My classifier detect the background of the object [closed]

asked 2013-12-12 13:14:04 -0500

ioanna gravatar image

updated 2013-12-15 06:04:08 -0500

I trained my own classifier using haar cascade to detect an object but the classifier, also, detect the object's background. Is there any solution for that?

Some images of my tools are https://www.dropbox.com/s/6ncqhnjkekgig5h/grabber_446.bmp https://www.dropbox.com/s/ihh953fheuk05w9/grabber_42.bmp

edit retag flag offensive reopen merge delete

Closed for the following reason question is not relevant or outdated by sturkmen
close date 2020-11-13 10:17:06.594376

Comments

1

Can you upload some example images?

GilLevi gravatar imageGilLevi ( 2013-12-15 04:08:22 -0500 )edit

@GilLevi I uploaded twoimages

ioanna gravatar imageioanna ( 2013-12-15 11:25:26 -0500 )edit

Hard to tell. Once you have the bounding box, perhaps you can try segmentation using the HSV color space (using Hue).

GilLevi gravatar imageGilLevi ( 2013-12-15 11:37:04 -0500 )edit

When I used HSV color space for threshing the result did not good cause the real background has many color, sometimes near to tool's color.

ioanna gravatar imageioanna ( 2013-12-15 12:15:07 -0500 )edit

The most obvious explanation is that you probably didn't train a good enough classifier. Try adding more positive and negative examples, and also try to train it with the lowest FA rate and highest hit rate possible.

Pedro Batista gravatar imagePedro Batista ( 2013-12-17 11:46:58 -0500 )edit

Is there a ratio for negative and positive images? Because I read that positive images is good to be less than negatives. Also, I don't have enough images so I create some samples from one image. Is that a problem?

ioanna gravatar imageioanna ( 2013-12-17 12:47:33 -0500 )edit

For real world, non-laboratory environments you really must have a considerable amount of positive samples. For you to have a n idea, I trained a aerial car detector with 1800 unique positive samples and 2800 negative samples.

Pedro Batista gravatar imagePedro Batista ( 2013-12-18 04:18:50 -0500 )edit