Tools for ROCCH DET Curves
 

These tools are to bridge the gap between DET-curves and the ROC Convex Hull(ROCCH) . In short, ROCCH can be used to produce smoother DET-curves in a principled way.


Although the ROCCH-DET curve shows lower error-rates than the traditional stepped DET curve in places, the ROCCH-DET curve is not overoptimistic, in the sense that no point on the ROC convex hull will give a lower DCF value than that which can be obtained on the traditional ROC/DET curve. This holds (see papers here) for any DCF parametrization (i.e. for any values of Cmiss, Cfa and Ptar).

The ROCCH also gives a nice way to find an equal-error-rate (EER) solution, where Pmiss = Pfa exactly (see the example image above).

Here is version 0.1 of the MATLAB toolkit: rocch_det.zip.

  • Install on your MATLAB path.
  • Type 'rocch', or 'rocchdet' at the commandline for demos.
  • Type 'help function_name', or browse the source code for further info.
Alternatively, this functionality and much more, is now contained in the BOSARIS Toolkit.