When making data, draw a red box in the center of where to put the objects for recognition
After taking images, shift the image in various random shifts (deleting the shifted-out sides), while ensuring that everything in the red box remains somewhere in the image.
I'm thinking the following: