Image Labeler ROI Selection

I am using the Image Labeler to select objects of interest in a larger image in order to train a cascade object detector.

The question is how background information affects training/detection. Specifically, should I try and make the ROI as small as possible (while still fitting in all of the object to be detected), or include more of the background to provide contextual information. As it is the object to be detected will occur in front of varying background.

Answer 1

[Greg Heath]

Somewhere in between. Exactly how much is determined by trial and error.

Hope this helps.

Greg

[Josef Kriegl]

Thanks Greg. I will try different ROI/object ratios to see how they affect detection and classification performance.

[Shamil Alakbarov]

Hey, Josef Kriegl, I'm dealing with the same issue, any conclusions?