Image Labeler ROI selection

Hi,

In my ongoing project in the study of Matlab, I am trying to create a CNN using digital images and which capable to recognize specific elements. I have created a set of positive images and negative images (each 20x20 pxl), cropped from a larger image (400x400 pxl). The structure I am interested in is relatively simple and really stands out from the background, so it is not hard to create positive and negative images. From the info in "Train a Cascade Object Detector" page, I understand that I can use these 2 image types to train an "Object detector".

My question is:

Does Image Labeler select the ROI of the selected images (20x20 pxl) automatically or do I have to select the complete ROI of each individual image manually? The info on the webpage is somewhat convoluted about this. It would help and save a lot of work if this was explained in a little more detail? Thanks!