L1 loss different scale for small or large boxes [object detetion]

In this paper there is a section where it is mentioned that “The most commonly used L1 loss will have different scales for small and large objects even if their relative errors are similar”.

I don’t understand this statement. Can anybody explain?