The ML methods in OpenCV usually give an overall balanced error rate for both positive and negative samples.
For some applications, a very high recall rate with medium precision is needed. In such case, what can be done?
Your ideas and suggestions are highly welcome.
The answer depends on a particular classifier that you want to use.
For example, if you use a classifier based on decision trees you can use priors parameter to adjust weights of classes (see CvDTreeParams).
In case of SVM you can use distance d to the separating hyper-plane (see CvSVM::predict). This distance d corresponds to a confidence of the classifier. By default the decision boundary is d = 0 but you can manually change it to d = c, where c is a threshold (positive or negative). By varying c you shift the separating hyper-plane in a desired direction.
Asked: 2012-07-11 02:01:47 -0600
Seen: 1,194 times
Last updated: Jul 20 '12
