How is acceptance ratio caclulated?

asked 2014-07-21 07:53:24 -0500

AbbeFaria gravatar image

updated 2014-07-28 10:23:26 -0500

berak gravatar image

I've been working with Object Training for a while now, and I've found that generally cascades with an acceptance ratio of .0002 to .0007 are the most refined generally. However, I don't quite understand how acceptance ratio is calculated using traincascade. Could someone explain it to me or point me to a good reference that explains it?

edit retag flag offensive close merge delete

Comments

Hmmm strange... i get best results when i go up to 10to the power -5...

StevenPuttemans gravatar imageStevenPuttemans ( 2014-07-28 10:42:14 -0500 )edit

Interesting, normally when mine are in that range the cascade picks up a lot of false positives, or sometimes nothing at all. How large are your sample sets generally?

AbbeFaria gravatar imageAbbeFaria ( 2014-07-28 13:31:51 -0500 )edit

it depends on the situation. Industrial cases about 1000 pos and 2500 neg. For more challenging cases about 5000 pos and 15000 neg.

StevenPuttemans gravatar imageStevenPuttemans ( 2014-07-28 14:51:51 -0500 )edit

Do you generate virtual samples? I mean, I am always afraid that generating new training samples from rotating, translating or maybe scaling relatively few basic samples doesn't provide various enough samples. I mean, imagine that the training learns artifacts from the sample? In other words, how do you generate your training samples without burning the cmos of a camera? ;)

Doombot gravatar imageDoombot ( 2014-11-11 14:36:52 -0500 )edit
1

No I never generate virtual samples. By far they introducing indeed artificial artefacts that are far from desired. I always use only positive samples of real life occurrences of the object, meaning I just burn CMOS time :)

StevenPuttemans gravatar imageStevenPuttemans ( 2014-11-12 02:31:21 -0500 )edit

Ok so I am not the only one :)

Doombot gravatar imageDoombot ( 2014-11-12 06:56:31 -0500 )edit