“…With a series of alternating optimization steps, the algorithm can discover the two separate domains and estimate the ellipse parameters over each domain. In contrast, gradient-based approaches (Aigner and Jüttler, 2009;Yang et al, 2005) cannot handle such a case: either they identify a single continuous curve outlining the composite shape, or the composite shape falls apart at intersections, depending on the starting parameter values and the number of initial continuous curves. Similarly, the moving least squares approach (Guennebaud and Gross, 2007) gives false estimates near the intersections, especially if the noise magnitude is more substantial.…”