What is mean FA?
When I trained my classifier I noticed that the FA was 0. I read a lot about that. Some people say that when the FA=0 the classifier is over-train and it can not detect anything. My classifier can detect my object. So what is the problem when FA is 0?
Classification techniques such as Haar Cascade assume that in a vast training set of samples it is impossible to classify correctly every single sample. So, in most cases, using haar cascade classification means accepting that there will be a number of false positives (negative examples classified as valid objects) and false negatives (positive examples not classified as such).
The False Alarm rate in the input cascade parameters is the percentage of false positives you are allowing the classifier to make in each stage during training. Setting FA to 0 while demanding a very high positive hit-rate will in many cases make it impossible for the training process to end.
To add to this clear remark, I want to say that the reason your training and detector is successfull is probably because you use very clinical environment to perform your detections. When having for example 50 apple images that need detection on a white background in all cases, then reaching a classifier with FA=0 will be quite easy and will work very good also. However, this reported number is only on your training data set.
Take for example 100 unseen examples, test your detector on it and them see how many wrong detections happen. You will immediatly see where problems arise.
If you take more complex training sets, for example cars in urban areas, using huge training sets, you will see that reaching a FA of 0 will be quite impossible.
Asked: 2013-12-03 19:36:30 -0600
Seen: 2,284 times
Last updated: Dec 04 '13
