Is there a way to control how many stages a cascade classifier uses?

asked 2015-03-24 13:33:48 -0500

cytokinesis gravatar image

So, if I train up a classifier to 20 stages, and find that it is too specific, can I load it and programatically tell it to use only 18 of those stages?

edit retag flag offensive close merge delete



If I remember it correctly you just need to delete then stage 19 and 20 from the xml file. @StevenPuttemans: do I remember correct?

Guanta gravatar imageGuanta ( 2015-03-25 05:35:37 -0500 )edit

You dont even have to delete stage 19 and 20, simply change the -numStages parameter to 18 and do a retrian which will merge previous stages into a 18 stage model. However keep in mind that the training criteria when first trained depend on the parameter that was given. For example the maximumStageFalseAlarm rate is calculated depending on the number of stages. Therefore a merging of previous stages into a weaker classifier works BUT it is NOT the same as retraining everything with -numStages 18. I am still investigating if this is correct behaviour and will adapt the interface if not, because to me it seems that training of equal stages should be independently of the number of stages.

StevenPuttemans gravatar imageStevenPuttemans ( 2015-03-25 07:25:31 -0500 )edit