These small image values are the positives (objects) for which you want to train the classifier. The object in your real frame can be of arbitrary size, because the cascading classifier works for different image scales.
Here is a good tutorial that helped me when I trained my own classifier. Your cropped images used for training may be larger, but when you createsamples you need to specify the size at which the positive textures are scaled. These new tiny patterns are the ones used for the classifier. It also affects the speed of the cascading classifier, so they are tiny in size.
Background images may be larger in size if I'm not mistaken, but I remember that I still cropped the background images that I had in smaller sizes.
When you run your classifier with real 640x480 data, you specify the limits of the minimum size that can be positive (of course, this value must be at least the size -w -h that you specified earlier), as well as the maximum expected size.
The haar detector will only search for objects in the window range of your test image, which can be as large as possible.