Revision history [back]

First of all, you cannot specify an accuracy on a classifier alone. Accuracy is always specified in function of the training, validation and test set. From those sets you need to calculate the false positives, true positives, false negatives and true negatives and create a precision recall curve on your data.

See this links:

1. http://en.wikipedia.org/wiki/Precision_and_recall
2. http://www.cs.cornell.edu/courses/cs578/2003fa/performance_measures.pdf

So basically it all depends on the data that you feed to it. For example, detecting faces will be much easier and thus having a better precision/recall position if it is in a controlled environment behind a laptop and needs to detect a single face. Take that same detector outside, detecting faces in the wild and the precision/recall will change drastically.

I guess you should read about machine learning and performance measurements. Google on these items for more information.

As for retrieve a kind of score for each detection, using a cascade classifier, what you could do is look at the amount of overlapping detections. The more overlapping detections, the better the end result will be. Try your detectMultiScale with 0 for the minNeighbors parameter and you will immediatly see what i mean here.