ROC curves illustrate performance on a binary classification problem where classification is based on simply thresholding a set of scores at varying levels. Lenient thresholds give high sensitivity but low specificity, strict thresholds give high specificity but low sensitivity; the ROC curve plots this trade-off over a range of thresholds (usually with sens vs 1-spec, but I prefer sens vs spec; this code gives you the option).

It is theoretically possible to operate anywhere on the convex hull of an ROC curve, so this is plotted too. The area under the curve (AUC) for a ROC plot is a measure of overall accuracy, and the area under the ROCCH is a kind of upper bound on what might be achievable with a weighted combination of differently thresholded results from the given classifier.