Ask Your Question

Object Detection Positive Samples Background

asked 2012-10-19 10:46:55 -0600

rphv gravatar image

I have a large database of images representing the object I want to detect. These images are carefully cropped to the (non-rectangular) outline of my object. However, this means that parts of my positive sample images can be considered "background." Right now, this "background" is black (RGB 0,0,0). Is this the best strategy, or should the "background" regions of my positive image samples be filled with some other value (e.g., white / random noise)?

edit retag flag offensive close merge delete

1 answer

Sort by » oldest newest most voted

answered 2013-06-14 04:48:03 -0600

vinayverma gravatar image

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


As an addemdum to that, if you can to detect your object in a specific scene, train the object in that scene also. By that you do not learn features that are created due to steep edges with the black background. It is the approach I am using and it seems working rather well.

StevenPuttemans gravatar imageStevenPuttemans ( 2013-06-14 04:50:25 -0600 )edit

Question Tools


Asked: 2012-10-19 10:46:55 -0600

Seen: 474 times

Last updated: Jun 14 '13