Space-varying color distributions for interactive multiregion segmentation: Discrete versus continuous approaches

State-of-the-art approaches in interactive image segmentation often fail for objects exhibiting complex color variability, similar colors or difficult lighting conditions. The reason is that they treat the given user information as independent and identically distributed in the input space yielding a single color distribution per region. Due to their strong overlap segmentation often fails. By statistically taking into account the local distribution of the scribbles we obtain spatially varying color distributions, which are locally separable and allow for weaker regularization assumptions. Starting from a Bayesian formulation for image segmentation, we derive a variational framework for multi-region segmentation, which incorporates spatially adaptive probability density functions. Minimization is done by three different optimization methods from the MRF and PDE community. We discuss advantages and drawbacks of respective algorithms and compare them experimentally in terms of segmentation accuracy, quantitative performance on the Graz benchmark and speed.