How do I generate a good cascade file to detect an image
What value should we use as -numNeg and -numPos. Why are negatives needed at all? For example, When I am trying to detect a particular car image, and not sure about the background,why is negative dataset required? Can I generate a cascade file with no negative images and one car image?
it is logic: when you learn that an icecream is an icecream, then you also learn what is not an icecream, that is why the background is needed, to say what not to detect.
thank you, but what should I do to detect an image (one particular image), which can be anywhere, on a wall, a desk, a bottle, etc. Then where do I get these negatives from? Any suggestion would be of great help.
The negatives are images (no matter the size, but at least the size of the samples) that do not contain the object in your positives (e.g. positives just faces, negatives images without faces, like bicycles, rooms, cars, cups, etc, but no faces on them)
I get it now, I have to have negatives, which can be anything other than my image. What should be the number of negatives when I have only one positive image?
I would start having at least 10 positives, otherwise you should do a template matching, better than cascade for one object only. And the number of negatives should be at least 10 times bigger than the positives, but using big images will solve this problem, because the negative samples will be picked from the big ones (cropped and resized)