Training Cascades with 2-3 varient items/increasing robustness.

asked 2016-10-26 18:22:21 -0500

Hello all!

I'm in the process of trying to train a cascade for a project I am working on, I have seen and followed through multiple examples of training a cascade that can be trained via a single image (for example of a watch face), the problem is that the cascade would only be able to find the watch face if it was the same style of watch. The other extreme of this problem is get hundreds of thousands of training images and train a general one to identify a watch face, but that's often tedious and difficult depending on the application. The problem is the training portion of opencv doesn't seem to like training images that look to similarly to each other, so you have to get a lot of data.

My question is this: Say that you know that in the application, you will have 2-3 different watch faces or 4-5 different styles of coffee cups, is it better to train 2-3 or 4-5 different cascade files? Or is there a way to train a good cascade with 1 image of each type in a single cascade?

This extends to the question of, if I were identifying a mug, and I want to be able to recognize it from a top down view as well as straight on...is there a way you can train with both sample images? Or is it better to do with with two separate cascades?

is there an advantage performance wise at run time to do it one way or the other?

Thanks in advance!

~Atomoclast

edit retag flag offensive close merge delete

Comments

"a cascade that can be trained via a single image" -- forget that idea, rather get a few hundred positives of the very same thing. . and no, you cannot train it on cats and dogs at the same time.

berak gravatar imageberak ( 2016-10-27 01:39:06 -0500 )edit