Ask Your Question
0

HaarTraining: Best way to minimally crop +ve images (for efficient detection)

asked 2013-05-12 23:37:22 -0500

vinayverma gravatar image

updated 2013-05-12 23:38:26 -0500

First of all, I am not talking about the utilities that can be used to crop positive images.

I want to know the best way to crop the object of interest so that minimal number of +ve images produce good detection results. There can be two types of objects,

  1. Which can be cropped easily with a rectangle including considerable number of object features (objects roughly in rectangular/circular shapes)

  2. Whose features cannot be captured easily in rectangle without keeping the background too in the cropping rectangle (objects in shapes like star maybe; like aircraft)

I have attached a sample cropped image of an aircraft. I am trying to build a cascade classifier with similarly cropped images but my training doesn't proceed beyond 4 stages (False detection rate achieved with 4000 +ves, 30000 -ves).

Please suggest:

  1. Is it because of my cropping style or something else is wrong, that the learning stops at stage:4

  2. I am using images from a single aircraft model at fairly close angles (extracted from a video), but still covering 180 deg view I require for detection. It it required to have +ve images only from same view angle?

  3. Please use the attached cropped image and suggest what would be the best cropping style for these type of objects.

Thanks.

image description

edit retag flag offensive close merge delete

1 answer

Sort by » oldest newest most voted
1

answered 2013-06-14 03:34:28 -0500

vinayverma gravatar image

I am answering it on the basis of my cropping experience of Haartraining for objects which are difficult to crop without keeping considerable amount of background.

If we have more background pixels on positive images in comparison with actual object’s pixels, it’s not good because the background would be learnt too, as a feature of our object. The best way in that case is to fill the background with random noise hence avoiding constant background. This would also avoid learning of the background during Haartraining.

edit flag offensive delete link more
Login/Signup to Answer

Question Tools

Stats

Asked: 2013-05-12 23:37:22 -0500

Seen: 1,258 times

Last updated: Jun 14 '13