Proposed clustering process SigMA, highlighted on a 2D toy data set of three Gaussians with variable covariance matrices and means. (1) The generated toy data set consisting of three bivariate Gaussians shown in white alongside 2σ confidence ellipses in color. (2) The clustering procedure starts off by estimating the density of the input data. (3) Next, a graph-based hill climbing step is performed in which points are propagated along gradient lines toward local peaks. (4) This gradient propagation results in a preliminary segmentation of input samples that typically is far too fine-grained. (5) These segmented regions are iteratively merged with a parent mode if a modality test along the MEP detects no significant density dip. (6) The final segmentation retains all three clusters.