Pea's profile - activity

About the ready detection model that offered by OpenCV, like lbpcascade_frontalface.xml and haarcascade_frontalface_default.xml, etc.

I know I can use myself databases to train the new model.

but... If I want to use the ready .xml model and "additional add" my own databases to "retrain" the ready detection model. Is it OK to do? And how can I do?

I want to detect frontal face and hope to use the ready lbpcascade_frontalface.xml and add other training samples to improve the model. I want to add the negative samples to retrain and reduce the false positive rate and increse the detect speed.

The following picture is cascade detect method.

BUT How the TRAINGING method of cascade ?

Cascade consists of many stages. If the first stage is already trained. Then the second stage training will use the rest of the samples which passed at first stage, or use all samples which are composed of the failing samples(at first stage) with increasing weight and passing samples(at first stage) with decreasing weight（like adaboost algorithm method）.

Q1: Which cascade training method is, status 1 or 2 ?

status 1: using the rest of the samples which are pass at first stage.

status 2: using all samples which are composed of the failing samples(at first stage) with increasing weight and passing samples(at first stage) with decreasing weight.

Q2: What is the different of training cascade between two situation.

situation 1: set the detect rate and false positive rate at each stage.

situation 2: instead of setting the detect rate and false positive rate, control the number of feature at each stage. For example, use only one feature in each stage, and is it possible to get good detect rate ?

Please experts explain me about cascade training steps, thank you!