Bug in ROC curves
I obtained a strange ROC curve using the evaluate.py script of Bob. The weird thing is the fact the curve doesn't cross the upper right corner (that is the 100%-100% coordinate) and it should be like this. Oleg and me already looked for the source of the issue and we think the problem is coming from the FAR and CAR computation. In our case, we have scores going from 0.0 to some positive value (1.0 being the maximum) and the higher, the more likely we match. We didn't check but it seems like it happens some genuine scores are 0.0 and they are not count when the threshold is defined to 0.0, causing the decay in the CAR when the FAR is 100%. I don't know if there is the same problem to the coordinate 0%-0% since the plot is in log scale. I could have reported the plot but I wasn't be able to share the score files. Just explain me how to proceed if needed.