best resolution for negatives on traincascade?

asked 2015-12-22 09:12:20 -0500

Delauney gravatar image

I'm training a cascade classifier now and it's on the 6th stage and has been there for more than 48 hours so far. It's slowly working through the negatives. I'm wondering if the resolution of the negatives is causing this? They are large-ish images (roughly desktop background resolution). I noticed that using cascade classifiers is much faster with smaller images so I can only imagine that training with large images slows it down as well. Any experience with this?

So what is the ideal resolution for negatives?

edit retag flag offensive close merge delete


I don't think the resolution of the negative images is the problem there. Negative images are used to extract negative windows of the same size as the provided positive samples. The larger the negative images, the more negative windows can be extracted, but that's not a problem (more like an advantage I'd say).

LorenaGdL gravatar imageLorenaGdL ( 2015-12-22 11:52:58 -0500 )edit

So this isn't unexpectedly slow?

Delauney gravatar imageDelauney ( 2015-12-22 12:11:54 -0500 )edit