What are Stages in traincascade.exe?

edit

To completely understand stages you should first read the Viola and Jones paper which discusses the principle of cascade classifiers which is a combination of weak classifiers to obtain a strong classifier. However compared to the official implementation, OpenCV has a slight addition, which are stages.

Basically

• From the complete pool of possible features the most discriminative feature between pos and neg training set is determined.
• That feature is used to train (in the default setting) a binary decision tree.
• However a single tree might not do good enough to reach a weak classifier (from the paper, a weak classifier is a classifier that does a bit better then random guessing, so higher accuracy then 50%).
• Therefore OpenCV combines stumps until the stumps combined reach the stage of a weak classifier existing out of multiple stump decision trees. Then it continues to the next stage.

Now how to define how much stages you need? Start by adding for example 25 stages, and see how the returned acceptanceRatio is small enough. I usually suggest until it goes below 10^-5. After that you are overfitting the actual model from my experience. I implemented an overload function that allows adding a acceptanceRatioBreakValue. Take a look here.

Addition: if you want a complete wrapup of the traincascade parameters and useful settings, you can get the just released book OpenCV 3 Blueprints which has a complete chapter discussing the interface.

Matlab Cascade Classifier is a wrapper over OpenCV one. There is a good tutorial for Matlab classifier.

Total hit and false alarm rates calculated using hit (stageHR) and false alarm (stageFAR) rates for each stage:

totalHR = stageHR^NumStages;    
totalFAR = stageFAR^NumStages.

Thus values of hit and false alarm rates for each stage depend on number of stages.

Number of stages influences the amount of training data: number of positive samples to use at each stage can be estimated using the following formula (see link):

number of positive samples = floor(totalPositiveSamples / (1 + (NumStages- 1) * (1 - stageHR)))

Following information is useful for choosing a number of stages:

• "Stages with a lower false positive rate are more complex because they contain a greater number of weak learners."
• "Stages with a higher false positive rate contain fewer weak learners."
• "Generally, it is better to have a greater number of simple stages because at each stage the overall false positive rate decreases exponentially. For example, if the false positive rate at each stage is 50%, then the overall false positive rate of a cascade classifier with two stages is 25%. With three stages, it becomes 12.5%, and so on."