How does number of training stage affect the performance of cascade file?

asked 2013-10-25 12:06:27 -0600

sodeq
11 ●2

I have train my own classifier using 3 different stages , (10, 20, 30). All with 399 positive images, and about 180 negative images. I've tested each the classifier on to a video file, and analyze the ratio of positives and the actual object counts. I got a result that those three give different result, 10 stages gives the bad performance, 20 stages is better. But the 30 stages almost gives nothing, or, almost never detect the object as seen in the 20 stages. How does this could happened? As far as my understanding, the performance should be better if number of boosting stages were increased.

Thanks

edit retag flag offensive close merge delete

add a comment

1 answer

Sort by » oldest newest most voted

answered 2013-10-25 20:03:27 -0600

mccunnj
11 ●2

With 30 stages you might have over trained your classifier. I'm running into a similar problem. Did you crop positives from the video you are testing with? Finding the perfect parameters for cascade training is the key. For my project I have a trial journal where I note all parameters and the result. Trial and error takes a long time when training takes a few hours.