Negative sample image sizes for cascading classifier training in OpenCV?

Question

So, following here , I now need to collect negative samples for cascading classification using OpenCV. With positive samples, I know that all samples should have the same aspect ratio.

How about negative samples?

Should they be bigger than positive samples (since OpenCV is going to insert positive expressions on top of negatives to create test images).

Should all be the same size?

Can they be arbitrary sizes?

Should they have the same aspect ratio among themselves?

Answer 1

From OpenCV Cascading Classifier Training Document :

Negative samples are taken from arbitrary images. These images should not contain detected objects. [...] The described images may be of different sizes . But each image must be (but not necessarily) larger than the size of the training window , because these images are used to subsample a negative image in the amount of training.