Figure 4.

ROC curves compare breakpoint detection of global models on two annotation data sets. ROC curves for the training error with respect to the breakpoint annotation data are shown as colored lines. The curves are shown in 3 panels zoomed to the upper left region of ROC space to avoid visual clutter. Each curve is traced by plotting the error of a model as the degree of smoothness is varied, and an empty black circle shows the global model chosen by minimizing the error with respect to all annotations. Algorithms with no tuning parameters are shown as black dots. Note that some ROC curves appear incomplete since some segmentation algorithms are not flexible enough for the task of breakpoint detection, even though we ran each algorithm on a very large range of smoothness parameter values.