods, e.g. Zhu-Yuille, Chan-Vese, GrabCut, etc., is presented as a generalized "probabilistic" K-means energy for color space clustering. This interpretation reveals some limitations, e.g. over-fitting. We propose an alternative approach to color clustering using kernel K-means energy with well-known properties such as non-linear separation and scalability to higher-dimensional feature spaces. Similarly to log-likelihoods, our kernel energy term for color space clustering can be combined with image grid regularization, e.g. boundary smoothness, and minimized using (pseudo-) bound optimization and max-flow algorithm. Unlike histogram or GMM fitting and implicit entropy minimization, our approach is closely related to general pairwise clustering such as average association and normalized cut. But, in contrast to previous pairwise clustering algorithms, our approach can incorporate any standard geometric regularization in the image domain. We analyze extreme cases for kernel bandwidth (e.g. Gini bias) and propose adaptive strategies. Our general kernel-based approach opens the door for many extensions/applications.